ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
#include "ggml-vulkan.h"
|
|
|
|
|
|
|
|
#ifdef VK_RUN_TESTS
|
|
|
|
#include <chrono>
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#include <vulkan/vulkan.hpp>
|
|
|
|
|
|
|
|
#include <algorithm>
|
|
|
|
#include <cmath>
|
|
|
|
#include <iostream>
|
|
|
|
#include <iomanip>
|
|
|
|
#include <limits>
|
|
|
|
#include <tuple>
|
|
|
|
#include <vector>
|
|
|
|
#include <sstream>
|
|
|
|
#include <utility>
|
|
|
|
|
|
|
|
#include "ggml.h"
|
|
|
|
#include "ggml-backend-impl.h"
|
|
|
|
|
|
|
|
#include "ggml-vulkan-shaders.hpp"
|
|
|
|
|
|
|
|
#define VK_API_VERSION VK_API_VERSION_1_2
|
|
|
|
|
|
|
|
#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
|
|
|
|
|
|
|
|
#define VK_VENDOR_ID_AMD 0x1002
|
|
|
|
#define VK_VENDOR_ID_INTEL 0x8086
|
|
|
|
#define VK_VENDOR_ID_NVIDIA 0x10de
|
|
|
|
|
|
|
|
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN 0
|
|
|
|
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI 1
|
|
|
|
#define VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE 2
|
|
|
|
|
|
|
|
#define VK_NUM_TYPES 16
|
|
|
|
|
|
|
|
#define GGML_VK_MAX_NODES 8192
|
|
|
|
|
|
|
|
#ifndef K_QUANTS_PER_ITERATION
|
|
|
|
#define K_QUANTS_PER_ITERATION 1
|
|
|
|
#else
|
|
|
|
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define VK_CHECK(err, msg) \
|
|
|
|
do { \
|
|
|
|
vk::Result err_ = (err); \
|
|
|
|
if (err_ != vk::Result::eSuccess) { \
|
|
|
|
fprintf(stderr, "ggml_vulkan: %s error %s at %s:%d\n", \
|
|
|
|
#err, to_string(err_).c_str(), __FILE__, __LINE__); \
|
|
|
|
exit(1); \
|
|
|
|
} \
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
struct vk_buffer {
|
|
|
|
vk::Buffer buffer;
|
|
|
|
vk::DeviceMemory device_memory;
|
|
|
|
vk::MemoryPropertyFlags memory_property_flags;
|
|
|
|
void * ptr;
|
|
|
|
size_t size = 0;
|
|
|
|
uint32_t qf_owner;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_subbuffer {
|
|
|
|
vk_buffer buffer;
|
|
|
|
uint64_t offset;
|
|
|
|
uint64_t size;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_pipeline {
|
|
|
|
std::string name;
|
|
|
|
vk::DescriptorSetLayout dsl;
|
|
|
|
std::vector<vk::DescriptorPool> descriptor_pools;
|
|
|
|
std::vector<vk::DescriptorSet> descriptor_sets;
|
|
|
|
uint32_t descriptor_set_idx;
|
|
|
|
vk::PipelineLayout layout;
|
|
|
|
vk::Pipeline pipeline;
|
|
|
|
uint32_t push_constant_size;
|
|
|
|
uint32_t parameter_count;
|
|
|
|
std::array<uint32_t, 3> wg_denoms;
|
|
|
|
uint32_t align;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_queue {
|
|
|
|
uint32_t queue_family_index;
|
|
|
|
vk::Queue queue;
|
|
|
|
vk::CommandPool pool;
|
|
|
|
uint32_t cmd_buffer_idx;
|
|
|
|
std::vector<vk::CommandBuffer> cmd_buffers;
|
|
|
|
|
|
|
|
vk::PipelineStageFlags stage_flags;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_semaphore {
|
|
|
|
vk::Semaphore s;
|
|
|
|
uint64_t value;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_submission {
|
|
|
|
vk::CommandBuffer buffer;
|
|
|
|
std::vector<vk_semaphore> wait_semaphores;
|
|
|
|
std::vector<vk_semaphore> signal_semaphores;
|
|
|
|
};
|
|
|
|
|
|
|
|
typedef std::vector<vk_submission> vk_sequence;
|
|
|
|
|
|
|
|
struct vk_device {
|
|
|
|
vk::PhysicalDevice physical_device;
|
|
|
|
vk::PhysicalDeviceProperties properties;
|
|
|
|
uint64_t max_memory_allocation_size;
|
|
|
|
bool fp16;
|
|
|
|
vk::Device device;
|
|
|
|
uint32_t vendor_id;
|
|
|
|
vk_queue compute_queue;
|
|
|
|
vk_queue transfer_queue;
|
|
|
|
uint32_t descriptor_set_mode;
|
|
|
|
uint32_t subgroup_size;
|
2024-01-30 12:59:30 +00:00
|
|
|
bool uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_op_push_constants {
|
|
|
|
uint32_t KX;
|
|
|
|
uint32_t KY;
|
|
|
|
float param1;
|
|
|
|
float param2;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_op_cpy_push_constants {
|
|
|
|
uint32_t ne;
|
|
|
|
uint32_t ne00; uint32_t ne01; uint32_t nb00; uint32_t nb01; uint32_t nb02;
|
|
|
|
uint32_t ne10; uint32_t ne11; uint32_t nb10; uint32_t nb11; uint32_t nb12;
|
|
|
|
uint32_t d_offset;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_op_diag_mask_push_constants {
|
|
|
|
uint32_t ncols;
|
|
|
|
uint32_t rows_per_channel;
|
|
|
|
int32_t n_past;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_op_rope_push_constants {
|
|
|
|
uint32_t ncols;
|
|
|
|
float freq_scale;
|
|
|
|
uint32_t p_delta_rows;
|
|
|
|
float freq_base;
|
|
|
|
float ext_factor;
|
|
|
|
float attn_factor;
|
|
|
|
float corr_dims[4];
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_op_rope_neox_push_constants {
|
|
|
|
uint32_t ncols;
|
|
|
|
uint32_t ndims;
|
|
|
|
float freq_scale;
|
|
|
|
uint32_t p_delta_rows;
|
|
|
|
float freq_base;
|
|
|
|
float ext_factor;
|
|
|
|
float attn_factor;
|
|
|
|
float corr_dims[4];
|
|
|
|
float theta_scale;
|
|
|
|
float inv_ndims;
|
|
|
|
};
|
|
|
|
|
|
|
|
// Allow pre-recording command buffers
|
|
|
|
struct vk_staging_memcpy {
|
|
|
|
vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
|
|
|
|
|
|
|
|
void * dst;
|
|
|
|
const void * src;
|
|
|
|
size_t n;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct vk_context {
|
|
|
|
size_t idx;
|
|
|
|
|
|
|
|
vk_submission * s;
|
|
|
|
std::vector<vk_sequence> seqs;
|
|
|
|
|
|
|
|
ggml_tensor * exit_tensor;
|
|
|
|
|
|
|
|
std::vector<vk_staging_memcpy> in_memcpys;
|
|
|
|
std::vector<vk_staging_memcpy> out_memcpys;
|
|
|
|
|
|
|
|
vk_queue * q;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct ggml_tensor_extra_gpu {
|
|
|
|
bool ready;
|
|
|
|
|
|
|
|
size_t ctx_idx;
|
|
|
|
|
|
|
|
vk_buffer buffer_gpu;
|
|
|
|
uint64_t offset;
|
|
|
|
|
|
|
|
void reset() {
|
|
|
|
ready = false;
|
|
|
|
ctx_idx = 0;
|
|
|
|
buffer_gpu.size = 0;
|
|
|
|
offset = 0;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
struct ggml_vk_garbage_collector {
|
|
|
|
std::vector<vk_pipeline *> pipelines;
|
|
|
|
std::vector<vk_semaphore> tl_semaphores;
|
|
|
|
std::vector<vk_semaphore> semaphores;
|
|
|
|
std::vector<vk::Event> events;
|
|
|
|
std::vector<vk_buffer> temp_buffers;
|
|
|
|
std::vector<vk_context> contexts;
|
|
|
|
};
|
|
|
|
|
|
|
|
typedef void (*ggml_vk_func_t)(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
|
|
|
|
|
|
|
|
vk::Instance vk_instance;
|
|
|
|
vk_device vk_device;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f32_l, vk_pipeline_matmul_f32_m, vk_pipeline_matmul_f32_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f32_aligned_l, vk_pipeline_matmul_f32_aligned_m, vk_pipeline_matmul_f32_aligned_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f16_l, vk_pipeline_matmul_f16_m, vk_pipeline_matmul_f16_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f16_aligned_l, vk_pipeline_matmul_f16_aligned_m, vk_pipeline_matmul_f16_aligned_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f16_f32_l, vk_pipeline_matmul_f16_f32_m, vk_pipeline_matmul_f16_f32_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_f16_f32_aligned_l, vk_pipeline_matmul_f16_f32_aligned_m, vk_pipeline_matmul_f16_f32_aligned_s;
|
|
|
|
vk_pipeline vk_pipeline_matmul_split_k_reduce;
|
|
|
|
vk_pipeline vk_pipeline_dequant[VK_NUM_TYPES];
|
|
|
|
vk_pipeline vk_pipeline_dequant_mul_mat_vec_f32[VK_NUM_TYPES];
|
|
|
|
vk_pipeline vk_pipeline_mul_mat_vec_p021_f16_f32;
|
|
|
|
vk_pipeline vk_pipeline_mul_mat_vec_nc_f16_f32;
|
|
|
|
vk_pipeline vk_pipeline_get_rows[VK_NUM_TYPES];
|
|
|
|
vk_pipeline vk_pipeline_get_rows_f32[VK_NUM_TYPES];
|
|
|
|
vk_pipeline vk_pipeline_mul_f32;
|
|
|
|
vk_pipeline vk_pipeline_add_f32;
|
|
|
|
vk_pipeline vk_pipeline_scale_f32;
|
|
|
|
vk_pipeline vk_pipeline_sqr_f32;
|
|
|
|
vk_pipeline vk_pipeline_clamp_f32;
|
|
|
|
vk_pipeline vk_pipeline_cpy_f32_f32, vk_pipeline_cpy_f32_f16, vk_pipeline_cpy_f16_f16;
|
|
|
|
vk_pipeline vk_pipeline_norm_f32;
|
|
|
|
vk_pipeline vk_pipeline_rms_norm_f32;
|
|
|
|
vk_pipeline vk_pipeline_gelu_f32;
|
|
|
|
vk_pipeline vk_pipeline_silu_f32;
|
|
|
|
vk_pipeline vk_pipeline_relu_f32;
|
|
|
|
vk_pipeline vk_pipeline_diag_mask_inf_f32;
|
|
|
|
vk_pipeline vk_pipeline_soft_max_f32;
|
|
|
|
vk_pipeline vk_pipeline_rope_f32, vk_pipeline_rope_f16;
|
|
|
|
vk_pipeline vk_pipeline_rope_neox_f32, vk_pipeline_rope_neox_f16;
|
|
|
|
|
|
|
|
static size_t vk_semaphore_idx, vk_event_idx;
|
|
|
|
static ggml_vk_garbage_collector vk_gc;
|
|
|
|
static std::vector<std::tuple<void*, size_t, vk_buffer>> vk_pinned_memory;
|
|
|
|
static size_t vk_prealloc_size_qx, vk_prealloc_size_qy, vk_prealloc_size_x, vk_prealloc_size_y, vk_prealloc_size_split_k;
|
|
|
|
static vk_buffer vk_prealloc_qx, vk_prealloc_qy, vk_prealloc_x, vk_prealloc_y, vk_prealloc_split_k;
|
|
|
|
static vk::Fence vk_fence;
|
|
|
|
static vk_buffer vk_staging;
|
|
|
|
static size_t vk_staging_size;
|
|
|
|
static size_t vk_staging_offset;
|
|
|
|
static vk_buffer vk_sync_staging;
|
|
|
|
|
|
|
|
static vk_context * vk_ctx;
|
|
|
|
|
|
|
|
static bool vk_disable;
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
|
|
|
size_t vk_skip_checks;
|
|
|
|
size_t vk_output_tensor;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static vk_pipeline ggml_vk_create_pipeline(const std::string& name, size_t spv_size, const void* spv_data, const std::string& entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_pipeline(" << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(parameter_count > 0);
|
|
|
|
GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
|
|
|
|
|
|
|
|
vk_pipeline pipeline;
|
|
|
|
|
|
|
|
pipeline.name = name;
|
|
|
|
pipeline.parameter_count = parameter_count;
|
|
|
|
pipeline.push_constant_size = push_constant_size;
|
|
|
|
pipeline.wg_denoms = wg_denoms;
|
|
|
|
pipeline.align = align;
|
|
|
|
|
|
|
|
vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
|
|
|
|
vk::ShaderModule shader_module = vk_device.device.createShaderModule(shader_module_create_info);
|
|
|
|
|
|
|
|
std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
|
|
|
|
std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
|
|
|
|
for (uint32_t i = 0; i < parameter_count; i++) {
|
|
|
|
dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
|
|
|
|
dsl_binding_flags.push_back({});
|
|
|
|
}
|
|
|
|
|
|
|
|
vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
|
|
|
|
|
|
|
|
vk::PushConstantRange pcr(
|
|
|
|
vk::ShaderStageFlagBits::eCompute,
|
|
|
|
0,
|
|
|
|
pipeline.push_constant_size
|
|
|
|
);
|
|
|
|
|
|
|
|
vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
|
|
|
|
{},
|
|
|
|
dsl_binding);
|
|
|
|
descriptor_set_layout_create_info.setPNext(&dslbfci);
|
|
|
|
pipeline.dsl = vk_device.device.createDescriptorSetLayout(descriptor_set_layout_create_info);
|
|
|
|
|
|
|
|
// Check if device supports multiple descriptors per pool
|
|
|
|
if (vk_device.descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN) {
|
|
|
|
const uint32_t alloc_count = 2;
|
|
|
|
|
|
|
|
// Try allocating multiple sets from one pool
|
|
|
|
// This fails on AMD for some reason, so add a fall back to allocating one pool per set
|
|
|
|
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count);
|
|
|
|
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, alloc_count, descriptor_pool_size);
|
|
|
|
vk::DescriptorPool pool = vk_device.device.createDescriptorPool(descriptor_pool_create_info);
|
|
|
|
|
|
|
|
std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
|
|
|
|
for (uint32_t i = 0; i < alloc_count; i++) {
|
|
|
|
layouts[i] = pipeline.dsl;
|
|
|
|
}
|
|
|
|
try {
|
|
|
|
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pool, alloc_count, layouts.data());
|
|
|
|
std::vector<vk::DescriptorSet> sets = vk_device.device.allocateDescriptorSets(descriptor_set_alloc_info);
|
|
|
|
} catch(vk::OutOfPoolMemoryError const&) {
|
|
|
|
vk_device.descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_SINGLE;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_device.device.destroyDescriptorPool(pool);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vk_device.descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) {
|
|
|
|
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count);
|
|
|
|
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 128, descriptor_pool_size);
|
|
|
|
pipeline.descriptor_pools.push_back(vk_device.device.createDescriptorPool(descriptor_pool_create_info));
|
|
|
|
}
|
|
|
|
|
|
|
|
pipeline.descriptor_set_idx = 0;
|
|
|
|
|
|
|
|
vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline.dsl, pcr);
|
|
|
|
pipeline.layout = vk_device.device.createPipelineLayout(pipeline_layout_create_info);
|
|
|
|
|
|
|
|
std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size());
|
|
|
|
|
|
|
|
for (size_t i = 0; i < specialization_constants.size(); i++) {
|
|
|
|
specialization_entries[i].constantID = i;
|
|
|
|
specialization_entries[i].offset = i * sizeof(uint32_t);
|
|
|
|
specialization_entries[i].size = sizeof(uint32_t);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk::SpecializationInfo specialization_info(
|
|
|
|
specialization_entries.size(),
|
|
|
|
specialization_entries.data(),
|
|
|
|
specialization_constants.size() * sizeof(uint32_t),
|
|
|
|
specialization_constants.data()
|
|
|
|
);
|
|
|
|
|
|
|
|
vk::PipelineShaderStageCreateInfo pipeline_shader_create_info(
|
|
|
|
vk::PipelineShaderStageCreateFlags(),
|
|
|
|
vk::ShaderStageFlagBits::eCompute,
|
|
|
|
shader_module,
|
|
|
|
entrypoint.c_str(),
|
|
|
|
&specialization_info);
|
|
|
|
vk::ComputePipelineCreateInfo compute_pipeline_create_info(
|
|
|
|
vk::PipelineCreateFlags(),
|
|
|
|
pipeline_shader_create_info,
|
|
|
|
pipeline.layout);
|
|
|
|
pipeline.pipeline = vk_device.device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
|
|
|
|
|
|
|
|
return pipeline;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_pipeline_allocate_descriptor_sets(vk_pipeline& pipeline, uint32_t n) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_pipeline_allocate_descriptor_sets(" << pipeline.name << ", " << n << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
// Check if gc already contains pipeline before adding it
|
|
|
|
bool gc_found = false;
|
|
|
|
for (auto * pl : vk_gc.pipelines) {
|
|
|
|
if (&pipeline == pl) {
|
|
|
|
gc_found = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!gc_found) {
|
|
|
|
vk_gc.pipelines.push_back(&pipeline);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (pipeline.descriptor_sets.size() >= pipeline.descriptor_set_idx + n) {
|
|
|
|
// Enough descriptors are available
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vk_device.descriptor_set_mode == VK_DEVICE_DESCRIPTOR_POOL_MODE_MULTI) {
|
|
|
|
const uint32_t alloc_count = pipeline.descriptor_set_idx + n - pipeline.descriptor_sets.size();
|
|
|
|
|
|
|
|
std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
|
|
|
|
for (uint32_t i = 0; i < alloc_count; i++) {
|
|
|
|
layouts[i] = pipeline.dsl;
|
|
|
|
}
|
|
|
|
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline.descriptor_pools[0], alloc_count, layouts.data());
|
|
|
|
std::vector<vk::DescriptorSet> sets = vk_device.device.allocateDescriptorSets(descriptor_set_alloc_info);
|
|
|
|
pipeline.descriptor_sets.insert(pipeline.descriptor_sets.end(), sets.begin(), sets.end());
|
|
|
|
} else {
|
|
|
|
for (uint32_t i = pipeline.descriptor_sets.size(); i < pipeline.descriptor_set_idx + n; i++) {
|
|
|
|
vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline.parameter_count);
|
|
|
|
vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, 1, descriptor_pool_size);
|
|
|
|
pipeline.descriptor_pools.push_back(vk_device.device.createDescriptorPool(descriptor_pool_create_info));
|
|
|
|
|
|
|
|
vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline.descriptor_pools[i], 1, &pipeline.dsl);
|
|
|
|
std::vector<vk::DescriptorSet> sets = vk_device.device.allocateDescriptorSets(descriptor_set_alloc_info);
|
|
|
|
pipeline.descriptor_sets.push_back(sets[0]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_pipeline_cleanup(vk_pipeline& pipeline) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_pipeline_cleanup(" << pipeline.name << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
pipeline.descriptor_set_idx = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_queue& q) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_cmd_buffer()" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (q.cmd_buffers.size() > q.cmd_buffer_idx) {
|
|
|
|
// Reuse command buffer
|
|
|
|
return q.cmd_buffers[q.cmd_buffer_idx++];
|
|
|
|
}
|
|
|
|
|
|
|
|
vk::CommandBufferAllocateInfo command_buffer_alloc_info(
|
|
|
|
q.pool,
|
|
|
|
vk::CommandBufferLevel::ePrimary,
|
|
|
|
1);
|
|
|
|
const std::vector<vk::CommandBuffer> cmd_buffers = vk_device.device.allocateCommandBuffers(command_buffer_alloc_info);
|
|
|
|
auto buf = cmd_buffers.front();
|
|
|
|
|
|
|
|
q.cmd_buffers.push_back(buf);
|
|
|
|
q.cmd_buffer_idx++;
|
|
|
|
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_submission ggml_vk_create_submission(vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_submission()" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_submission s;
|
|
|
|
s.buffer = ggml_vk_create_cmd_buffer(q);
|
|
|
|
s.wait_semaphores = std::move(wait_semaphores);
|
|
|
|
s.signal_semaphores = std::move(signal_semaphores);
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_sequence ggml_vk_create_sequence_1(vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_sequence_1()" << std::endl;
|
|
|
|
#endif
|
|
|
|
return { ggml_vk_create_submission(q, std::move(wait_semaphores), std::move(signal_semaphores)) };
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_submit(vk_context * ctx, vk::Fence fence) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_submit(" << ctx->seqs.size() << ", " << fence << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (ctx->seqs.empty()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::vector<std::vector<uint64_t>> tl_wait_vals;
|
|
|
|
std::vector<std::vector<uint64_t>> tl_signal_vals;
|
|
|
|
std::vector<std::vector<vk::Semaphore>> tl_wait_semaphores;
|
|
|
|
std::vector<std::vector<vk::Semaphore>> tl_signal_semaphores;
|
|
|
|
std::vector<vk::TimelineSemaphoreSubmitInfo> tl_submit_infos;
|
|
|
|
std::vector<vk::SubmitInfo> submit_infos;
|
|
|
|
int idx = -1;
|
|
|
|
std::vector<std::vector<vk::PipelineStageFlags>> stage_flags;
|
|
|
|
|
|
|
|
size_t reserve = 0;
|
|
|
|
|
|
|
|
for (const auto& sequence : ctx->seqs) {
|
|
|
|
reserve += sequence.size();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Pre-reserve vectors to prevent reallocation, which invalidates pointers
|
|
|
|
tl_wait_semaphores.reserve(reserve);
|
|
|
|
tl_wait_vals.reserve(reserve);
|
|
|
|
tl_signal_semaphores.reserve(reserve);
|
|
|
|
tl_signal_vals.reserve(reserve);
|
|
|
|
tl_submit_infos.reserve(reserve);
|
|
|
|
submit_infos.reserve(reserve);
|
|
|
|
stage_flags.reserve(reserve);
|
|
|
|
|
|
|
|
for (const auto& sequence : ctx->seqs) {
|
|
|
|
for (const auto& submission : sequence) {
|
|
|
|
stage_flags.push_back({});
|
|
|
|
idx++;
|
|
|
|
tl_wait_vals.push_back({});
|
|
|
|
tl_wait_semaphores.push_back({});
|
|
|
|
tl_signal_vals.push_back({});
|
|
|
|
tl_signal_semaphores.push_back({});
|
|
|
|
for (size_t i = 0; i < submission.wait_semaphores.size(); i++) {
|
|
|
|
stage_flags[idx].push_back(ctx->q->stage_flags);
|
|
|
|
tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value);
|
|
|
|
tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s);
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < submission.signal_semaphores.size(); i++) {
|
|
|
|
tl_signal_vals[idx].push_back(submission.signal_semaphores[i].value);
|
|
|
|
tl_signal_semaphores[idx].push_back(submission.signal_semaphores[i].s);
|
|
|
|
}
|
|
|
|
tl_submit_infos.push_back({
|
|
|
|
(uint32_t) submission.wait_semaphores.size(),
|
|
|
|
tl_wait_vals[idx].data(),
|
|
|
|
(uint32_t) submission.signal_semaphores.size(),
|
|
|
|
tl_signal_vals[idx].data(),
|
|
|
|
});
|
|
|
|
tl_submit_infos[idx].sType = vk::StructureType::eTimelineSemaphoreSubmitInfo;
|
|
|
|
tl_submit_infos[idx].pNext = nullptr;
|
|
|
|
vk::SubmitInfo si{
|
|
|
|
(uint32_t) submission.wait_semaphores.size(),
|
|
|
|
tl_wait_semaphores[idx].data(),
|
|
|
|
stage_flags[idx].data(),
|
|
|
|
1,
|
|
|
|
&submission.buffer,
|
|
|
|
(uint32_t) submission.signal_semaphores.size(),
|
|
|
|
tl_signal_semaphores[idx].data(),
|
|
|
|
};
|
|
|
|
si.setPNext(&tl_submit_infos[idx]);
|
|
|
|
submit_infos.push_back(si);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ctx->q->queue.submit(submit_infos, fence);
|
|
|
|
|
|
|
|
ctx->seqs.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyProperties>& queue_family_props, const vk::QueueFlags& required, const vk::QueueFlags& avoid, int32_t compute_index, uint32_t min_num_queues) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_find_queue_family_index()" << std::endl;
|
|
|
|
#endif
|
|
|
|
const uint32_t qfsize = queue_family_props.size();
|
|
|
|
|
|
|
|
// Try with avoid preferences first
|
|
|
|
for (uint32_t i = 0; i < qfsize; i++) {
|
|
|
|
if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required && !(queue_family_props[i].queueFlags & avoid)) {
|
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fall back to only required
|
|
|
|
for (size_t i = 0; i < qfsize; i++) {
|
|
|
|
if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required) {
|
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fall back to reusing compute queue
|
|
|
|
for (size_t i = 0; i < qfsize; i++) {
|
|
|
|
if (queue_family_props[i].queueCount >= min_num_queues && queue_family_props[i].queueFlags & required) {
|
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Fall back to ignoring min_num_queries
|
|
|
|
for (size_t i = 0; i < qfsize; i++) {
|
|
|
|
if (queue_family_props[i].queueFlags & required) {
|
|
|
|
return i;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;
|
|
|
|
|
|
|
|
for(auto &q_family : queue_family_props) {
|
|
|
|
std::cerr << "Queue number: " + std::to_string(q_family.queueCount) << " flags: " + to_string(q_family.queueFlags) << std::endl;
|
|
|
|
}
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_queue ggml_vk_create_queue(uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_queue()" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_queue q;
|
|
|
|
q.queue_family_index = queue_family_index;
|
|
|
|
|
|
|
|
vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index);
|
|
|
|
q.pool = vk_device.device.createCommandPool(command_pool_create_info_compute);
|
|
|
|
|
|
|
|
q.cmd_buffer_idx = 0;
|
|
|
|
|
|
|
|
q.queue = vk_device.device.getQueue(queue_family_index, queue_index);
|
|
|
|
|
|
|
|
q.stage_flags = stage_flags;
|
|
|
|
|
|
|
|
return q;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_context * ggml_vk_create_context(vk_queue& q) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_context()" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_gc.contexts.emplace_back();
|
|
|
|
vk_context * result = &vk_gc.contexts[vk_gc.contexts.size() - 1];
|
|
|
|
memset((void *) result, 0, sizeof(vk_context));
|
|
|
|
result->idx = vk_gc.contexts.size() - 1;
|
|
|
|
result->q = &q;
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_semaphore * ggml_vk_create_binary_semaphore() {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_timeline_semaphore()" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eBinary, 0 };
|
|
|
|
vk::SemaphoreCreateInfo ci{};
|
|
|
|
ci.setPNext(&tci);
|
|
|
|
vk::Semaphore semaphore = vk_device.device.createSemaphore(ci);
|
|
|
|
vk_gc.semaphores.push_back({ semaphore, 0 });
|
|
|
|
return &vk_gc.semaphores[vk_gc.semaphores.size() - 1];
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_semaphore * ggml_vk_create_timeline_semaphore() {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_timeline_semaphore()" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (vk_semaphore_idx >= vk_gc.tl_semaphores.size()) {
|
|
|
|
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
|
|
|
|
vk::SemaphoreCreateInfo ci{};
|
|
|
|
ci.setPNext(&tci);
|
|
|
|
vk::Semaphore semaphore = vk_device.device.createSemaphore(ci);
|
|
|
|
vk_gc.tl_semaphores.push_back({ semaphore, 0 });
|
|
|
|
}
|
|
|
|
return &vk_gc.tl_semaphores[vk_semaphore_idx++];
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk::Event ggml_vk_create_event() {
|
|
|
|
if (vk_event_idx >= vk_gc.events.size()) {
|
|
|
|
vk_gc.events.push_back(vk_device.device.createEvent({}));
|
|
|
|
}
|
|
|
|
return vk_gc.events[vk_event_idx++];
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_queue_cleanup(vk_queue& q) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_queue_cleanup()" << std::endl;
|
|
|
|
#endif
|
|
|
|
// Requires command buffers to be done
|
|
|
|
|
|
|
|
vk_device.device.resetCommandPool(q.pool);
|
|
|
|
q.cmd_buffer_idx = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_buffer ggml_vk_create_buffer(size_t size, vk::MemoryPropertyFlags req_flags) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(size > 0);
|
|
|
|
|
|
|
|
vk_buffer buf;
|
|
|
|
|
|
|
|
buf.size = size;
|
|
|
|
vk::BufferCreateInfo buffer_create_info{
|
|
|
|
vk::BufferCreateFlags(),
|
|
|
|
size,
|
|
|
|
vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst,
|
|
|
|
vk::SharingMode::eExclusive,
|
|
|
|
0,
|
|
|
|
nullptr,
|
|
|
|
};
|
|
|
|
|
|
|
|
buf.buffer = vk_device.device.createBuffer(buffer_create_info);
|
|
|
|
|
|
|
|
vk::MemoryRequirements mem_req = vk_device.device.getBufferMemoryRequirements(buf.buffer);
|
|
|
|
|
|
|
|
vk::PhysicalDeviceMemoryProperties mem_props = vk_device.physical_device.getMemoryProperties();
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
uint32_t memory_type_index = UINT32_MAX;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
for (uint32_t i = 0; i < mem_props.memoryTypeCount; ++i) {
|
|
|
|
vk::MemoryType memory_type = mem_props.memoryTypes[i];
|
|
|
|
if ((mem_req.memoryTypeBits & ((uint64_t)1 << i)) && (req_flags & memory_type.propertyFlags) == req_flags && mem_props.memoryHeaps[memory_type.heapIndex].size >= mem_req.size) {
|
|
|
|
memory_type_index = i;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
if (memory_type_index >= mem_props.memoryTypeCount) {
|
|
|
|
throw vk::OutOfDeviceMemoryError("No suitable memory type found");
|
|
|
|
}
|
|
|
|
|
|
|
|
try {
|
|
|
|
buf.device_memory = vk_device.device.allocateMemory({ mem_req.size, memory_type_index });
|
|
|
|
} catch (const vk::SystemError& e) {
|
|
|
|
// Out of Host/Device memory, clean up buffer
|
|
|
|
vk_device.device.destroyBuffer(buf.buffer);
|
|
|
|
buf.size = 0;
|
|
|
|
throw e;
|
|
|
|
}
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
buf.memory_property_flags = req_flags;
|
|
|
|
buf.ptr = nullptr;
|
|
|
|
|
|
|
|
if (req_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
|
|
|
|
buf.ptr = vk_device.device.mapMemory(buf.device_memory, 0, VK_WHOLE_SIZE);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_device.device.bindBufferMemory(buf.buffer, buf.device_memory, 0);
|
|
|
|
|
|
|
|
buf.qf_owner = VK_QUEUE_FAMILY_IGNORED;
|
|
|
|
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static vk_buffer ggml_vk_create_buffer_check(size_t size, vk::MemoryPropertyFlags req_flags) {
|
|
|
|
try {
|
|
|
|
return ggml_vk_create_buffer(size, req_flags);
|
|
|
|
} catch (const vk::SystemError& e) {
|
|
|
|
std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
|
|
|
|
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
|
|
|
|
throw e;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_buffer ggml_vk_create_buffer_device(size_t size) {
|
|
|
|
vk_buffer buf;
|
|
|
|
try {
|
|
|
|
buf = ggml_vk_create_buffer(size, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
|
|
|
} catch (const vk::SystemError& e) {
|
|
|
|
if (vk_device.uma) {
|
|
|
|
// Fall back to host memory type
|
|
|
|
buf = ggml_vk_create_buffer_check(size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
|
|
|
|
} else {
|
|
|
|
std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
|
|
|
|
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
|
|
|
|
throw e;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_destroy_buffer(vk_buffer& buf) {
|
|
|
|
if (buf.size == 0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_destroy_buffer(" << buf.size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
buf.size = 0;
|
|
|
|
vk_device.device.freeMemory(buf.device_memory);
|
|
|
|
vk_device.device.destroyBuffer(buf.buffer);
|
|
|
|
}
|
|
|
|
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
|
|
|
|
return { buf, 0, VK_WHOLE_SIZE };
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_sync_buffers(vk_context * ctx) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_sync_buffers()" << std::endl;
|
|
|
|
#endif
|
|
|
|
const std::vector<vk::MemoryBarrier> mem_barriers{ { { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite }, { vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite } } };
|
|
|
|
|
|
|
|
ctx->s->buffer.pipelineBarrier(
|
|
|
|
ctx->q->stage_flags,
|
|
|
|
ctx->q->stage_flags,
|
|
|
|
{},
|
|
|
|
mem_barriers,
|
|
|
|
{},
|
|
|
|
{}
|
|
|
|
);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_wait_events(vk::CommandBuffer& cmd_buffer, std::vector<vk::Event>&& events, vk::PipelineStageFlags src_stages, vk::PipelineStageFlags dst_stages) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_wait_events()" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (events.empty()) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
cmd_buffer.waitEvents(
|
|
|
|
events,
|
|
|
|
src_stages,
|
|
|
|
dst_stages,
|
|
|
|
{},
|
|
|
|
{},
|
|
|
|
{}
|
|
|
|
);
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool ggml_vk_build_shader(ggml_type type) {
|
|
|
|
switch(type) {
|
|
|
|
case GGML_TYPE_F16:
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_load_shaders() {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_load_shaders()" << std::endl;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// mulmat
|
|
|
|
std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, vk_device.subgroup_size * 2, 64, 2, 4, 4, vk_device.subgroup_size };
|
|
|
|
std::initializer_list<uint32_t> warptile_m = { 128, 64, 64, 16, vk_device.subgroup_size, 32, 2, 4, 2, vk_device.subgroup_size };
|
2024-01-31 10:44:19 +00:00
|
|
|
std::initializer_list<uint32_t> warptile_s = { vk_device.subgroup_size, 32, 32, 16, 32, 32, 2, 2, 2, vk_device.subgroup_size };
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
|
|
|
|
std::array<uint32_t, 3> m_wg_denoms = { 64, 64, 1 };
|
|
|
|
std::array<uint32_t, 3> s_wg_denoms = { 32, 32, 1 };
|
|
|
|
|
|
|
|
uint32_t l_align = 128;
|
|
|
|
uint32_t m_align = 64;
|
|
|
|
uint32_t s_align = 32;
|
|
|
|
|
|
|
|
if (vk_device.fp16) {
|
|
|
|
vk_pipeline_matmul_f32_l = ggml_vk_create_pipeline("matmul_f32_l", matmul_f32_l_len, matmul_f32_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f32_m = ggml_vk_create_pipeline("matmul_f32_m", matmul_f32_m_len, matmul_f32_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f32_s = ggml_vk_create_pipeline("matmul_f32_s", matmul_f32_s_len, matmul_f32_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
vk_pipeline_matmul_f32_aligned_l = ggml_vk_create_pipeline("matmul_f32_aligned_l", matmul_f32_aligned_l_len, matmul_f32_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f32_aligned_m = ggml_vk_create_pipeline("matmul_f32_aligned_m", matmul_f32_aligned_m_len, matmul_f32_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f32_aligned_s = ggml_vk_create_pipeline("matmul_f32_aligned_s", matmul_f32_aligned_s_len, matmul_f32_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_l = ggml_vk_create_pipeline("matmul_f16_l", matmul_f16_l_len, matmul_f16_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f16_m = ggml_vk_create_pipeline("matmul_f16_m", matmul_f16_m_len, matmul_f16_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f16_s = ggml_vk_create_pipeline("matmul_f16_s", matmul_f16_s_len, matmul_f16_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_aligned_l = ggml_vk_create_pipeline("matmul_f16_aligned_l", matmul_f16_aligned_l_len, matmul_f16_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f16_aligned_m = ggml_vk_create_pipeline("matmul_f16_aligned_m", matmul_f16_aligned_m_len, matmul_f16_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f16_aligned_s = ggml_vk_create_pipeline("matmul_f16_aligned_s", matmul_f16_aligned_s_len, matmul_f16_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_f32_l = ggml_vk_create_pipeline("matmul_f16_f32_l", matmul_f16_f32_l_len, matmul_f16_f32_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_m = ggml_vk_create_pipeline("matmul_f16_f32_m", matmul_f16_f32_m_len, matmul_f16_f32_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_s = ggml_vk_create_pipeline("matmul_f16_f32_s", matmul_f16_f32_s_len, matmul_f16_f32_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_l = ggml_vk_create_pipeline("matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_l_len, matmul_f16_f32_aligned_l_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_m = ggml_vk_create_pipeline("matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_m_len, matmul_f16_f32_aligned_m_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_s = ggml_vk_create_pipeline("matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_s_len, matmul_f16_f32_aligned_s_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
// Build dequant shaders
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_F32] = ggml_vk_create_pipeline("f32_to_f16", f32_to_f16_len, f32_to_f16_data, "main", 2, 4 * sizeof(int), {64, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_F16] = ggml_vk_create_pipeline("dequant_f16", dequant_f16_len, dequant_f16_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("dequant_q4_0", dequant_q4_0_len, dequant_q4_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("dequant_q4_1", dequant_q4_1_len, dequant_q4_1_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("dequant_q2_K", dequant_q2_K_len, dequant_q2_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("dequant_q3_K", dequant_q3_K_len, dequant_q3_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("dequant_q4_K", dequant_q4_K_len, dequant_q4_K_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("dequant_q5_K", dequant_q5_K_len, dequant_q5_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("dequant_q6_K", dequant_q6_K_len, dequant_q6_K_data, "main", 2, 4 * sizeof(int), {256 * 64, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
// get_rows
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
} else {
|
|
|
|
vk_pipeline_matmul_f32_l = ggml_vk_create_pipeline("matmul_f32_l", matmul_f32_l_fp32_len, matmul_f32_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f32_m = ggml_vk_create_pipeline("matmul_f32_m", matmul_f32_m_fp32_len, matmul_f32_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f32_s = ggml_vk_create_pipeline("matmul_f32_s", matmul_f32_s_fp32_len, matmul_f32_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
vk_pipeline_matmul_f32_aligned_l = ggml_vk_create_pipeline("matmul_f32_aligned_l", matmul_f32_aligned_l_fp32_len, matmul_f32_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f32_aligned_m = ggml_vk_create_pipeline("matmul_f32_aligned_m", matmul_f32_aligned_m_fp32_len, matmul_f32_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f32_aligned_s = ggml_vk_create_pipeline("matmul_f32_aligned_s", matmul_f32_aligned_s_fp32_len, matmul_f32_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_l = ggml_vk_create_pipeline("matmul_f16_l", matmul_f16_l_fp32_len, matmul_f16_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f16_m = ggml_vk_create_pipeline("matmul_f16_m", matmul_f16_m_fp32_len, matmul_f16_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f16_s = ggml_vk_create_pipeline("matmul_f16_s", matmul_f16_s_fp32_len, matmul_f16_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_aligned_l = ggml_vk_create_pipeline("matmul_f16_aligned_l", matmul_f16_aligned_l_fp32_len, matmul_f16_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f16_aligned_m = ggml_vk_create_pipeline("matmul_f16_aligned_m", matmul_f16_aligned_m_fp32_len, matmul_f16_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f16_aligned_s = ggml_vk_create_pipeline("matmul_f16_aligned_s", matmul_f16_aligned_s_fp32_len, matmul_f16_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_f16_f32_l = ggml_vk_create_pipeline("matmul_f16_f32_l", matmul_f16_f32_l_fp32_len, matmul_f16_f32_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_m = ggml_vk_create_pipeline("matmul_f16_f32_m", matmul_f16_f32_m_fp32_len, matmul_f16_f32_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_s = ggml_vk_create_pipeline("matmul_f16_f32_s", matmul_f16_f32_s_fp32_len, matmul_f16_f32_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, 1);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_l = ggml_vk_create_pipeline("matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_l_fp32_len, matmul_f16_f32_aligned_l_fp32_data, "main", 3, 14 * sizeof(uint32_t), l_wg_denoms, warptile_l, l_align);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_m = ggml_vk_create_pipeline("matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_m_fp32_len, matmul_f16_f32_aligned_m_fp32_data, "main", 3, 14 * sizeof(uint32_t), m_wg_denoms, warptile_m, m_align);
|
|
|
|
vk_pipeline_matmul_f16_f32_aligned_s = ggml_vk_create_pipeline("matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_s_fp32_len, matmul_f16_f32_aligned_s_fp32_data, "main", 3, 14 * sizeof(uint32_t), s_wg_denoms, warptile_s, s_align);
|
|
|
|
|
|
|
|
// Build dequant shaders
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_F32] = ggml_vk_create_pipeline("f32_to_f16", f32_to_f16_fp32_len, f32_to_f16_fp32_data, "main", 2, 4 * sizeof(int), {64, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_F16] = ggml_vk_create_pipeline("dequant_f16", dequant_f16_fp32_len, dequant_f16_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("dequant_q4_0", dequant_q4_0_fp32_len, dequant_q4_0_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("dequant_q4_1", dequant_q4_1_fp32_len, dequant_q4_1_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("dequant_q5_0", dequant_q5_0_fp32_len, dequant_q5_0_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("dequant_q5_1", dequant_q5_1_fp32_len, dequant_q5_1_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("dequant_q8_0", dequant_q8_0_fp32_len, dequant_q8_0_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("dequant_q2_K", dequant_q2_K_fp32_len, dequant_q2_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("dequant_q3_K", dequant_q3_K_fp32_len, dequant_q3_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("dequant_q4_K", dequant_q4_K_fp32_len, dequant_q4_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("dequant_q5_K", dequant_q5_K_fp32_len, dequant_q5_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("dequant_q6_K", dequant_q6_K_fp32_len, dequant_q6_K_fp32_data, "main", 2, 4 * sizeof(int), {256 * 32, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
// get_rows
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16", get_rows_f16_fp32_len, get_rows_f16_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0", get_rows_q4_0_fp32_len, get_rows_q4_0_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("get_rows_q4_1", get_rows_q4_1_fp32_len, get_rows_q4_1_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("get_rows_q5_0", get_rows_q5_0_fp32_len, get_rows_q5_0_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("get_rows_q5_1", get_rows_q5_1_fp32_len, get_rows_q5_1_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("get_rows_q8_0", get_rows_q8_0_fp32_len, get_rows_q8_0_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("get_rows_f16_f32", get_rows_f16_f32_fp32_len, get_rows_f16_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("get_rows_q4_0_f32", get_rows_q4_0_f32_fp32_len, get_rows_q4_0_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("get_rows_q4_1_f32", get_rows_q4_1_f32_fp32_len, get_rows_q4_1_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("get_rows_q5_0_f32", get_rows_q5_0_f32_fp32_len, get_rows_q5_0_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("get_rows_q5_1_f32", get_rows_q5_1_f32_fp32_len, get_rows_q5_1_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_get_rows_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("get_rows_q8_0_f32", get_rows_q8_0_f32_fp32_len, get_rows_q8_0_f32_fp32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_F16] = ggml_vk_create_pipeline("mul_mat_vec_f16_f32", mul_mat_vec_f16_f32_len, mul_mat_vec_f16_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_0] = ggml_vk_create_pipeline("mul_mat_vec_q4_0_f32", mul_mat_vec_q4_0_f32_len, mul_mat_vec_q4_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_1] = ggml_vk_create_pipeline("mul_mat_vec_q4_1_f32", mul_mat_vec_q4_1_f32_len, mul_mat_vec_q4_1_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_0] = ggml_vk_create_pipeline("mul_mat_vec_q5_0_f32", mul_mat_vec_q5_0_f32_len, mul_mat_vec_q5_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_1] = ggml_vk_create_pipeline("mul_mat_vec_q5_1_f32", mul_mat_vec_q5_1_f32_len, mul_mat_vec_q5_1_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q8_0] = ggml_vk_create_pipeline("mul_mat_vec_q8_0_f32", mul_mat_vec_q8_0_f32_len, mul_mat_vec_q8_0_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q2_K] = ggml_vk_create_pipeline("mul_mat_vec_q2_K_f32", mul_mat_vec_q2_K_f32_len, mul_mat_vec_q2_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q3_K] = ggml_vk_create_pipeline("mul_mat_vec_q3_K_f32", mul_mat_vec_q3_K_f32_len, mul_mat_vec_q3_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q4_K] = ggml_vk_create_pipeline("mul_mat_vec_q4_K_f32", mul_mat_vec_q4_K_f32_len, mul_mat_vec_q4_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q5_K] = ggml_vk_create_pipeline("mul_mat_vec_q5_K_f32", mul_mat_vec_q5_K_f32_len, mul_mat_vec_q5_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_dequant_mul_mat_vec_f32[GGML_TYPE_Q6_K] = ggml_vk_create_pipeline("mul_mat_vec_q6_K_f32", mul_mat_vec_q6_K_f32_len, mul_mat_vec_q6_K_f32_data, "main", 3, 3 * sizeof(int), {1, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_matmul_split_k_reduce = ggml_vk_create_pipeline("split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_mul_mat_vec_p021_f16_f32 = ggml_vk_create_pipeline("mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_mul_mat_vec_nc_f16_f32 = ggml_vk_create_pipeline("mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_norm_f32 = ggml_vk_create_pipeline("norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_rms_norm_f32 = ggml_vk_create_pipeline("rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_cpy_f32_f32 = ggml_vk_create_pipeline("cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_cpy_f32_f16 = ggml_vk_create_pipeline("cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_cpy_f16_f16 = ggml_vk_create_pipeline("cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_cpy_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_add_f32 = ggml_vk_create_pipeline("add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_mul_f32 = ggml_vk_create_pipeline("mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_scale_f32 = ggml_vk_create_pipeline("scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_sqr_f32 = ggml_vk_create_pipeline("sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_clamp_f32 = ggml_vk_create_pipeline("clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_gelu_f32 = ggml_vk_create_pipeline("gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_silu_f32 = ggml_vk_create_pipeline("silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
vk_pipeline_relu_f32 = ggml_vk_create_pipeline("relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_diag_mask_inf_f32 = ggml_vk_create_pipeline("diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_soft_max_f32 = ggml_vk_create_pipeline("soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_rope_f32 = ggml_vk_create_pipeline("rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
|
|
|
|
vk_pipeline_rope_f16 = ggml_vk_create_pipeline("rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
|
|
|
|
|
|
|
|
vk_pipeline_rope_neox_f32 = ggml_vk_create_pipeline("rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
|
|
|
|
vk_pipeline_rope_neox_f16 = ggml_vk_create_pipeline("rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_init() {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_init()" << std::endl;
|
|
|
|
#endif
|
|
|
|
static bool initialized = false;
|
|
|
|
|
|
|
|
if (initialized) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
initialized = true;
|
|
|
|
|
|
|
|
const char* GGML_VULKAN_DEVICE = getenv("GGML_VULKAN_DEVICE");
|
|
|
|
int dev_num = (GGML_VULKAN_DEVICE == NULL ? 0 : atoi(GGML_VULKAN_DEVICE));
|
|
|
|
|
|
|
|
vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION };
|
|
|
|
const std::vector<const char*> layers = {
|
|
|
|
#ifdef VK_VALIDATE
|
|
|
|
"VK_LAYER_KHRONOS_validation",
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
const std::vector<const char*> extensions = {
|
|
|
|
#ifdef VK_VALIDATE
|
|
|
|
"VK_EXT_validation_features",
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags(), &app_info, layers, extensions);
|
|
|
|
#ifdef VK_VALIDATE
|
|
|
|
const std::vector<vk::ValidationFeatureEnableEXT> features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
|
|
|
|
vk::ValidationFeaturesEXT validation_features = {
|
|
|
|
features_enable,
|
|
|
|
{},
|
|
|
|
};
|
|
|
|
validation_features.setPNext(nullptr);
|
|
|
|
instance_create_info.setPNext(&validation_features);
|
|
|
|
|
|
|
|
std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_instance = vk::createInstance(instance_create_info);
|
|
|
|
|
|
|
|
vk_device.physical_device = vk_instance.enumeratePhysicalDevices()[dev_num];
|
|
|
|
std::vector<vk::ExtensionProperties> ext_props = vk_device.physical_device.enumerateDeviceExtensionProperties();
|
|
|
|
|
|
|
|
bool maintenance4_support = false;
|
|
|
|
|
|
|
|
// Check if maintenance4 is supported
|
|
|
|
for (auto properties : ext_props) {
|
|
|
|
if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
|
|
|
|
maintenance4_support = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
vk::PhysicalDeviceProperties2 props2;
|
|
|
|
vk::PhysicalDeviceMaintenance3Properties props3;
|
|
|
|
vk::PhysicalDeviceMaintenance4Properties props4;
|
|
|
|
vk::PhysicalDeviceSubgroupProperties subgroup_props;
|
|
|
|
props2.pNext = &props3;
|
|
|
|
props3.pNext = &subgroup_props;
|
|
|
|
if (maintenance4_support) {
|
|
|
|
subgroup_props.pNext = &props4;
|
|
|
|
}
|
|
|
|
vk_device.physical_device.getProperties2(&props2);
|
|
|
|
vk_device.properties = props2.properties;
|
|
|
|
|
|
|
|
if (maintenance4_support) {
|
|
|
|
vk_device.max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize);
|
|
|
|
} else {
|
|
|
|
vk_device.max_memory_allocation_size = props3.maxMemoryAllocationSize;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_device.vendor_id = vk_device.properties.vendorID;
|
|
|
|
vk_device.subgroup_size = subgroup_props.subgroupSize;
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_device.uma = vk_device.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
bool fp16_storage = false;
|
|
|
|
bool fp16_compute = false;
|
|
|
|
|
|
|
|
for (auto properties : ext_props) {
|
|
|
|
if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
|
|
|
|
fp16_storage = true;
|
|
|
|
} else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
|
|
|
|
fp16_compute = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
const char* GGML_VULKAN_DISABLE_F16 = getenv("GGML_VULKAN_DISABLE_F16");
|
|
|
|
bool force_disable_f16 = GGML_VULKAN_DISABLE_F16 != NULL;
|
|
|
|
|
|
|
|
vk_device.fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
|
|
|
|
|
|
|
|
std::vector<vk::QueueFamilyProperties> queue_family_props = vk_device.physical_device.getQueueFamilyProperties();
|
|
|
|
|
|
|
|
// Try to find a non-graphics compute queue and transfer-focused queues
|
|
|
|
const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1);
|
|
|
|
const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1);
|
|
|
|
|
|
|
|
const float priorities[] = { 1.0f, 1.0f };
|
|
|
|
const bool single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
|
|
|
|
|
|
|
|
std::vector<vk::DeviceQueueCreateInfo> device_queue_create_infos;
|
|
|
|
if (compute_queue_family_index != transfer_queue_family_index) {
|
|
|
|
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
|
|
|
|
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1});
|
|
|
|
} else if(!single_queue) {
|
|
|
|
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities});
|
|
|
|
} else {
|
|
|
|
device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
|
|
|
|
}
|
|
|
|
vk::DeviceCreateInfo device_create_info;
|
|
|
|
std::vector<const char *> device_extensions;
|
|
|
|
vk::PhysicalDeviceFeatures device_features = vk_device.physical_device.getFeatures();
|
|
|
|
|
|
|
|
VkPhysicalDeviceFeatures2 device_features2;
|
|
|
|
device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
|
|
|
|
device_features2.pNext = nullptr;
|
|
|
|
device_features2.features = (VkPhysicalDeviceFeatures)device_features;
|
|
|
|
|
|
|
|
VkPhysicalDeviceVulkan11Features vk11_features;
|
|
|
|
vk11_features.pNext = nullptr;
|
|
|
|
vk11_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES;
|
|
|
|
device_features2.pNext = &vk11_features;
|
|
|
|
|
|
|
|
VkPhysicalDeviceVulkan12Features vk12_features;
|
|
|
|
vk12_features.pNext = nullptr;
|
|
|
|
vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
|
|
|
|
vk11_features.pNext = &vk12_features;
|
|
|
|
|
|
|
|
vkGetPhysicalDeviceFeatures2(vk_device.physical_device, &device_features2);
|
|
|
|
|
|
|
|
vk_device.fp16 = vk_device.fp16 && vk12_features.shaderFloat16;
|
|
|
|
|
|
|
|
if (!vk11_features.storageBuffer16BitAccess) {
|
|
|
|
std::cerr << "ggml_vulkan: device does not support 16-bit storage" << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
device_extensions.push_back("VK_KHR_16bit_storage");
|
|
|
|
|
|
|
|
#ifdef VK_VALIDATE
|
|
|
|
device_extensions.push_back("VK_KHR_shader_non_semantic_info");
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if (vk_device.fp16) {
|
|
|
|
device_extensions.push_back("VK_KHR_shader_float16_int8");
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
std::cerr << "ggml_vulkan: Using " << vk_device.properties.deviceName << " | uma: " << vk_device.uma << " | fp16: " << vk_device.fp16 << " | warp size: " << vk_device.subgroup_size << std::endl;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
device_create_info = {
|
|
|
|
vk::DeviceCreateFlags(),
|
|
|
|
device_queue_create_infos,
|
|
|
|
{},
|
|
|
|
device_extensions
|
|
|
|
};
|
|
|
|
device_create_info.setPNext(&device_features2);
|
|
|
|
vk_device.device = vk_device.physical_device.createDevice(device_create_info);
|
|
|
|
|
|
|
|
vk_device.descriptor_set_mode = VK_DEVICE_DESCRIPTOR_POOL_MODE_UNKNOWN;
|
|
|
|
|
|
|
|
// Shaders
|
|
|
|
ggml_vk_load_shaders();
|
|
|
|
|
|
|
|
// Queues
|
|
|
|
vk_device.compute_queue = ggml_vk_create_queue(compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer });
|
|
|
|
if (!single_queue) {
|
|
|
|
const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
|
|
|
|
vk_device.transfer_queue = ggml_vk_create_queue(transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer });
|
|
|
|
} else {
|
|
|
|
vk_device.transfer_queue = vk_device.compute_queue;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_fence = vk_device.device.createFence({});
|
|
|
|
|
|
|
|
vk_ctx = nullptr;
|
|
|
|
|
|
|
|
vk_disable = false;
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
|
|
|
const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS");
|
|
|
|
vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks));
|
|
|
|
const char* output_tensor = getenv("GGML_VULKAN_OUTPUT_TENSOR");
|
|
|
|
vk_output_tensor = (output_tensor == NULL ? 0 : atoi(output_tensor));
|
|
|
|
#endif
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_pipeline* ggml_vk_get_to_fp16(ggml_type type) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_get_to_fp16()" << std::endl;
|
|
|
|
#endif
|
|
|
|
switch (type) {
|
|
|
|
case GGML_TYPE_F32:
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
return &vk_pipeline_dequant[type];
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_pipeline* ggml_vk_get_dequantize_mul_mat_vec(ggml_type type) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_get_dequantize_mul_mat_vec()" << std::endl;
|
|
|
|
#endif
|
|
|
|
switch (type) {
|
|
|
|
case GGML_TYPE_F16:
|
|
|
|
case GGML_TYPE_Q4_0:
|
|
|
|
case GGML_TYPE_Q4_1:
|
|
|
|
case GGML_TYPE_Q5_0:
|
|
|
|
case GGML_TYPE_Q5_1:
|
|
|
|
case GGML_TYPE_Q8_0:
|
|
|
|
case GGML_TYPE_Q2_K:
|
|
|
|
case GGML_TYPE_Q3_K:
|
|
|
|
case GGML_TYPE_Q4_K:
|
|
|
|
case GGML_TYPE_Q5_K:
|
|
|
|
case GGML_TYPE_Q6_K:
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
return &vk_pipeline_dequant_mul_mat_vec_f32[type];
|
|
|
|
}
|
|
|
|
|
|
|
|
// buffer pool for vulkan
|
|
|
|
#define MAX_VK_BUFFERS 256
|
|
|
|
|
|
|
|
static vk_buffer g_vk_buffer_pool[MAX_VK_BUFFERS];
|
|
|
|
|
|
|
|
static vk_buffer ggml_vk_pool_malloc(size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_pool_malloc(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
int best_i = -1;
|
|
|
|
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
|
|
|
|
int worst_i = -1;
|
|
|
|
size_t worst_size = 0; //largest unused buffer seen so far
|
|
|
|
for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
|
|
|
|
vk_buffer &b = g_vk_buffer_pool[i];
|
|
|
|
if (b.size > 0 && b.size >= size && b.size < best_size) {
|
|
|
|
best_i = i;
|
|
|
|
best_size = b.size;
|
|
|
|
}
|
|
|
|
if (b.size > 0 && b.size > worst_size) {
|
|
|
|
worst_i = i;
|
|
|
|
worst_size = b.size;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if(best_i != -1) {
|
|
|
|
//found the smallest buffer that fits our needs
|
|
|
|
vk_buffer b = g_vk_buffer_pool[best_i];
|
|
|
|
g_vk_buffer_pool[best_i].size = 0;
|
|
|
|
return b;
|
|
|
|
}
|
|
|
|
if(worst_i != -1) {
|
|
|
|
//no buffer that fits our needs, resize largest one to save memory
|
|
|
|
vk_buffer& b = g_vk_buffer_pool[worst_i];
|
|
|
|
ggml_vk_destroy_buffer(b);
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
return ggml_vk_create_buffer_check(size, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_pool_free(vk_buffer& buffer) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_pool_free(" << buffer.size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
|
|
|
|
vk_buffer& b = g_vk_buffer_pool[i];
|
|
|
|
if (b.size == 0) {
|
|
|
|
b = buffer;
|
|
|
|
// Set owning queue family index to ignored to avoid synchronization on next use
|
|
|
|
b.qf_owner = VK_QUEUE_FAMILY_IGNORED;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf(stderr, "WARNING: vk buffer pool full, increase MAX_VK_BUFFERS\n");
|
|
|
|
ggml_vk_destroy_buffer(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Returns an available temporary buffer that may only be used temporarily, it will be reused
|
|
|
|
static vk_buffer ggml_vk_create_buffer_temp(size_t size) {
|
|
|
|
// Try to find existing temp buffer with enough capacity
|
|
|
|
for (auto& buffer : vk_gc.temp_buffers) {
|
|
|
|
if (buffer.size >= size) {
|
|
|
|
return buffer;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Otherwise create new buffer
|
|
|
|
vk_buffer buf = ggml_vk_pool_malloc(size);
|
|
|
|
vk_gc.temp_buffers.push_back(buf);
|
|
|
|
|
|
|
|
return buf;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void * ggml_vk_host_malloc(size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_host_malloc(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_buffer buf = ggml_vk_create_buffer(size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
|
|
|
|
|
|
|
|
if(!(buf.memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
|
|
|
|
fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
|
|
|
|
size/1024.0/1024.0);
|
|
|
|
buf.size = 0;
|
|
|
|
vk_device.device.freeMemory(buf.device_memory);
|
|
|
|
vk_device.device.destroyBuffer(buf.buffer);
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_pinned_memory.push_back(std::make_tuple(buf.ptr, size, buf));
|
|
|
|
|
|
|
|
return buf.ptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_host_free(void* ptr) {
|
|
|
|
if (ptr == nullptr) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_host_free(" << ptr << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_buffer* buf = nullptr;
|
|
|
|
size_t index;
|
|
|
|
for (size_t i = 0; i < vk_pinned_memory.size(); i++) {
|
|
|
|
const uint8_t* addr = (const uint8_t*) std::get<0>(vk_pinned_memory[i]);
|
|
|
|
const uint8_t* endr = addr + std::get<1>(vk_pinned_memory[i]);
|
|
|
|
if (ptr >= addr && ptr < endr) {
|
|
|
|
buf = &std::get<2>(vk_pinned_memory[i]);
|
|
|
|
index = i;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (buf == nullptr) {
|
|
|
|
fprintf(stderr, "WARNING: failed to free pinned memory: memory not in map\n");
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_destroy_buffer(*buf);
|
|
|
|
|
|
|
|
vk_pinned_memory.erase(vk_pinned_memory.begin() + index);
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_host_get(const void * ptr, vk_buffer *& buf, size_t& buf_offset) {
|
|
|
|
buf = nullptr;
|
|
|
|
buf_offset = 0;
|
|
|
|
for (size_t i = 0; i < vk_pinned_memory.size(); i++) {
|
|
|
|
const uint8_t* addr = (const uint8_t*) std::get<0>(vk_pinned_memory[i]);
|
|
|
|
const uint8_t* endr = addr + std::get<1>(vk_pinned_memory[i]);
|
|
|
|
if (ptr >= addr && ptr < endr) {
|
|
|
|
buf = &std::get<2>(vk_pinned_memory[i]);
|
|
|
|
buf_offset = ((const uint8_t *)ptr) - addr;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
static vk_submission ggml_vk_begin_submission(vk_queue& q, bool one_time = true) {
|
|
|
|
vk_submission s;
|
|
|
|
s.buffer = ggml_vk_create_cmd_buffer(q);
|
|
|
|
if (one_time) {
|
|
|
|
s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit });
|
|
|
|
} else {
|
|
|
|
s.buffer.begin({ vk::CommandBufferUsageFlags{} });
|
|
|
|
}
|
|
|
|
|
|
|
|
return s;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_dispatch_pipeline(vk_context * ctx, vk_pipeline& pipeline, std::vector<vk_subbuffer>&& buffers, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
|
|
|
|
const uint32_t wg0 = CEIL_DIV(elements[0], pipeline.wg_denoms[0]);
|
|
|
|
const uint32_t wg1 = CEIL_DIV(elements[1], pipeline.wg_denoms[1]);
|
|
|
|
const uint32_t wg2 = CEIL_DIV(elements[2], pipeline.wg_denoms[2]);
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_dispatch_pipeline(" << pipeline.name << ", (" << wg0 << "," << wg1 << "," << wg2 << "))" << std::endl;
|
|
|
|
#endif
|
|
|
|
std::vector<vk::DescriptorBufferInfo> descriptor_buffer_infos;
|
|
|
|
std::vector<vk::WriteDescriptorSet> write_descriptor_sets;
|
|
|
|
GGML_ASSERT(pipeline.descriptor_set_idx < pipeline.descriptor_sets.size());
|
|
|
|
GGML_ASSERT(buffers.size() == pipeline.parameter_count);
|
|
|
|
vk::DescriptorSet& descriptor_set = pipeline.descriptor_sets[pipeline.descriptor_set_idx++];
|
|
|
|
for (uint32_t i = 0; i < pipeline.parameter_count; i++) {
|
|
|
|
descriptor_buffer_infos.push_back({buffers[i].buffer.buffer, buffers[i].offset, buffers[i].size});
|
|
|
|
}
|
|
|
|
for (uint32_t i = 0; i < pipeline.parameter_count; i++) {
|
|
|
|
write_descriptor_sets.push_back({descriptor_set, i, 0, 1, vk::DescriptorType::eStorageBuffer, nullptr, &descriptor_buffer_infos[i]});
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_device.device.updateDescriptorSets(write_descriptor_sets, {});
|
|
|
|
|
|
|
|
ctx->s->buffer.pushConstants(pipeline.layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
|
|
|
|
ctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
|
|
|
|
ctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
|
|
|
|
pipeline.layout,
|
|
|
|
0,
|
|
|
|
{ descriptor_set },
|
|
|
|
{});
|
|
|
|
ctx->s->buffer.dispatch(wg0, wg1, wg2);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_end_submission(vk_submission& s, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
|
|
|
|
s.buffer.end();
|
|
|
|
|
|
|
|
s.wait_semaphores = std::move(wait_semaphores);
|
|
|
|
s.signal_semaphores = std::move(signal_semaphores);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_ctx_end(vk_context * ctx) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_ctx_end(" << ctx << ", " << ctx->seqs.size() << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (ctx->s == nullptr) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ctx->s->buffer.end();
|
|
|
|
ctx->s = nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_ctx_begin(vk_context * ctx) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_ctx_begin(" << ctx << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (ctx->s != nullptr) {
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
ctx->seqs.push_back({ ggml_vk_begin_submission(*ctx->q) });
|
|
|
|
ctx->s = ctx->seqs[ctx->seqs.size() - 1].data();
|
|
|
|
}
|
|
|
|
|
|
|
|
static size_t ggml_vk_align_size(size_t width, size_t align) {
|
|
|
|
return CEIL_DIV(width, align) * align;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void deferred_memcpy(void * dst, const void * src, size_t size, std::vector<vk_staging_memcpy>* memcpys = nullptr) {
|
|
|
|
if (memcpys == nullptr) {
|
|
|
|
memcpy(dst, src, size);
|
|
|
|
} else {
|
|
|
|
memcpys->emplace_back(dst, src, size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ensure_sync_staging_buffer(size_t size) {
|
|
|
|
if (vk_sync_staging.size < size) {
|
|
|
|
ggml_vk_destroy_buffer(vk_sync_staging);
|
|
|
|
vk_sync_staging = ggml_vk_create_buffer_check(size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
static void ggml_vk_buffer_write_nc_async(vk_context * ctx, vk_buffer* dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_write_nc_async(" << tensor << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(!ggml_is_contiguous(tensor));
|
|
|
|
// Buffer is already mapped
|
|
|
|
if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
|
2024-01-30 12:59:30 +00:00
|
|
|
std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
// Check if src is pinned memory
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer * buf = nullptr;
|
|
|
|
size_t buf_offset;
|
|
|
|
ggml_vk_host_get(tensor->data, buf, buf_offset);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
const uint64_t ne0 = tensor->ne[0];
|
|
|
|
const uint64_t ne1 = tensor->ne[1];
|
|
|
|
const uint64_t ne2 = tensor->ne[2];
|
|
|
|
const uint64_t ne3 = tensor->ne[3];
|
|
|
|
const uint64_t nb0 = tensor->nb[0];
|
|
|
|
const uint64_t nb1 = tensor->nb[1];
|
|
|
|
const uint64_t nb2 = tensor->nb[2];
|
|
|
|
const uint64_t nb3 = tensor->nb[3];
|
|
|
|
const ggml_type type = tensor->type;
|
|
|
|
const uint64_t ts = ggml_type_size(type);
|
|
|
|
const uint64_t bs = ggml_blck_size(type);
|
|
|
|
|
|
|
|
const uint64_t dstnb0 = ts;
|
|
|
|
const uint64_t dstnb1 = dstnb0*(ne0/bs);
|
|
|
|
const uint64_t dstnb2 = dstnb1*ne1;
|
|
|
|
const uint64_t dstnb3 = dstnb2*ne2;
|
|
|
|
|
|
|
|
const uint64_t ne = ggml_nelements(tensor);
|
|
|
|
|
|
|
|
if (buf != nullptr) {
|
|
|
|
// Memory is pinned, use as staging buffer
|
|
|
|
std::vector<vk::BufferCopy> slices;
|
|
|
|
|
|
|
|
for (uint64_t i3 = 0; i3 < ne3; i3++) {
|
|
|
|
for (uint64_t i2 = 0; i2 < ne2; i2++) {
|
|
|
|
// Find longest contiguous slice
|
|
|
|
if (ne1*nb1 == dstnb2) {
|
|
|
|
slices.push_back({ buf_offset + i3*nb3 + i2*nb2, offset + i3*dstnb3 + i2*dstnb2, dstnb2 });
|
|
|
|
} else {
|
|
|
|
for (uint64_t i1 = 0; i1 < ne1; i1++) {
|
|
|
|
if (ne0*nb0/bs == dstnb1) {
|
|
|
|
slices.push_back({ buf_offset + i3*nb3 + i2*nb2 + i1*nb1, offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, dstnb1 });
|
|
|
|
} else {
|
|
|
|
const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
|
|
|
|
const uint64_t d_off = offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
|
|
|
|
for (uint64_t i0 = 0; i0 < ne0; i0++) {
|
|
|
|
slices.push_back({ s_off + i1*nb0, d_off + i0*dstnb0, dstnb0 });
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Staging buffer required
|
|
|
|
vk_buffer * staging = &vk_staging;
|
|
|
|
size_t staging_offset = vk_staging_offset;
|
|
|
|
const size_t copy_size = ts*ne/bs;
|
|
|
|
if (vk_staging.size < vk_staging_offset + copy_size) {
|
|
|
|
if (sync_staging) {
|
|
|
|
// Create temporary larger buffer
|
2024-01-30 12:59:30 +00:00
|
|
|
ensure_sync_staging_buffer(copy_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
staging = &vk_sync_staging;
|
|
|
|
staging_offset = 0;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
VkBufferCopy buf_copy{ staging_offset, offset, copy_size };
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
vkCmdCopyBuffer(ctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
|
|
|
|
|
|
|
|
for (uint64_t i3 = 0; i3 < ne3; i3++) {
|
|
|
|
for (uint64_t i2 = 0; i2 < ne2; i2++) {
|
|
|
|
// Find longest contiguous slice
|
|
|
|
if (ne1*nb1 == dstnb2) {
|
|
|
|
deferred_memcpy((uint8_t *)staging->ptr + staging_offset + i3*dstnb3 + i2*dstnb2, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2, dstnb2, &ctx->in_memcpys);
|
|
|
|
} else {
|
|
|
|
for (uint64_t i1 = 0; i1 < ne1; i1++) {
|
|
|
|
if (ne0*nb0/bs == dstnb1) {
|
|
|
|
deferred_memcpy((uint8_t *)staging->ptr + staging_offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2 + i1*nb1, dstnb1, &ctx->in_memcpys);
|
|
|
|
} else {
|
|
|
|
const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
|
|
|
|
const uint64_t d_off = staging_offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
|
|
|
|
for (uint64_t i0 = 0; i0 < ne0; i0++) {
|
|
|
|
deferred_memcpy((uint8_t *)staging->ptr + d_off + i0*dstnb0, (const uint8_t *) tensor->data + s_off + i0*nb0, dstnb0, &ctx->in_memcpys);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_write_2d_async(vk_context * ctx, vk_buffer* dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, bool sync_staging = false) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_write_2d_async(" << width << ", " << height << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
// Buffer is already mapped
|
|
|
|
if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
|
|
|
|
std::cerr << "ggml_vulkan: buffer_write_async dst buffer is host_visible. Use synchronous write." << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
// Check if src is pinned memory
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer * buf = nullptr;
|
|
|
|
size_t buf_offset;
|
|
|
|
ggml_vk_host_get(src, buf, buf_offset);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
if (buf != nullptr) {
|
|
|
|
// Memory is pinned, use as staging buffer
|
|
|
|
std::vector<vk::BufferCopy> slices(1);
|
|
|
|
if (width == spitch) {
|
|
|
|
// Only do single write if stride is equal
|
|
|
|
slices[0].srcOffset = buf_offset;
|
|
|
|
slices[0].dstOffset = offset;
|
|
|
|
slices[0].size = width * height;
|
|
|
|
} else {
|
|
|
|
slices.resize(height);
|
|
|
|
for (size_t i = 0; i < height; i++) {
|
|
|
|
slices[i].srcOffset = buf_offset + i * spitch;
|
|
|
|
slices[i].dstOffset = offset + i * width;
|
|
|
|
slices[i].size = width;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "STAGING" << std::endl;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Staging buffer required
|
|
|
|
vk_buffer * staging = &vk_staging;
|
|
|
|
size_t staging_offset = vk_staging_offset;
|
|
|
|
const size_t copy_size = width*height;
|
|
|
|
if (vk_staging.size < vk_staging_offset + copy_size) {
|
|
|
|
if (sync_staging) {
|
2024-01-30 12:59:30 +00:00
|
|
|
ensure_sync_staging_buffer(copy_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
staging = &vk_sync_staging;
|
|
|
|
staging_offset = 0;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
VkBufferCopy buf_copy = {
|
|
|
|
staging_offset,
|
|
|
|
offset,
|
|
|
|
copy_size};
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
vkCmdCopyBuffer(ctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
|
|
|
|
|
|
|
|
if (width == spitch) {
|
|
|
|
deferred_memcpy((uint8_t *)staging->ptr + staging_offset, src, width * height, &ctx->in_memcpys);
|
|
|
|
} else {
|
|
|
|
for (size_t i = 0; i < height; i++) {
|
|
|
|
deferred_memcpy((uint8_t *)staging->ptr + staging_offset + i * width, (const uint8_t *) src + i * spitch, width, &ctx->in_memcpys);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_write_async(vk_context * ctx, vk_buffer* dst, size_t offset, const void * src, size_t size, bool sync_staging = false) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_write_async(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
return ggml_vk_buffer_write_2d_async(ctx, dst, offset, src, size, size, 1, sync_staging);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_write_2d(vk_buffer* dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_write_2d(" << width << ", " << height << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
// Buffer is already mapped
|
|
|
|
if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
|
|
|
|
GGML_ASSERT(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
|
|
|
|
|
|
|
|
for (size_t i = 0; i < height; i++) {
|
|
|
|
memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
ggml_vk_buffer_write_2d_async(ctx, dst, offset, src, spitch, width, height, true);
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
|
|
|
|
for (auto& cpy : ctx->in_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_write(vk_buffer* dst, size_t offset, const void * src, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_write(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
ggml_vk_buffer_write_2d(dst, offset, src, 0, size, 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_read_2d_async(vk_context * ctx, vk_buffer* src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(width > 0);
|
|
|
|
GGML_ASSERT(height > 0);
|
|
|
|
GGML_ASSERT(src->size > 0);
|
|
|
|
// Check if dst is pinned memory
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer * buf = nullptr;
|
|
|
|
size_t buf_offset;
|
|
|
|
ggml_vk_host_get(dst, buf, buf_offset);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
std::vector<vk::BufferCopy> slices(1);
|
|
|
|
if (width == spitch && width == dpitch) {
|
|
|
|
// Only do single write if stride is equal
|
|
|
|
slices[0].srcOffset = offset;
|
|
|
|
slices[0].dstOffset = buf_offset;
|
|
|
|
slices[0].size = width * height;
|
|
|
|
} else {
|
|
|
|
slices.resize(height);
|
|
|
|
for (size_t i = 0; i < height; i++) {
|
|
|
|
slices[i].srcOffset = offset + i * spitch;
|
|
|
|
slices[i].dstOffset = buf_offset + i * dpitch;
|
|
|
|
slices[i].size = width;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (buf != nullptr) {
|
|
|
|
// Memory is pinned, use as staging buffer
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.copyBuffer(src->buffer, buf->buffer, slices);
|
|
|
|
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "STAGING" << std::endl;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
// Fall back to staging buffer
|
|
|
|
vk_buffer * staging = &vk_staging;
|
|
|
|
const size_t copy_size = dpitch * height;
|
|
|
|
if (vk_staging.size < vk_staging_offset + copy_size) {
|
|
|
|
if (sync_staging) {
|
|
|
|
// Create temporary larger buffer
|
2024-01-30 12:59:30 +00:00
|
|
|
ensure_sync_staging_buffer(copy_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
staging = &vk_sync_staging;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.copyBuffer(src->buffer, staging->buffer, slices);
|
|
|
|
|
|
|
|
deferred_memcpy(dst, staging->ptr, copy_size, &ctx->out_memcpys);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_read_async(vk_context * ctx, vk_buffer* src, size_t offset, void * dst, size_t size, bool sync_staging = false) {
|
|
|
|
return ggml_vk_buffer_read_2d_async(ctx, src, offset, dst, size, size, size, 1, sync_staging);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_read(vk_buffer* src, size_t offset, void * dst, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_read(" << offset << ", " << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
|
|
|
|
GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
|
|
|
|
|
|
|
|
memcpy(dst, (uint8_t *) src->ptr + offset, size);
|
|
|
|
} else {
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
ggml_vk_buffer_read_async(ctx, src, offset, dst, size, true);
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "vk_buffer_read waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
for (auto& cpy : ctx->out_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_copy_async(vk_context * ctx, vk_buffer * dst, size_t dst_offset, vk_buffer * src, size_t src_offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_copy_async(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
VkBufferCopy bc{ src_offset, dst_offset, size };
|
|
|
|
|
|
|
|
vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_copy(vk_buffer * dst, size_t dst_offset, vk_buffer * src, size_t src_offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_copy(" << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
VkBufferCopy bc{ src_offset, dst_offset, size };
|
|
|
|
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
|
|
|
|
ggml_vk_buffer_copy_async(ctx, dst, dst_offset, src, src_offset, size);
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_buffer_memset(vk_buffer* dst, size_t offset, uint32_t c, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
ctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "vk_memset waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_h2d_tensor_2d(vk_context * ctx, vk_buffer * dst, size_t offset, const ggml_tensor * src, uint64_t i3, uint64_t i2, uint64_t i1) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_h2d_tensor_2d(dst=" << dst << ", offset=" << offset << ", src=" << src << ", i3=" << i3 << ", i2=" << i2 << ", i1=" << i1 << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
const uint64_t ne0 = src->ne[0];
|
|
|
|
const uint64_t ne1 = src->ne[1];
|
|
|
|
const uint64_t nb0 = src->nb[0];
|
|
|
|
const uint64_t nb1 = src->nb[1];
|
|
|
|
const uint64_t nb2 = src->nb[2];
|
|
|
|
const uint64_t nb3 = src->nb[3];
|
|
|
|
const enum ggml_type type = src->type;
|
|
|
|
const size_t ts = ggml_type_size(type);
|
|
|
|
const size_t bs = ggml_blck_size(type);
|
|
|
|
const size_t row_length = ts*ne0/bs;
|
|
|
|
|
|
|
|
const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
|
|
|
|
if (nb0 == ts && nb1 == row_length) {
|
|
|
|
return ggml_vk_buffer_write_async(ctx, dst, offset, x, i1*nb1);
|
|
|
|
}
|
|
|
|
if (nb0 == ts && (i1 == ne1 || !ggml_is_permuted(src))) {
|
|
|
|
return ggml_vk_buffer_write_2d_async(ctx, dst, offset, x, nb1, row_length, i1);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_ASSERT(i3 == 0);
|
|
|
|
GGML_ASSERT(i2 == 0);
|
|
|
|
GGML_ASSERT(i1 == (uint64_t) ggml_nrows(src));
|
|
|
|
|
|
|
|
return ggml_vk_buffer_write_nc_async(ctx, dst, offset, src);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_d2h_tensor_2d(vk_context * ctx, vk_buffer * src, size_t offset, const ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_d2h_tensor_2d()" << std::endl;
|
|
|
|
#endif
|
|
|
|
const uint64_t ne0 = dst->ne[0];
|
|
|
|
const uint64_t ne1 = dst->ne[1];
|
|
|
|
const uint64_t ne2 = dst->ne[2];
|
|
|
|
const uint64_t ne3 = dst->ne[3];
|
|
|
|
const uint64_t nb0 = dst->nb[0];
|
|
|
|
const uint64_t nb1 = dst->nb[1];
|
|
|
|
// const uint64_t nb2 = dst->nb[2];
|
|
|
|
// const uint64_t nb3 = dst->nb[3];
|
|
|
|
const enum ggml_type type = dst->type;
|
|
|
|
const size_t ts = ggml_type_size(type);
|
|
|
|
const size_t bs = ggml_blck_size(type);
|
|
|
|
const size_t row_length = ts*ne0/bs;
|
|
|
|
|
|
|
|
if (ggml_is_contiguous(dst)) {
|
|
|
|
return ggml_vk_buffer_read_async(ctx, src, offset, dst->data, ne1*nb1*ne2*ne3);
|
|
|
|
}
|
|
|
|
if (nb0 == ts) {
|
|
|
|
return ggml_vk_buffer_read_2d_async(ctx, src, offset, dst->data, nb1, nb1, row_length, ne1*ne2*ne3);
|
|
|
|
}
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t ggml_vk_guess_split_k(int m, int n, int k) {
|
|
|
|
#ifdef VK_DEBUG
|
2024-01-30 12:59:30 +00:00
|
|
|
std::cerr << "ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")";
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
#endif
|
|
|
|
if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " = 4" << std::endl;
|
|
|
|
#endif
|
|
|
|
return 4;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " = 1" << std::endl;
|
|
|
|
#endif
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static uint32_t ggml_vk_guess_matmul_pipeline_align(int m, int n) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (m <= 32 || n <= 32) {
|
|
|
|
return vk_pipeline_matmul_f32_aligned_s.align;
|
|
|
|
}
|
|
|
|
if (vk_device.subgroup_size == 64 || m <= 64 || n <= 64) {
|
|
|
|
return vk_pipeline_matmul_f32_aligned_m.align;
|
|
|
|
}
|
|
|
|
return vk_pipeline_matmul_f32_aligned_l.align;
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_pipeline* ggml_vk_guess_matmul_pipeline(bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
|
|
|
|
#endif
|
|
|
|
if (bit16_x && bit16_y) {
|
|
|
|
if (m <= 32 || n <= 32) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " S" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_aligned_s : &vk_pipeline_matmul_f16_s;
|
|
|
|
}
|
|
|
|
if (vk_device.subgroup_size == 64 || m <= 64 || n <= 64) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " M" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_aligned_m : &vk_pipeline_matmul_f16_m;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " L" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_aligned_l : &vk_pipeline_matmul_f16_l;
|
|
|
|
}
|
|
|
|
if (bit16_x && !bit16_y) {
|
|
|
|
if (m <= 32 || n <= 32) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " S" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_f32_aligned_s : &vk_pipeline_matmul_f16_f32_s;
|
|
|
|
}
|
|
|
|
if (vk_device.subgroup_size == 64 || m <= 64 || n <= 64) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " M" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_f32_aligned_m : &vk_pipeline_matmul_f16_f32_m;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " L" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f16_f32_aligned_l : &vk_pipeline_matmul_f16_f32_l;
|
|
|
|
}
|
|
|
|
if (!bit16_x && bit16_y) {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (m <= 32 || n <= 32) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " S" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f32_aligned_s : &vk_pipeline_matmul_f32_s;
|
|
|
|
}
|
|
|
|
if (vk_device.subgroup_size == 64 || m <= 64 || n <= 64) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " M" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f32_aligned_m : &vk_pipeline_matmul_f32_m;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << " L" << std::endl;
|
|
|
|
#endif
|
|
|
|
return aligned ? &vk_pipeline_matmul_f32_aligned_l : &vk_pipeline_matmul_f32_l;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_matmul(vk_context * ctx, vk_pipeline& pipeline, vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_matmul(a: (" << a.buffer.buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer.buffer << ", " << b.offset << ", " << b.size << "), c: (" << d.buffer.buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << split_k_buffer.buffer.buffer << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (split_k == 1) {
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
const std::array<uint32_t, 14> pc = { m, n, k, stride_a, stride_b, stride_d, k, ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d };
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, pipeline, { a, b, d }, pc.size() * sizeof(uint32_t), pc.data(), { m, n, batch });
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_ASSERT(batch_stride_d == m * n);
|
|
|
|
|
|
|
|
// Synchronize the two submissions
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.fillBuffer(split_k_buffer.buffer.buffer, 0, split_k_buffer.size, 0);
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
const std::array<uint32_t, 14> pc1 = { m, n, k, stride_a, stride_b, stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, batch_stride_a, batch_stride_b, batch_stride_d };
|
|
|
|
// Make sure enough workgroups get assigned for split k to work
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, pipeline, { a, b, split_k_buffer }, pc1.size() * sizeof(uint32_t), pc1.data(), { (CEIL_DIV(m, pipeline.wg_denoms[0]) * pipeline.wg_denoms[0]) * split_k, n, batch });
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, vk_pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
|
|
|
|
return
|
|
|
|
tensor->nb[0] == ggml_type_size(tensor->type) &&
|
|
|
|
tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
|
|
|
|
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
|
|
|
|
}
|
|
|
|
|
|
|
|
static vk_pipeline * ggml_vk_get_cpy_pipeline(ggml_type from, ggml_type to) {
|
|
|
|
if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_cpy_f32_f32;
|
|
|
|
}
|
|
|
|
if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
|
|
|
|
return &vk_pipeline_cpy_f32_f16;
|
|
|
|
}
|
|
|
|
if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
|
|
|
|
return &vk_pipeline_cpy_f16_f16;
|
|
|
|
}
|
|
|
|
|
|
|
|
std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_cpy_to_contiguous(vk_context * ctx, vk_pipeline * pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out, ggml_type buffer_type, bool aligned=true) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
|
|
|
|
std::cerr << "buffer in size=" << in.buffer.size << ", buffer out size=" << out.buffer.size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
const int tensor_type_size = ggml_type_size(tensor->type);
|
|
|
|
const int dst_type_size = ggml_type_size(buffer_type);
|
|
|
|
|
|
|
|
const uint32_t ne = tensor->ne[0] * tensor->ne[1] * tensor->ne[2];
|
|
|
|
|
|
|
|
const uint32_t nb2 = aligned ? ggml_vk_align_size(dst_type_size * tensor->ne[0] * tensor->ne[1], vk_device.properties.limits.minStorageBufferOffsetAlignment) / dst_type_size : tensor->ne[0] * tensor->ne[1];
|
|
|
|
|
|
|
|
const vk_op_cpy_push_constants pc = {
|
|
|
|
(uint32_t)ne,
|
|
|
|
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size,
|
|
|
|
(uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], 1 , (uint32_t)tensor->ne[0] , nb2,
|
|
|
|
0,
|
|
|
|
};
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { in, out }, sizeof(vk_op_cpy_push_constants), &pc, { ne, 1, 1 });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul_mat_q_f16(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
|
|
|
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
|
|
|
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT
|
|
|
|
GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT
|
|
|
|
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
const uint64_t ne03 = src0->ne[3];
|
|
|
|
|
|
|
|
const uint64_t ne10 = src1->ne[0];
|
|
|
|
const uint64_t ne11 = src1->ne[1];
|
|
|
|
const uint64_t ne12 = src1->ne[2];
|
|
|
|
const uint64_t ne13 = src1->ne[3];
|
|
|
|
|
|
|
|
const uint64_t ne20 = dst->ne[0];
|
|
|
|
const uint64_t ne21 = dst->ne[1];
|
|
|
|
|
|
|
|
const uint64_t r2 = ne12 / ne02;
|
|
|
|
const uint64_t r3 = ne13 / ne03;
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
|
|
|
|
vk_buffer * d_Qx = nullptr;
|
|
|
|
size_t qx_buf_offset = 0;
|
|
|
|
vk_buffer * d_Qy = nullptr;
|
|
|
|
size_t qy_buf_offset = 0;
|
|
|
|
|
|
|
|
bool src0_uma = false;
|
|
|
|
bool src1_uma = false;
|
|
|
|
|
|
|
|
if (vk_device.uma) {
|
|
|
|
ggml_vk_host_get(src0->data, d_Qx, qx_buf_offset);
|
|
|
|
ggml_vk_host_get(src1->data, d_Qy, qy_buf_offset);
|
|
|
|
src0_uma = d_Qx != nullptr;
|
|
|
|
src1_uma = d_Qy != nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
|
|
|
|
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
|
|
|
|
const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
|
|
|
|
|
|
|
|
const bool f16_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
|
|
|
|
|
|
|
|
const bool qx_needs_dequant = src0->type != GGML_TYPE_F16 || x_non_contig;
|
|
|
|
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
|
|
|
|
|
|
|
|
// Not implemented
|
|
|
|
GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT
|
|
|
|
|
|
|
|
const int x_ne = ne01 * ne00;
|
|
|
|
const int y_ne = ne11 * ne10;
|
|
|
|
const int d_ne = ne11 * ne01;
|
|
|
|
|
|
|
|
const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ne01, ne11));
|
|
|
|
const bool aligned = ne10 == kpad;
|
|
|
|
|
|
|
|
const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
|
|
|
|
|
|
|
|
vk_pipeline * pipeline = ggml_vk_guess_matmul_pipeline(true, !f16_f32_kernel, ne01, ne11, aligned);
|
|
|
|
|
|
|
|
const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
|
|
|
|
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
|
|
|
|
const uint64_t x_sz = sizeof(ggml_fp16_t) * x_ne;
|
|
|
|
const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
|
|
|
|
const uint64_t d_sz = sizeof(float) * d_ne;
|
|
|
|
|
|
|
|
vk_buffer* d_D = &extra->buffer_gpu;
|
|
|
|
const uint64_t d_buf_offset = extra->offset;
|
|
|
|
GGML_ASSERT(d_D != nullptr);
|
|
|
|
GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
|
|
|
|
vk_buffer* d_X;
|
|
|
|
uint64_t x_buf_offset = 0;
|
|
|
|
vk_buffer* d_Y;
|
|
|
|
uint64_t y_buf_offset = 0;
|
|
|
|
if (load_x) {
|
|
|
|
d_Qx = &vk_prealloc_qx;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if (!src0_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Qx = &extra_src0->buffer_gpu;
|
|
|
|
qx_buf_offset = extra_src0->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
}
|
|
|
|
if (load_y) {
|
|
|
|
d_Qy = &vk_prealloc_qy;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if (!src1_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Qy = &extra_src1->buffer_gpu;
|
|
|
|
qy_buf_offset = extra_src1->offset;
|
|
|
|
GGML_ASSERT(d_Qy != nullptr);
|
|
|
|
}
|
|
|
|
if (qx_needs_dequant) {
|
|
|
|
d_X = &vk_prealloc_x;
|
|
|
|
GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
|
|
|
|
} else {
|
|
|
|
d_X = d_Qx;
|
|
|
|
x_buf_offset = qx_buf_offset;
|
|
|
|
GGML_ASSERT(qx_sz == x_sz); // NOLINT
|
|
|
|
}
|
|
|
|
if (qy_needs_dequant) {
|
|
|
|
d_Y = &vk_prealloc_y;
|
|
|
|
GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
|
|
|
|
} else {
|
|
|
|
d_Y = d_Qy;
|
|
|
|
y_buf_offset = qy_buf_offset;
|
|
|
|
GGML_ASSERT(qy_sz == y_sz);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_pipeline * to_fp16_vk_0 = nullptr;
|
|
|
|
vk_pipeline * to_fp16_vk_1 = nullptr;
|
|
|
|
|
|
|
|
if (x_non_contig) {
|
|
|
|
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(src0->type, GGML_TYPE_F16);
|
|
|
|
} else {
|
|
|
|
to_fp16_vk_0 = ggml_vk_get_to_fp16(src0->type);
|
|
|
|
}
|
|
|
|
if (y_non_contig) {
|
|
|
|
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(src1->type, GGML_TYPE_F16);
|
|
|
|
} else {
|
|
|
|
to_fp16_vk_1 = ggml_vk_get_to_fp16(src1->type);
|
|
|
|
}
|
|
|
|
GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT
|
|
|
|
GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT
|
|
|
|
|
|
|
|
// Allocate descriptor sets
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*pipeline, ne12 * ne13);
|
|
|
|
if (qx_needs_dequant) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*to_fp16_vk_0, x_non_contig ? 1 : ne12 * ne13);
|
|
|
|
}
|
|
|
|
if (qy_needs_dequant) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13);
|
|
|
|
}
|
|
|
|
if (split_k > 1) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(vk_pipeline_matmul_split_k_reduce, ne12 * ne13);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (x_non_contig) {
|
|
|
|
ggml_vk_cpy_to_contiguous(ctx, to_fp16_vk_0, src0, { *d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { *d_X, 0, VK_WHOLE_SIZE }, dst->type, false);
|
|
|
|
} else if (load_x || qx_needs_dequant) {
|
|
|
|
if (load_x) {
|
|
|
|
// copy data to device
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qx, 0, src0, 0, 0, ggml_nrows(src0));
|
|
|
|
vk_staging_offset = qx_sz * ne02 * ne03;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (qx_needs_dequant) {
|
|
|
|
const std::vector<int> pc = { (int)ne01, (int)ne10, (int)ne10, (int)ne10 };
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *to_fp16_vk_0, { { *d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, { *d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (y_non_contig) {
|
|
|
|
ggml_vk_cpy_to_contiguous(ctx, to_fp16_vk_1, src1, { *d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { *d_Y, 0, VK_WHOLE_SIZE }, dst->type);
|
|
|
|
} else if (load_y) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qy, 0, src1, 0, 0, ggml_nrows(src1));
|
|
|
|
}
|
|
|
|
|
|
|
|
uint32_t stride_batch_x = ne00*ne01;
|
|
|
|
uint32_t stride_batch_y = ne10*ne11;
|
|
|
|
|
|
|
|
if (!ggml_vk_dim01_contiguous(src0) && !load_x && !qx_needs_dequant) {
|
|
|
|
stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!ggml_vk_dim01_contiguous(src1) && !load_y && !qy_needs_dequant) {
|
|
|
|
stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
|
|
|
|
}
|
|
|
|
|
|
|
|
// compute
|
|
|
|
ggml_vk_matmul(ctx, *pipeline, { *d_X, x_buf_offset, x_sz * ne02 * ne03 }, { *d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { *d_D, d_buf_offset, d_sz * ne12 * ne13 }, { vk_prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21); // NOLINT
|
|
|
|
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
float * d = (float *) ((char *) dst->data);
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, 0, d, sizeof(float) * d_ne * ne12 * ne13);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul_mat_vec_q_f16(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
|
|
|
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
|
|
|
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); // NOLINT
|
|
|
|
GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); // NOLINT
|
|
|
|
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
const uint64_t ne03 = src0->ne[3];
|
|
|
|
|
|
|
|
const uint64_t ne10 = src1->ne[0];
|
|
|
|
const uint64_t ne11 = src1->ne[1];
|
|
|
|
const uint64_t ne12 = src1->ne[2];
|
|
|
|
const uint64_t ne13 = src1->ne[3];
|
|
|
|
|
|
|
|
GGML_ASSERT(ne11 == 1);
|
|
|
|
|
|
|
|
const uint64_t nb2 = dst->nb[2];
|
|
|
|
const uint64_t nb3 = dst->nb[3];
|
|
|
|
|
|
|
|
const uint64_t r2 = ne12 / ne02;
|
|
|
|
const uint64_t r3 = ne13 / ne03;
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
|
|
|
|
vk_buffer * d_Qx = nullptr;
|
|
|
|
size_t qx_buf_offset = 0;
|
|
|
|
vk_buffer * d_Qy = nullptr;
|
|
|
|
size_t qy_buf_offset = 0;
|
|
|
|
|
|
|
|
bool src0_uma = false;
|
|
|
|
bool src1_uma = false;
|
|
|
|
|
|
|
|
if (vk_device.uma) {
|
|
|
|
ggml_vk_host_get(src0->data, d_Qx, qx_buf_offset);
|
|
|
|
ggml_vk_host_get(src1->data, d_Qy, qy_buf_offset);
|
|
|
|
src0_uma = d_Qx != nullptr;
|
|
|
|
src1_uma = d_Qy != nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
|
|
|
|
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
|
|
|
|
const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
|
|
|
|
|
|
|
|
const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
|
|
|
|
|
|
|
|
const bool qx_needs_dequant = x_non_contig;
|
|
|
|
const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
|
|
|
|
|
|
|
|
const uint64_t x_ne = ne01 * ne00;
|
|
|
|
const uint64_t y_ne = ne11 * ne10;
|
|
|
|
const uint64_t d_ne = ne11 * ne01;
|
|
|
|
|
|
|
|
const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), vk_device.properties.limits.minStorageBufferOffsetAlignment);
|
|
|
|
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
|
|
|
|
const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, vk_device.properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
|
|
|
|
const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
|
|
|
|
const uint64_t d_sz = sizeof(float) * d_ne;
|
|
|
|
|
|
|
|
vk_buffer* d_D = &extra->buffer_gpu;
|
|
|
|
const uint64_t d_buf_offset = extra->offset;
|
|
|
|
GGML_ASSERT(d_D != nullptr);
|
|
|
|
vk_buffer* d_X;
|
|
|
|
uint64_t x_buf_offset = 0;
|
|
|
|
vk_buffer* d_Y;
|
|
|
|
uint64_t y_buf_offset = 0;
|
|
|
|
if (load_x) {
|
|
|
|
d_Qx = &vk_prealloc_qx;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if(!src1_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Qx = &extra_src0->buffer_gpu;
|
|
|
|
qx_buf_offset = extra_src0->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
}
|
|
|
|
if (load_y) {
|
|
|
|
d_Qy = &vk_prealloc_qy;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if(!src1_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Qy = &extra_src1->buffer_gpu;
|
|
|
|
qy_buf_offset = extra_src1->offset;
|
|
|
|
GGML_ASSERT(d_Qy != nullptr);
|
|
|
|
}
|
|
|
|
if (qx_needs_dequant) {
|
|
|
|
d_X = &vk_prealloc_x;
|
|
|
|
} else {
|
|
|
|
d_X = d_Qx;
|
|
|
|
x_buf_offset = qx_buf_offset;
|
|
|
|
GGML_ASSERT(qx_sz == x_sz);
|
|
|
|
}
|
|
|
|
if (qy_needs_dequant) {
|
|
|
|
d_Y = &vk_prealloc_y;
|
|
|
|
} else {
|
|
|
|
d_Y = d_Qy;
|
|
|
|
y_buf_offset = qy_buf_offset;
|
|
|
|
GGML_ASSERT(qy_sz == y_sz);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_pipeline * to_fp16_vk_0 = nullptr;
|
|
|
|
vk_pipeline* to_fp16_vk_1 = nullptr;
|
|
|
|
if (x_non_contig) {
|
|
|
|
to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(src0->type, src0->type);
|
|
|
|
}
|
|
|
|
if (y_non_contig) {
|
|
|
|
to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(src1->type, src1->type);
|
|
|
|
} else {
|
|
|
|
to_fp16_vk_1 = ggml_vk_get_to_fp16(src1->type);
|
|
|
|
}
|
|
|
|
vk_pipeline* dmmv = ggml_vk_get_dequantize_mul_mat_vec(src0->type);
|
|
|
|
GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT
|
|
|
|
GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT
|
|
|
|
GGML_ASSERT(dmmv != nullptr);
|
|
|
|
|
|
|
|
// Allocate descriptor sets
|
|
|
|
if (qx_needs_dequant) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*to_fp16_vk_0, 1);
|
|
|
|
}
|
|
|
|
if (qy_needs_dequant) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*to_fp16_vk_1, y_non_contig ? 1 : ne12 * ne13);
|
|
|
|
}
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*dmmv, ne12 * ne13);
|
|
|
|
|
|
|
|
if (x_non_contig) {
|
|
|
|
GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, vk_device.properties.limits.minStorageBufferOffsetAlignment));
|
|
|
|
ggml_vk_cpy_to_contiguous(ctx, to_fp16_vk_0, src0, { *d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { *d_X, 0, VK_WHOLE_SIZE }, src0->type);
|
|
|
|
} else if (load_x) {
|
|
|
|
// copy data to device
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qx, 0, src0, 0, 0, ggml_nrows(src0));
|
|
|
|
}
|
|
|
|
if (y_non_contig) {
|
|
|
|
GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
|
|
|
|
ggml_vk_cpy_to_contiguous(ctx, to_fp16_vk_1, src1, { *d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { *d_Y, 0, VK_WHOLE_SIZE }, src1->type);
|
|
|
|
} else if (load_y) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qy, 0, src1, 0, 0, ggml_nrows(src1));
|
|
|
|
}
|
|
|
|
|
|
|
|
for (uint64_t i13 = 0; i13 < ne13; i13++) {
|
|
|
|
const uint64_t i03 = i13 / r3;
|
|
|
|
for (uint64_t i12 = 0; i12 < ne12; i12++) {
|
|
|
|
const uint64_t i02 = i12 / r2;
|
|
|
|
|
|
|
|
const uint64_t it_idx0 = (i03 * ne02 + i02);
|
|
|
|
const uint64_t it_idx1 = (i13 * ne12 + i12);
|
|
|
|
const uint64_t x_offset = x_buf_offset + x_sz * it_idx0;
|
|
|
|
const uint64_t qy_offset = qy_buf_offset + qy_sz * it_idx1;
|
|
|
|
const uint64_t y_offset = y_buf_offset + y_sz * it_idx1;
|
|
|
|
const uint64_t d_offset = d_buf_offset + d_sz * it_idx1;
|
|
|
|
|
|
|
|
const uint64_t y_buffer_offset = (y_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t y_shader_offset = y_offset - y_buffer_offset;
|
|
|
|
|
|
|
|
const uint64_t d_buffer_offset = (d_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t d_shader_offset = d_offset - d_buffer_offset;
|
|
|
|
|
|
|
|
if (!y_non_contig && qy_needs_dequant) {
|
|
|
|
const std::vector<int> pc = { (int)ne11, (int)ne10, (int)ne10, (int)ne10 };
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *to_fp16_vk_1, { { *d_Qy, qy_offset, qy_sz }, { *d_Y, y_offset, y_sz } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)y_ne, 1, 1});
|
|
|
|
}
|
|
|
|
|
|
|
|
// compute
|
|
|
|
const std::array<int, 3> pc = { (int)ne00, (int)(y_shader_offset / ggml_type_size(src1->type)), (int)(d_shader_offset / ggml_type_size(dst->type))};
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *dmmv, { { *d_X, x_offset, x_sz }, { *d_Y, y_buffer_offset, y_sz + y_shader_offset }, { *d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1});
|
|
|
|
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, d_offset, d, sizeof(float) * d_ne);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul_mat_vec_p021_f16_f32(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_mul_mat_p021_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
|
|
|
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
|
|
|
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
|
|
|
|
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
|
|
|
|
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // NOLINT
|
|
|
|
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // NOLINT
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
// const uint64_t ne03 = src0->ne[3];
|
|
|
|
|
|
|
|
const uint64_t ne10 = src1->ne[0];
|
|
|
|
const uint64_t ne11 = src1->ne[1];
|
|
|
|
const uint64_t ne12 = src1->ne[2];
|
|
|
|
// const uint64_t ne13 = src1->ne[3];
|
|
|
|
|
|
|
|
GGML_ASSERT(ne11 == 1);
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
|
|
|
|
vk_buffer * d_Qy = nullptr;
|
|
|
|
size_t qy_buf_offset = 0;
|
|
|
|
|
|
|
|
bool src1_uma = false;
|
|
|
|
|
|
|
|
if (vk_device.uma) {
|
|
|
|
ggml_vk_host_get(src1->data, d_Qy, qy_buf_offset);
|
|
|
|
src1_uma = d_Qy != nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
const uint64_t x_ne = ne00 * ne01 * ne02;
|
|
|
|
const uint64_t y_ne = ne10 * ne11 * ne12;
|
|
|
|
const uint64_t d_ne = ne01 * ne11 * ne12;
|
|
|
|
|
|
|
|
const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), vk_device.properties.limits.minStorageBufferOffsetAlignment);
|
|
|
|
const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
|
|
|
|
const uint64_t d_sz = sizeof(float) * d_ne;
|
|
|
|
|
|
|
|
vk_buffer* d_D = &extra->buffer_gpu;
|
|
|
|
const uint64_t d_buf_offset = extra->offset;
|
|
|
|
GGML_ASSERT(d_D != nullptr);
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer* d_Qx = &extra_src0->buffer_gpu;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
const uint64_t qx_buf_offset = extra_src0->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
if (load_y) {
|
|
|
|
d_Qy = &vk_prealloc_qy;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if (!src1_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Qy = &extra_src1->buffer_gpu;
|
|
|
|
qy_buf_offset = extra_src1->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Allocate descriptor sets
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(vk_pipeline_mul_mat_vec_p021_f16_f32, 1);
|
|
|
|
|
|
|
|
const uint64_t qy_buffer_offset = (qy_buf_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
|
|
|
|
|
|
|
|
const uint64_t d_buffer_offset = (d_buf_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
|
|
|
|
|
|
|
|
if (load_y) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qy, qy_buf_offset, src1, 0, 0, ggml_nrows(src1));
|
|
|
|
}
|
|
|
|
|
|
|
|
// compute
|
|
|
|
const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, vk_pipeline_mul_mat_vec_p021_f16_f32, { { *d_Qx, qx_buf_offset, qx_sz }, { *d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { *d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
|
|
|
|
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
float * d = (float *) dst->data;
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul_mat_vec_nc_f16_f32(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
|
|
|
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
|
|
|
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src0));
|
|
|
|
GGML_ASSERT(!ggml_is_transposed(src1));
|
|
|
|
GGML_ASSERT(!ggml_is_permuted(src0));
|
|
|
|
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
|
|
|
|
GGML_ASSERT(src0->type == GGML_TYPE_F16);
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_F32);
|
|
|
|
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
// const uint64_t ne03 = src0->ne[3];
|
|
|
|
|
|
|
|
const uint64_t nb01 = src0->nb[1];
|
|
|
|
const uint64_t nb02 = src0->nb[2];
|
|
|
|
|
|
|
|
// const uint64_t ne10 = src1->ne[0];
|
|
|
|
const uint64_t ne11 = src1->ne[1];
|
|
|
|
const uint64_t ne12 = src1->ne[2];
|
|
|
|
// const uint64_t ne13 = src1->ne[3];
|
|
|
|
|
|
|
|
GGML_ASSERT(ne11 == 1);
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src1 = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
|
|
|
|
vk_buffer * d_Qy = nullptr;
|
|
|
|
size_t qy_buf_offset = 0;
|
|
|
|
|
|
|
|
bool src1_uma = false;
|
|
|
|
|
|
|
|
if (vk_device.uma) {
|
|
|
|
ggml_vk_host_get(src1->data, d_Qy, qy_buf_offset);
|
|
|
|
src1_uma = d_Qy != nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
const uint64_t d_ne = ne01 * ne11 * ne12;
|
|
|
|
|
|
|
|
const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
|
|
|
|
const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
|
|
|
|
|
|
|
|
const uint64_t qx_sz = ggml_nbytes(src0);
|
|
|
|
const uint64_t qy_sz = ggml_nbytes(src1);
|
|
|
|
const uint64_t d_sz = sizeof(float) * d_ne;
|
|
|
|
|
|
|
|
vk_buffer* d_D = &extra->buffer_gpu;
|
|
|
|
const uint64_t d_buf_offset = extra->offset;
|
|
|
|
GGML_ASSERT(d_D != nullptr);
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer* d_Qx = &extra_src0->buffer_gpu;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
const uint64_t qx_buf_offset = extra_src0->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
if (load_y) {
|
|
|
|
d_Qy = &vk_prealloc_qy;
|
|
|
|
} else {
|
|
|
|
d_Qy = &extra_src1->buffer_gpu;
|
|
|
|
qy_buf_offset = extra_src1->offset;
|
|
|
|
GGML_ASSERT(d_Qx != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Allocate descriptor sets
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(vk_pipeline_mul_mat_vec_nc_f16_f32, 1);
|
|
|
|
|
|
|
|
const uint64_t qy_buffer_offset = (qy_buf_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
|
|
|
|
|
|
|
|
const uint64_t d_buffer_offset = (d_buf_offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
|
|
|
|
|
|
|
|
if (load_y) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Qy, qy_buf_offset, src1, 0, 0, ggml_nrows(src1));
|
|
|
|
}
|
|
|
|
|
|
|
|
// compute
|
|
|
|
const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, vk_pipeline_mul_mat_vec_nc_f16_f32, { { *d_Qx, qx_buf_offset, qx_sz }, { *d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { *d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
|
|
|
|
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
float * d = (float *) dst->data;
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, d_buf_offset, d, sizeof(float) * d_ne);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * dst) {
|
|
|
|
const uint64_t ne10 = src1->ne[0];
|
|
|
|
|
|
|
|
const uint64_t ne0 = dst->ne[0];
|
|
|
|
const uint64_t ne1 = dst->ne[1];
|
|
|
|
|
|
|
|
// TODO: find the optimal values for these
|
|
|
|
return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
|
|
|
|
(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || ggml_is_quantized(src1->type)) &&
|
|
|
|
dst->type == GGML_TYPE_F32 &&
|
|
|
|
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul_mat(vk_context * ctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
|
|
|
|
ggml_vk_mul_mat_vec_p021_f16_f32(ctx, src0, src1, dst);
|
|
|
|
} else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
|
|
|
|
ggml_vk_mul_mat_vec_nc_f16_f32(ctx, src0, src1, dst);
|
|
|
|
} else if (src1->ne[1] == 1 && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
|
|
|
|
ggml_vk_mul_mat_vec_q_f16(ctx, src0, src1, dst);
|
|
|
|
} else {
|
|
|
|
ggml_vk_mul_mat_q_f16(ctx, src0, src1, dst);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_op_repeat(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
// guaranteed to be an integer due to the check in ggml_can_repeat
|
|
|
|
const uint64_t ne0 = dst->ne[0];
|
|
|
|
const uint64_t ne1 = dst->ne[1];
|
|
|
|
const uint64_t ne2 = dst->ne[2];
|
|
|
|
const uint64_t ne3 = dst->ne[3];
|
|
|
|
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
const uint64_t ne03 = src0->ne[3];
|
|
|
|
|
|
|
|
const uint64_t nb0 = dst->nb[0];
|
|
|
|
const uint64_t nb1 = dst->nb[1];
|
|
|
|
const uint64_t nb2 = dst->nb[2];
|
|
|
|
const uint64_t nb3 = dst->nb[3];
|
|
|
|
|
|
|
|
const uint64_t nb00 = src0->nb[0];
|
|
|
|
const uint64_t nb01 = src0->nb[1];
|
|
|
|
const uint64_t nb02 = src0->nb[2];
|
|
|
|
const uint64_t nb03 = src0->nb[3];
|
|
|
|
|
|
|
|
const uint64_t nr0 = ne0/ne00;
|
|
|
|
const uint64_t nr1 = ne1/ne01;
|
|
|
|
const uint64_t nr2 = ne2/ne02;
|
|
|
|
const uint64_t nr3 = ne3/ne03;
|
|
|
|
|
|
|
|
// TODO: support for transposed / permuted tensors
|
|
|
|
GGML_ASSERT(nb0 == sizeof(float));
|
|
|
|
GGML_ASSERT(nb00 == sizeof(float));
|
|
|
|
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
|
|
|
|
GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
|
|
|
|
const vk_buffer* src_buf = &extra_src0->buffer_gpu;
|
|
|
|
const uint64_t src_offset = extra_src0->offset;
|
|
|
|
vk_buffer* dst_buf = &extra->buffer_gpu;
|
|
|
|
const uint64_t dst_offset = extra->offset;
|
|
|
|
|
|
|
|
std::vector<vk::BufferCopy> copies;
|
|
|
|
|
|
|
|
for (uint64_t i3 = 0; i3 < nr3; i3++) {
|
|
|
|
for (uint64_t k3 = 0; k3 < ne03; k3++) {
|
|
|
|
for (uint64_t i2 = 0; i2 < nr2; i2++) {
|
|
|
|
for (uint64_t k2 = 0; k2 < ne02; k2++) {
|
|
|
|
for (uint64_t i1 = 0; i1 < nr1; i1++) {
|
|
|
|
for (uint64_t k1 = 0; k1 < ne01; k1++) {
|
|
|
|
for (uint64_t i0 = 0; i0 < nr0; i0++) {
|
|
|
|
copies.push_back({
|
|
|
|
src_offset + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0,
|
|
|
|
dst_offset + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01,
|
|
|
|
ne00*nb0,
|
|
|
|
});
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ctx->s->buffer.copyBuffer(src_buf->buffer, dst_buf->buffer, copies);
|
|
|
|
|
|
|
|
(void) src1;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static vk_pipeline* ggml_vk_op_get_pipeline(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op) {
|
|
|
|
switch (op) {
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_add_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
GGML_ASSERT(src1->type == GGML_TYPE_I32);
|
|
|
|
if (dst->type == GGML_TYPE_F16) {
|
|
|
|
return &vk_pipeline_get_rows[src0->type];
|
|
|
|
}
|
|
|
|
if (dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_get_rows_f32[src0->type];
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_mul_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_scale_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_sqr_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_clamp_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DUP:
|
|
|
|
return ggml_vk_get_cpy_pipeline(src0->type, dst->type);
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_norm_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_rms_norm_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(dst)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_silu_f32;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_gelu_f32;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_relu_f32;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_diag_mask_inf_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_soft_max_f32;
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
{
|
|
|
|
const int mode = ((const int32_t *) dst->op_params)[2];
|
|
|
|
const bool is_neox = mode & 2;
|
|
|
|
const bool is_glm = mode & 4;
|
|
|
|
|
|
|
|
if (is_glm) {
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (is_neox) {
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_rope_neox_f32;
|
|
|
|
}
|
|
|
|
if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
|
|
|
|
return &vk_pipeline_rope_neox_f16;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
|
|
|
return &vk_pipeline_rope_f32;
|
|
|
|
}
|
|
|
|
if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
|
|
|
|
return &vk_pipeline_rope_f16;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_vk_func_t ggml_vk_op_get_func(ggml_op op) {
|
|
|
|
switch(op) {
|
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
return ggml_vk_op_repeat;
|
|
|
|
default:
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
|
|
|
|
static void ggml_vk_check_results_0(ggml_compute_params * params, ggml_tensor * tensor);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
#endif
|
|
|
|
|
|
|
|
template<typename PC>
|
|
|
|
static void ggml_vk_op_f32(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op, const PC&& pc) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", backend=" << src0->backend << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", backend=" << src1->backend << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
|
|
|
|
}
|
|
|
|
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))); // NOLINT
|
|
|
|
GGML_ASSERT(op == GGML_OP_CPY || ggml_vk_dim01_contiguous(src0)); // NOLINT
|
|
|
|
GGML_ASSERT(src1 == nullptr || ggml_vk_dim01_contiguous(src1)); // NOLINT
|
|
|
|
GGML_ASSERT(dst->extra != nullptr);
|
|
|
|
const uint64_t ne00 = src0->ne[0];
|
|
|
|
const uint64_t ne01 = src0->ne[1];
|
|
|
|
const uint64_t ne02 = src0->ne[2];
|
|
|
|
const uint64_t ne03 = src0->ne[3];
|
|
|
|
const uint64_t ne0 = ne00 * ne01;
|
|
|
|
const bool use_src1 = src1 != nullptr;
|
|
|
|
const uint64_t ne10 = use_src1 ? src1->ne[0] : 0;
|
|
|
|
const uint64_t ne11 = use_src1 ? src1->ne[1] : 0;
|
|
|
|
const uint64_t ne12 = use_src1 ? src1->ne[2] : 0;
|
|
|
|
const uint64_t ne13 = use_src1 ? src1->ne[3] : 0;
|
|
|
|
const uint64_t ne1 = ne10 * ne11;
|
|
|
|
// const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
|
|
|
|
const uint64_t nb2 = dst->nb[2];
|
|
|
|
const uint64_t nb3 = dst->nb[3];
|
|
|
|
|
|
|
|
vk_pipeline * pipeline = ggml_vk_op_get_pipeline(src0, src1, dst, op);
|
|
|
|
ggml_vk_func_t op_func;
|
|
|
|
|
|
|
|
if (pipeline == nullptr) {
|
|
|
|
op_func = ggml_vk_op_get_func(op);
|
|
|
|
if (op_func == nullptr) {
|
|
|
|
std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(op) << " for " << ggml_type_name(src0->type);
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
std::cerr << " and " << ggml_type_name(src1->type);
|
|
|
|
}
|
|
|
|
std::cerr << " to " << ggml_type_name(dst->type) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
op_func(ctx, src0, src1, dst);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
|
|
|
|
|
|
|
|
vk_buffer * d_X = nullptr;
|
|
|
|
size_t x_buf_offset = 0;
|
|
|
|
vk_buffer * d_Y = nullptr;
|
|
|
|
size_t y_buf_offset = 0;
|
|
|
|
|
|
|
|
bool src0_uma = false;
|
|
|
|
bool src1_uma = false;
|
|
|
|
|
|
|
|
if (vk_device.uma) {
|
|
|
|
ggml_vk_host_get(src0->data, d_X, x_buf_offset);
|
|
|
|
src0_uma = d_X != nullptr;
|
|
|
|
if (use_src1) {
|
|
|
|
ggml_vk_host_get(src1->data, d_Y, y_buf_offset);
|
|
|
|
src1_uma = d_Y != nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
const bool transfer_src0 = src0->backend != GGML_BACKEND_GPU && !src0_uma;
|
|
|
|
const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_GPU && !src1_uma;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type) * ne0, vk_device.properties.limits.minStorageBufferOffsetAlignment);
|
|
|
|
uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, vk_device.properties.limits.minStorageBufferOffsetAlignment) : 0;
|
|
|
|
uint64_t d_sz = ggml_type_size(dst->type) * ne0;
|
|
|
|
|
|
|
|
// Workaround for tiny tensor inputs on ROPE
|
|
|
|
if (use_src1 && src1->backend == GGML_BACKEND_GPU && y_sz > extra_src1->buffer_gpu.size) {
|
|
|
|
y_sz = VK_WHOLE_SIZE;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_buffer* d_D = &extra->buffer_gpu;
|
|
|
|
GGML_ASSERT(d_D != nullptr);
|
|
|
|
uint64_t d_buf_offset = (extra->offset / vk_device.properties.limits.minStorageBufferOffsetAlignment) * vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
GGML_ASSERT(d_buf_offset == extra->offset || op == GGML_OP_CPY); // NOLINT
|
|
|
|
if (transfer_src0) {
|
|
|
|
d_X = &vk_prealloc_qx;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if(!src0_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_X = &extra_src0->buffer_gpu;
|
|
|
|
x_buf_offset = extra_src0->offset;
|
|
|
|
GGML_ASSERT(d_X != nullptr);
|
|
|
|
}
|
|
|
|
if (transfer_src1) {
|
|
|
|
d_Y = &vk_prealloc_qy;
|
2024-01-30 12:59:30 +00:00
|
|
|
} else if (use_src1 && !src1_uma) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
d_Y = &extra_src1->buffer_gpu;
|
|
|
|
y_buf_offset = extra_src1->offset;
|
|
|
|
GGML_ASSERT(d_Y != nullptr);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (op == GGML_OP_CPY) {
|
|
|
|
GGML_ASSERT(!transfer_src0);
|
|
|
|
GGML_ASSERT(!transfer_src1);
|
2024-01-31 10:44:19 +00:00
|
|
|
x_sz = ggml_nbytes(src0);
|
|
|
|
d_sz = ggml_nbytes(dst);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
if (extra->offset + d_sz >= d_D->size) {
|
|
|
|
d_sz = VK_WHOLE_SIZE;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
std::array<uint32_t, 3> elements;
|
|
|
|
|
|
|
|
// copy src0 to device
|
|
|
|
if (transfer_src0) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_X, 0, src0, 0, 0, ggml_nrows(src0));
|
|
|
|
vk_staging_offset = x_sz * ne02 * ne03;
|
|
|
|
}
|
|
|
|
if (transfer_src1) {
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, d_Y, 0, src1, 0, 0, ggml_nrows(src1));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Single call if dimension 2 is contiguous
|
|
|
|
if (op == GGML_OP_CPY || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1)))) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*pipeline, 1);
|
|
|
|
|
|
|
|
switch (dst->op) {
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
elements = { (uint32_t)ggml_nrows(src0), 1, 1 };
|
|
|
|
break;
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
x_sz *= ne02 * ne03;
|
|
|
|
if (y_sz != VK_WHOLE_SIZE) {
|
|
|
|
y_sz *= ne12 * ne13;
|
|
|
|
}
|
|
|
|
if (op != GGML_OP_CPY) {
|
|
|
|
d_sz *= ne02 * ne03;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!use_src1 && op == GGML_OP_SOFT_MAX) {
|
|
|
|
// Empty src1 is possible on soft_max, but the shader needs a buffer
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset, x_sz }, { vk_prealloc_y, 0, vk_prealloc_y.size }, { *d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
} else if (use_src1) {
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset, x_sz }, { *d_Y, y_buf_offset, y_sz }, { *d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
} else {
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset, x_sz }, { *d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
}
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU && op == GGML_OP_CPY) {
|
|
|
|
ggml_vk_d2h_tensor_2d(ctx, d_D, 0, dst);
|
|
|
|
} else if(dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
float * d = (float *) dst->data;
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, 0, d, d_sz);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*pipeline, ne02 * ne03);
|
|
|
|
|
|
|
|
switch (dst->op) {
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
elements = { (uint32_t)ne01, 1, 1 };
|
|
|
|
break;
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
elements = { (uint32_t)ne01, (uint32_t)ne00, 1 };
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
elements = { (uint32_t)ne0, 1, 1 };
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (uint64_t i03 = 0; i03 < ne03; i03++) {
|
|
|
|
for (uint64_t i02 = 0; i02 < ne02; i02++) {
|
|
|
|
const uint32_t it_idx0 = (i03 * ne02 + i02);
|
|
|
|
const uint32_t it_idx1 = use_src1 ? ((i03 % ne13) * ne12 + (i02 % ne12)) : 0;
|
|
|
|
const uint32_t x_offset = x_sz * it_idx0;
|
|
|
|
const uint32_t y_offset = y_sz * it_idx1;
|
|
|
|
const uint32_t d_offset = d_sz * it_idx0;
|
|
|
|
|
|
|
|
if (!use_src1 && op == GGML_OP_SOFT_MAX) {
|
|
|
|
// Empty src1 is possible on soft_max, but the shader needs a buffer
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset, x_sz }, { vk_prealloc_y, 0, vk_prealloc_y.size }, { *d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
} else if (use_src1) {
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset + x_offset, x_sz }, { *d_Y, y_buf_offset + y_offset, y_sz }, { *d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
} else {
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_dispatch_pipeline(ctx, *pipeline, { { *d_X, x_buf_offset + x_offset, x_sz }, { *d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
|
|
|
|
}
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
// copy dst to host
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, d_buf_offset + d_offset, (char *) dst->data + i02*nb2 + i03*nb3, d_sz);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_repeat(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, src1, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_get_rows(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, src1, dst, GGML_OP_GET_ROWS, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_add(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, src1, dst, GGML_OP_ADD, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_mul(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, src1, dst, GGML_OP_MUL, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_scale(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
float * op_params = (float *)dst->op_params;
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_SCALE, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_sqr(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_SQR, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_clamp(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
float * op_params = (float *)dst->op_params;
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_CLAMP, { (uint32_t)ggml_nelements(src0), 0, op_params[0], op_params[1] });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_cpy(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
const int src0_type_size = ggml_type_size(src0->type);
|
|
|
|
const int dst_type_size = ggml_type_size(dst->type);
|
|
|
|
const uint32_t d_offset = (extra->offset % vk_device.properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
|
|
|
|
ggml_vk_op_f32<vk_op_cpy_push_constants>(ctx, src0, nullptr, dst, GGML_OP_CPY, {
|
|
|
|
(uint32_t)ggml_nelements(src0),
|
|
|
|
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size,
|
|
|
|
(uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size,
|
|
|
|
d_offset,
|
|
|
|
});
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_norm(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_rms_norm(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
float * op_params = (float *)dst->op_params;
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_unary(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_diag_mask_inf(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
int32_t * op_params = (int32_t *)dst->op_params;
|
|
|
|
ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, src0, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_soft_max(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
float * op_params = (float *)dst->op_params;
|
|
|
|
ggml_vk_op_f32<vk_op_push_constants>(ctx, src0, src1, dst, GGML_OP_SOFT_MAX, { (uint32_t)src0->ne[0], (uint32_t)(src1 != nullptr ? ggml_nrows(src1) : 0), op_params[0], 0.0f });
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_rope(vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
|
|
|
|
const int n_dims = ((int32_t *) dst->op_params)[1];
|
|
|
|
const int mode = ((int32_t *) dst->op_params)[2];
|
|
|
|
// const int n_ctx = ((int32_t *) dst->op_params)[3];
|
|
|
|
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
|
|
|
|
const float freq_base = ((float *) dst->op_params)[5];
|
|
|
|
const float freq_scale = ((float *) dst->op_params)[6];
|
|
|
|
const float ext_factor = ((float *) dst->op_params)[7];
|
|
|
|
const float attn_factor = ((float *) dst->op_params)[8];
|
|
|
|
const float beta_fast = ((float *) dst->op_params)[9];
|
|
|
|
const float beta_slow = ((float *) dst->op_params)[10];
|
|
|
|
|
|
|
|
const bool is_neox = mode & 2;
|
|
|
|
const bool is_glm = mode & 4;
|
|
|
|
|
|
|
|
GGML_ASSERT(!is_glm);
|
|
|
|
|
|
|
|
float corr_dims[2];
|
|
|
|
ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
|
|
|
|
|
|
|
|
if (is_neox) {
|
|
|
|
const float theta_scale = powf(freq_base, -2.0f/n_dims);
|
|
|
|
const float inv_ndims = -1.0f / n_dims;
|
|
|
|
ggml_vk_op_f32<vk_op_rope_neox_push_constants>(ctx, src0, src1, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f, theta_scale, inv_ndims });
|
|
|
|
} else {
|
|
|
|
ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, src0, src1, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, corr_dims[0], corr_dims[1], 0.0f, 0.0f });
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_nop(vk_context * ctx, const ggml_tensor * src0, ggml_tensor * dst) {
|
|
|
|
// If backend is CPU, data from src0 has to be copied off the device
|
|
|
|
if (dst->backend == GGML_BACKEND_CPU) {
|
|
|
|
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
vk_buffer * d_D = &extra_src0->buffer_gpu;
|
|
|
|
ggml_vk_sync_buffers(ctx);
|
|
|
|
ggml_vk_buffer_read_async(ctx, d_D, 0, dst->data, d_D->size);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef VK_RUN_TESTS
|
|
|
|
static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0, int ne1, int i0, int i1, int i2) {
|
|
|
|
if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
i0 = std::max(i0, 5);
|
|
|
|
i1 = std::max(i1, 5);
|
|
|
|
i2 = std::max(i2, 0);
|
|
|
|
fprintf(stderr, " ");
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
fprintf(stderr, "%7d ", idx1);
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
|
|
|
|
fprintf(stderr, "%7d: ", idx0);
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
if (idx0 >= 0 && idx0 < ne0 && idx1 >= 0 && idx1 < ne1) {
|
|
|
|
float val;
|
|
|
|
if (type == GGML_TYPE_F32) {
|
|
|
|
val = *((const float *) data + i2*ne1*ne0 + idx1*ne0 + idx0);
|
|
|
|
} else if (type == GGML_TYPE_F16) {
|
|
|
|
val = ggml_fp16_to_fp32(*((const ggml_fp16_t *) data + i2*ne1*ne0 + idx1*ne0 + idx0));
|
|
|
|
}
|
|
|
|
fprintf(stderr, "% 7.2f ", val);
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, " ");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
template <typename X_TYPE, typename Y_TYPE>
|
|
|
|
static void ggml_vk_test_matmul(size_t m, size_t n, size_t k, size_t batch, size_t num_it, int split_k, int shader_size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_test_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << shader_size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
const size_t x_ne = m * k * batch;
|
|
|
|
const size_t y_ne = k * n * batch;
|
|
|
|
const size_t d_ne = m * n * batch;
|
|
|
|
|
|
|
|
vk_pipeline * p;
|
|
|
|
std::string shname;
|
|
|
|
if (shader_size == 0) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_aligned_s;
|
|
|
|
shname = "F32_ALIGNED_S";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_aligned_s;
|
|
|
|
shname = "F16_F32_ALIGNED_S";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_aligned_s;
|
|
|
|
shname = "F16_ALIGNED_S";
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
} else if (shader_size == 1) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_aligned_m;
|
|
|
|
shname = "F32_ALIGNED_M";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_aligned_m;
|
|
|
|
shname = "F16_F32_ALIGNED_M";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_aligned_m;
|
|
|
|
shname = "F16_ALIGNED_M";
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
} else if (shader_size == 2) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_aligned_l;
|
|
|
|
shname = "F32_ALIGNED_L";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_aligned_l;
|
|
|
|
shname = "F16_F32_ALIGNED_L";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_aligned_l;
|
|
|
|
shname = "F16_ALIGNED_L";
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(0);
|
|
|
|
}
|
|
|
|
|
|
|
|
const size_t kpad = ggml_vk_align_size(k, p->align);
|
|
|
|
|
|
|
|
if (k != kpad) {
|
|
|
|
if (shader_size == 0) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_s;
|
|
|
|
shname = "F32_S";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_s;
|
|
|
|
shname = "F16_F32_S";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_s;
|
|
|
|
shname = "F16_S";
|
|
|
|
}
|
|
|
|
} else if (shader_size == 1) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_m;
|
|
|
|
shname = "F32_M";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_m;
|
|
|
|
shname = "F16_F32_M";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_m;
|
|
|
|
shname = "F16_M";
|
|
|
|
}
|
|
|
|
} else if (shader_size == 2) {
|
|
|
|
if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f32_l;
|
|
|
|
shname = "F32_L";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_f32_l;
|
|
|
|
shname = "F16_F32_L";
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
p = &vk_pipeline_matmul_f16_l;
|
|
|
|
shname = "F16_L";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(*p, num_it);
|
|
|
|
if (split_k > 1) {
|
|
|
|
ggml_vk_pipeline_allocate_descriptor_sets(vk_pipeline_matmul_split_k_reduce, num_it);
|
|
|
|
|
|
|
|
if (vk_prealloc_split_k.size < sizeof(float) * d_ne * split_k) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_split_k.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_split_k);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_split_k = ggml_vk_create_buffer_check(sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer d_X = ggml_vk_create_buffer_check(sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
|
|
|
vk_buffer d_Y = ggml_vk_create_buffer_check(sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
|
|
|
vk_buffer d_D = ggml_vk_create_buffer_check(sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
X_TYPE* x = (X_TYPE *) malloc(sizeof(X_TYPE) * x_ne);
|
|
|
|
Y_TYPE* y = (Y_TYPE *) malloc(sizeof(Y_TYPE) * y_ne);
|
|
|
|
float* d = (float *) malloc(sizeof(float) * d_ne);
|
|
|
|
|
|
|
|
for (size_t i = 0; i < x_ne; i++) {
|
|
|
|
if (std::is_same<float, X_TYPE>()) {
|
|
|
|
x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
|
|
|
|
x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for (size_t i = 0; i < y_ne; i++) {
|
|
|
|
if (std::is_same<float, Y_TYPE>()) {
|
|
|
|
y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
|
|
|
|
} else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_buffer_write(&d_X, 0, x, sizeof(X_TYPE) * k * m * batch);
|
|
|
|
ggml_vk_buffer_write(&d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
|
|
|
|
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.compute_queue);
|
|
|
|
for (size_t i = 0; i < num_it; i++) {
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
ggml_vk_matmul(ctx, *p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(vk_prealloc_split_k), m, n, k, k, k, m, split_k, batch, batch, batch, 1, 1, k*m, k*n, m*n);
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
auto begin = std::chrono::high_resolution_clock::now();
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
auto end = std::chrono::high_resolution_clock::now();
|
|
|
|
double time = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
|
|
|
|
|
|
|
|
// copy dst to host
|
|
|
|
ggml_vk_buffer_read(&d_D, 0, d, sizeof(float) * d_ne);
|
|
|
|
|
|
|
|
float * d_chk = (float *) malloc(sizeof(float) * d_ne);
|
|
|
|
|
|
|
|
ggml_init_params iparams = {
|
|
|
|
/*.mem_size =*/ 1024*1024*1024,
|
|
|
|
/*.mem_buffer =*/ NULL,
|
|
|
|
/*.no_alloc =*/ true,
|
|
|
|
};
|
|
|
|
|
|
|
|
ggml_context * ggml_ctx = ggml_init(iparams);
|
|
|
|
|
|
|
|
ggml_type src0_type;
|
|
|
|
ggml_type src1_type;
|
|
|
|
|
|
|
|
if (std::is_same<float, X_TYPE>()) {
|
|
|
|
src0_type = GGML_TYPE_F32;
|
|
|
|
} else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
|
|
|
|
src0_type = GGML_TYPE_F16;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
if (std::is_same<float, Y_TYPE>()) {
|
|
|
|
src1_type = GGML_TYPE_F32;
|
|
|
|
} else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
|
|
|
|
src1_type = GGML_TYPE_F16;
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, src0_type, k, m, batch);
|
|
|
|
ggml_tensor * src1_ggml = ggml_new_tensor_3d(ggml_ctx, src1_type, k, n, batch);
|
|
|
|
ggml_tensor * tensor_ggml = ggml_mul_mat(ggml_ctx, src0_ggml, src1_ggml);
|
|
|
|
|
|
|
|
src0_ggml->data = x;
|
|
|
|
src1_ggml->data = y;
|
|
|
|
tensor_ggml->data = d_chk;
|
|
|
|
|
|
|
|
vk_disable = true;
|
|
|
|
|
|
|
|
ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
|
|
|
|
ggml_build_forward_expand(cgraph, tensor_ggml);
|
|
|
|
|
|
|
|
ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
|
|
|
|
|
|
|
|
vk_disable = false;
|
|
|
|
|
|
|
|
ggml_free(ggml_ctx);
|
|
|
|
|
|
|
|
double avg_err = 0.0;
|
|
|
|
int first_err_n = -1;
|
|
|
|
int first_err_m = -1;
|
|
|
|
int first_err_b = -1;
|
|
|
|
|
|
|
|
for (size_t i = 0; i < m*n*batch; i++) {
|
|
|
|
double err = std::fabs(d[i] - d_chk[i]);
|
|
|
|
avg_err += err;
|
|
|
|
|
|
|
|
if (err > 0.05f && first_err_n == -1) {
|
|
|
|
first_err_b = i / (m * n);
|
|
|
|
first_err_n = (i % (m * n)) / m;
|
|
|
|
first_err_m = (i % (m * n)) % m;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
avg_err /= m * n;
|
|
|
|
|
|
|
|
std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms avg_err=" << avg_err << std::endl;
|
|
|
|
|
|
|
|
if (avg_err > 0.1) {
|
|
|
|
std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
|
|
|
|
std::cerr << "Actual result: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
std::cerr << "Expected result: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
|
|
|
|
if (split_k > 1) {
|
|
|
|
float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k);
|
|
|
|
ggml_vk_buffer_read(&vk_prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k);
|
|
|
|
|
|
|
|
std::cerr << "d_buf0: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(split_k_buf, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
|
|
|
|
std::cerr << "d_buf1: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(split_k_buf + d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
|
|
|
|
std::cerr << "d_buf2: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(split_k_buf + 2 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
|
|
|
|
std::cerr << "d_buf3: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_matrix_area(split_k_buf + 3 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
|
|
|
|
|
|
|
|
free(split_k_buf);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
free(d_chk);
|
|
|
|
|
|
|
|
ggml_vk_queue_cleanup(vk_device.transfer_queue);
|
|
|
|
ggml_vk_queue_cleanup(vk_device.compute_queue);
|
|
|
|
|
|
|
|
ggml_vk_destroy_buffer(d_X);
|
|
|
|
ggml_vk_destroy_buffer(d_Y);
|
|
|
|
ggml_vk_destroy_buffer(d_D);
|
|
|
|
|
|
|
|
ggml_vk_pipeline_cleanup(*p);
|
|
|
|
ggml_vk_pipeline_cleanup(vk_pipeline_matmul_split_k_reduce);
|
|
|
|
|
|
|
|
free(x);
|
|
|
|
free(y);
|
|
|
|
free(d);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, int i0, int i1, int i2, int i3) {
|
|
|
|
if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
|
|
|
|
return;
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
i0 = std::max(i0, 5);
|
|
|
|
i1 = std::max(i1, 5);
|
|
|
|
i2 = std::max(i2, 0);
|
|
|
|
i3 = std::max(i3, 0);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
fprintf(stderr, " ");
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
fprintf(stderr, "%7d ", idx1);
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
|
|
|
|
fprintf(stderr, "%7d: ", idx0);
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
if (idx0 >= 0 && idx0 < tensor->ne[0] && idx1 >= 0 && idx1 < tensor->ne[1] && i2 >= 0 && i2 < tensor->ne[2] && i3 >= 0 && i3 < tensor->ne[3]) {
|
|
|
|
float val;
|
|
|
|
if (tensor->type == GGML_TYPE_F32) {
|
|
|
|
val = *(float *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
|
|
|
|
} else if (tensor->type == GGML_TYPE_F16) {
|
|
|
|
val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
|
|
|
|
}
|
|
|
|
fprintf(stderr, "% 7.2f ", val);
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, " ");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_test_h2d_nc(size_t ne0, size_t ne1, size_t ne2, size_t ne3) {
|
|
|
|
const size_t ne = ne0 * ne1 * ne2 * ne3;
|
|
|
|
|
|
|
|
ggml_init_params iparams = {
|
|
|
|
/*.mem_size =*/ 1024*1024*1024,
|
|
|
|
/*.mem_buffer =*/ NULL,
|
|
|
|
/*.no_alloc =*/ true,
|
|
|
|
};
|
|
|
|
|
|
|
|
ggml_context * ggml_ctx = ggml_init(iparams);
|
|
|
|
|
|
|
|
ggml_tensor * tensor = ggml_new_tensor_4d(ggml_ctx, GGML_TYPE_F32, ne0, ne2, ne1, ne3); // NOLINT
|
|
|
|
ggml_tensor * result_tensor = ggml_new_tensor_4d(ggml_ctx, GGML_TYPE_F32, ne0, ne1, ne2, ne3);
|
|
|
|
|
|
|
|
float * data = (float *) ggml_vk_host_malloc(ggml_nbytes(tensor));
|
|
|
|
tensor->data = data;
|
|
|
|
|
|
|
|
float * result_data = (float *) malloc(ggml_nbytes(tensor));
|
|
|
|
result_tensor->data = result_data;
|
|
|
|
|
|
|
|
// Permute
|
|
|
|
{
|
|
|
|
size_t tmp = tensor->nb[2];
|
|
|
|
tensor->nb[2] = tensor->nb[1];
|
|
|
|
tensor->nb[1] = tmp;
|
|
|
|
|
|
|
|
tensor->ne[2] = ne2;
|
|
|
|
tensor->ne[1] = ne1;
|
|
|
|
}
|
|
|
|
|
|
|
|
for (size_t i = 0; i < ne; i++) {
|
|
|
|
data[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.compute_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer buffer = ggml_vk_create_buffer_check(ggml_nbytes(tensor), vk::MemoryPropertyFlagBits::eDeviceLocal);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
ggml_vk_h2d_tensor_2d(ctx, &buffer, 0, tensor, 0, 0, ggml_nrows(tensor));
|
|
|
|
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
ggml_vk_buffer_read(&buffer, 0, result_data, ggml_nbytes(tensor));
|
|
|
|
|
|
|
|
double avg_err = 0.0;
|
|
|
|
int first_err_i0 = -1;
|
|
|
|
int first_err_i1 = -1;
|
|
|
|
int first_err_i2 = -1;
|
|
|
|
int first_err_i3 = -1;
|
|
|
|
|
|
|
|
for (size_t i3 = 0; i3 < ne3; i3++) {
|
|
|
|
for (size_t i2 = 0; i2 < ne2; i2++) {
|
|
|
|
for (size_t i1 = 0; i1 < ne1; i1++) {
|
|
|
|
for (size_t i0 = 0; i0 < ne0; i0++) {
|
|
|
|
float correct = *(float *) ((char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
|
|
|
|
float result = *(float *) ((char *) result_data + i3*ne2*ne1*ne0*sizeof(float) + i2*ne1*ne0*sizeof(float) + i1*ne0*sizeof(float) + i0*sizeof(float));
|
|
|
|
double err = std::fabs(result - correct);
|
|
|
|
|
|
|
|
avg_err += err;
|
|
|
|
|
|
|
|
if (err > 0.05f && first_err_i0 == -1) {
|
|
|
|
first_err_i0 = i0;
|
|
|
|
first_err_i1 = i1;
|
|
|
|
first_err_i2 = i2;
|
|
|
|
first_err_i3 = i3;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
avg_err /= ne;
|
|
|
|
|
|
|
|
std::cerr << "TEST nc copy ne0=" << ne0 << " ne1=" << ne1 << " ne2=" << ne2 << " ne3=" << ne3 << " avg_err=" << avg_err << std::endl;
|
|
|
|
|
|
|
|
if (avg_err > 0.1) {
|
|
|
|
std::cerr << "i0 = " << first_err_i0 << " i1 = " << first_err_i1 << " i2 = " << first_err_i2 << " i3 = " << first_err_i3 << std::endl;
|
|
|
|
std::cerr << "Actual result: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(result_tensor, first_err_i0, first_err_i1, first_err_i2, first_err_i3);
|
|
|
|
std::cerr << "Expected result: " << std::endl << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, first_err_i0, first_err_i1, first_err_i2, first_err_i3);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_free(ggml_ctx);
|
|
|
|
|
|
|
|
ggml_vk_destroy_buffer(buffer);
|
|
|
|
|
|
|
|
ggml_vk_host_free(data);
|
|
|
|
free(result_data);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_test_transfer(size_t ne, bool pinned) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_test_transfer(" << ne << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
// Check transfers are correct
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer buffer = ggml_vk_create_buffer_check(sizeof(float) * ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
float * x;
|
|
|
|
float * y;
|
|
|
|
if (pinned) {
|
|
|
|
x = (float *) ggml_vk_host_malloc(sizeof(float) * ne);
|
|
|
|
y = (float *) ggml_vk_host_malloc(sizeof(float) * ne);
|
|
|
|
} else {
|
|
|
|
x = (float *) malloc(sizeof(float) * ne);
|
|
|
|
y = (float *) malloc(sizeof(float) * ne);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (size_t i = 0; i < ne; i++) {
|
|
|
|
x[i] = rand() / (float)RAND_MAX;
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_context * ctx = ggml_vk_create_context(vk_device.compute_queue);
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
|
|
|
|
auto begin = std::chrono::high_resolution_clock::now();
|
|
|
|
|
|
|
|
ggml_vk_buffer_write_async(ctx, &buffer, 0, x, sizeof(float) * ne);
|
|
|
|
|
|
|
|
for (auto& cpy : ctx->in_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
ctx->in_memcpys.clear();
|
|
|
|
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
auto end = std::chrono::high_resolution_clock::now();
|
|
|
|
|
|
|
|
double ms_to_gpu = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
|
|
|
|
|
|
|
|
ggml_vk_ctx_begin(ctx);
|
|
|
|
|
|
|
|
begin = std::chrono::high_resolution_clock::now();
|
|
|
|
|
|
|
|
ggml_vk_buffer_read_async(ctx, &buffer, 0, y, sizeof(float) * ne);
|
|
|
|
|
|
|
|
ggml_vk_ctx_end(ctx);
|
|
|
|
ggml_vk_submit(ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
for (auto& cpy : ctx->out_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
ctx->out_memcpys.clear();
|
|
|
|
|
|
|
|
end = std::chrono::high_resolution_clock::now();
|
|
|
|
|
|
|
|
double ms_from_gpu = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
|
|
|
|
|
|
|
|
double avg_err = 0.0;
|
|
|
|
for (size_t i = 0; i < ne; i++) {
|
|
|
|
avg_err += std::fabs(x[i] - y[i]);
|
|
|
|
}
|
|
|
|
|
|
|
|
double kb = ne * sizeof(float) / 1024.0;
|
|
|
|
|
|
|
|
std::cerr << "TEST TRANSFER " << kb << " KB to_gpu " << ms_to_gpu << "ms (" << kb / ms_to_gpu * 1000.0 / 1024.0 << " MB/s) from_gpu " << ms_from_gpu << "ms (" << kb / ms_from_gpu * 1000.0 / 1024.0 << " MB/s) avg_err=" << avg_err / ne << std::endl;
|
|
|
|
|
|
|
|
ggml_vk_destroy_buffer(buffer);
|
|
|
|
|
|
|
|
if (pinned) {
|
|
|
|
ggml_vk_host_free(x);
|
|
|
|
ggml_vk_host_free(y);
|
|
|
|
} else {
|
|
|
|
free(x);
|
|
|
|
free(y);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
static ggml_tensor_extra_gpu * ggml_vk_tensor_create_extra(ggml_tensor * tensor) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_create_extra(" << tensor << " (" << tensor->name << ", " << ggml_op_name(tensor->op) << "))" << std::endl;
|
|
|
|
#endif
|
|
|
|
ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu;
|
|
|
|
extra->reset();
|
|
|
|
tensor->extra = extra;
|
|
|
|
return extra;
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_tensor * ggml_vk_find_last_use(const ggml_tensor * node, ggml_cgraph * graph) {
|
|
|
|
GGML_ASSERT(node != nullptr);
|
|
|
|
|
|
|
|
for (int i = graph->n_nodes - 1; i >= 0; i--) {
|
|
|
|
for (int j = 0; j < GGML_MAX_SRC; j++) {
|
|
|
|
if (graph->nodes[i]->src[j] == node) {
|
|
|
|
return graph->nodes[i];
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return nullptr;
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_preallocate_buffers_graph(ggml_tensor * node){
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
const bool any_on_device = node->backend == GGML_BACKEND_GPU
|
|
|
|
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|
|
|
|
|| (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_GPU));
|
|
|
|
|
|
|
|
if (vk_disable || (!any_on_device && node->op != GGML_OP_MUL_MAT)) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
|
|
|
|
if (extra == nullptr) {
|
|
|
|
// Workaround for CPU backend BLAS matmul calls
|
|
|
|
extra = ggml_vk_tensor_create_extra(node);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor * src0 = node->src[0];
|
|
|
|
ggml_tensor * src1 = node->src[1];
|
|
|
|
|
|
|
|
const bool use_src0 = src0 != nullptr;
|
|
|
|
const int64_t ne00 = use_src0 ? src0->ne[0] : 0;
|
|
|
|
const int64_t ne01 = use_src0 ? src0->ne[1] : 0;
|
|
|
|
const int64_t ne02 = use_src0 ? src0->ne[2] : 0;
|
|
|
|
const int64_t ne03 = use_src0 ? src0->ne[3] : 0;
|
|
|
|
const bool use_src1 = src1 != nullptr && node->op != GGML_OP_CPY && node->op != GGML_OP_CONT && node->op != GGML_OP_DUP;
|
|
|
|
const int64_t ne10 = use_src1 ? src1->ne[0] : 0;
|
|
|
|
const int64_t ne11 = use_src1 ? src1->ne[1] : 0;
|
|
|
|
const int64_t ne12 = use_src1 ? src1->ne[2] : 0;
|
|
|
|
const int64_t ne13 = use_src1 ? src1->ne[3] : 0;
|
|
|
|
const int64_t ne20 = node->ne[0];
|
|
|
|
const int64_t ne21 = node->ne[1];
|
|
|
|
const int64_t ne22 = node->ne[2];
|
|
|
|
const int64_t ne23 = node->ne[3];
|
|
|
|
|
|
|
|
const bool f16_f32_kernel = use_src1 && src1->type == GGML_TYPE_F32;
|
|
|
|
|
|
|
|
int split_k;
|
|
|
|
if (node->op == GGML_OP_MUL_MAT) {
|
|
|
|
split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
|
|
|
|
} else {
|
|
|
|
split_k = 1;
|
|
|
|
}
|
|
|
|
const uint32_t x_ne = ne00 * ne01;
|
|
|
|
const uint32_t y_ne = ne10 * ne11;
|
|
|
|
const uint32_t d_ne = ne20 * ne21;
|
|
|
|
|
|
|
|
const uint64_t qx_sz = use_src0 ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), vk_device.properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0;
|
|
|
|
const uint64_t qy_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type), vk_device.properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0;
|
|
|
|
const uint64_t x_sz = use_src0 ? ggml_vk_align_size(sizeof(ggml_fp16_t) * x_ne, vk_device.properties.limits.minStorageBufferOffsetAlignment) * ne02 * ne03 : 0;
|
|
|
|
const uint64_t y_sz = use_src1 ? ggml_vk_align_size(f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne, vk_device.properties.limits.minStorageBufferOffsetAlignment) * ne12 * ne13 : 0;
|
|
|
|
uint64_t d_sz = ggml_vk_align_size(ggml_type_size(node->type) * d_ne, vk_device.properties.limits.minStorageBufferOffsetAlignment) * ne22 * ne23;
|
|
|
|
const uint64_t split_k_size = split_k > 1 ? d_sz * 4 : 0;
|
|
|
|
|
|
|
|
if (extra->buffer_gpu.size == 0) {
|
|
|
|
// Workaround for CPU backend BLAS matmul calls
|
|
|
|
extra->buffer_gpu = ggml_vk_create_buffer_temp(d_sz);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (node->op) {
|
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DUP:
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
break;
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(node)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
if (vk_prealloc_size_qx < qx_sz) {
|
|
|
|
vk_prealloc_size_qx = qx_sz;
|
|
|
|
}
|
|
|
|
if (vk_prealloc_size_qy < qy_sz) {
|
|
|
|
vk_prealloc_size_qy = qy_sz;
|
|
|
|
}
|
|
|
|
if (vk_prealloc_size_x < x_sz) {
|
|
|
|
vk_prealloc_size_x = x_sz;
|
|
|
|
}
|
|
|
|
if (vk_prealloc_size_y < y_sz) {
|
|
|
|
vk_prealloc_size_y = y_sz;
|
|
|
|
}
|
|
|
|
if (vk_prealloc_size_split_k < split_k_size) {
|
|
|
|
vk_prealloc_size_split_k = split_k_size;
|
|
|
|
}
|
|
|
|
if (vk_staging_size < x_sz + y_sz) {
|
|
|
|
vk_staging_size = x_sz + y_sz;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_preallocate_buffers() {
|
|
|
|
if (vk_disable) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_preallocate_buffers()" << std::endl;
|
|
|
|
std::cerr << "qx_size: " << vk_prealloc_size_qx << " qy_size: " << vk_prealloc_size_qy << " x_size: " << vk_prealloc_size_x << " y_size: " << vk_prealloc_size_y << " split_k_size: " << vk_prealloc_size_split_k << std::endl;
|
|
|
|
#endif
|
|
|
|
#if defined(VK_RUN_TESTS)
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_staging = ggml_vk_create_buffer_check(100ul * 1024ul * 1024ul, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
ggml_vk_test_transfer(8192 * 1000, false);
|
|
|
|
ggml_vk_test_transfer(8192 * 1000, true);
|
|
|
|
|
|
|
|
const std::vector<size_t> vals {
|
|
|
|
8, 8, 8,
|
|
|
|
100, 46, 576,
|
|
|
|
623, 111, 128,
|
|
|
|
100, 46, 558,
|
|
|
|
512, 1, 256,
|
|
|
|
128, 110, 622,
|
|
|
|
511, 511, 127,
|
|
|
|
511, 511, 7,
|
|
|
|
511, 511, 17,
|
|
|
|
49, 49, 128,
|
|
|
|
128, 49, 49,
|
|
|
|
4096, 49, 4096,
|
|
|
|
11008, 49, 4096,
|
|
|
|
4096, 49, 11008,
|
|
|
|
32000, 49, 4096,
|
|
|
|
512, 512, 128,
|
|
|
|
128, 512, 512,
|
|
|
|
4096, 512, 4096,
|
|
|
|
11008, 512, 4096,
|
|
|
|
4096, 512, 11008,
|
|
|
|
32000, 512, 4096,
|
|
|
|
};
|
|
|
|
const size_t num_it = 1;
|
|
|
|
for (size_t i = 0; i < vals.size(); i += 3) {
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0);
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1);
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2);
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
|
|
|
|
ggml_vk_test_matmul<ggml_fp16_t, float>(vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if (vk_prealloc_size_qx > 0 && vk_prealloc_qx.size < vk_prealloc_size_qx) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_qx.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_qx);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_qx = ggml_vk_create_buffer_device(vk_prealloc_size_qx);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
if (vk_prealloc_size_qy > 0 && vk_prealloc_qy.size < vk_prealloc_size_qy) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_qy.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_qy);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_qy = ggml_vk_create_buffer_device(vk_prealloc_size_qy);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
if (vk_prealloc_size_x > 0 && vk_prealloc_x.size < vk_prealloc_size_x) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_x.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_x);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_x = ggml_vk_create_buffer_device(vk_prealloc_size_x);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
if (vk_prealloc_size_y > 0 && vk_prealloc_y.size < vk_prealloc_size_y) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_y.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_y);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_y = ggml_vk_create_buffer_device(vk_prealloc_size_y);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
if (vk_prealloc_size_split_k > 0 && vk_prealloc_split_k.size < vk_prealloc_size_split_k) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_prealloc_split_k.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_split_k);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_prealloc_split_k = ggml_vk_create_buffer_device(vk_prealloc_size_split_k);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
if (vk_staging_size > 0 && vk_staging.size < vk_staging_size) {
|
|
|
|
// Resize buffer
|
|
|
|
if (vk_staging.size > 0) {
|
|
|
|
ggml_vk_destroy_buffer(vk_staging);
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_staging = ggml_vk_create_buffer_check(vk_staging_size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_build_graph(ggml_tensor * node, bool last_node){
|
|
|
|
const bool any_on_device = node->backend == GGML_BACKEND_GPU
|
|
|
|
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|
|
|
|
|| (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
if (vk_disable || (!any_on_device && node->op != GGML_OP_MUL_MAT) || (node->op == GGML_OP_MUL_MAT && !any_on_device && !ggml_vk_can_mul_mat(node->src[0], node->src[1], node))) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
vk_semaphore_idx = 0;
|
|
|
|
vk_staging_offset = 0;
|
|
|
|
|
|
|
|
const ggml_tensor * src0 = node->src[0];
|
|
|
|
const ggml_tensor * src1 = node->src[1];
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) node->extra;
|
|
|
|
|
|
|
|
switch (node->op) {
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(node)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
// case GGML_OP_GET_ROWS:
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DUP:
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
case GGML_OP_NONE:
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
if (any_on_device) {
|
|
|
|
std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vk_ctx == nullptr) {
|
|
|
|
vk_ctx = ggml_vk_create_context(vk_device.compute_queue);
|
|
|
|
ggml_vk_ctx_begin(vk_ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (node->op) {
|
|
|
|
case GGML_OP_REPEAT:
|
|
|
|
ggml_vk_repeat(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
ggml_vk_get_rows(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
ggml_vk_add(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
ggml_vk_mul(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
ggml_vk_scale(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
ggml_vk_sqr(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
ggml_vk_clamp(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DUP:
|
|
|
|
ggml_vk_cpy(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_NONE:
|
|
|
|
ggml_vk_nop(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
ggml_vk_norm(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
ggml_vk_rms_norm(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(node)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
ggml_vk_unary(vk_ctx, src0, node);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
ggml_vk_diag_mask_inf(vk_ctx, src0, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
ggml_vk_soft_max(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
ggml_vk_rope(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
ggml_vk_mul_mat(vk_ctx, src0, src1, node);
|
|
|
|
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
extra->ready = true;
|
|
|
|
extra->ctx_idx = vk_ctx->idx;
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
|
|
|
// Force context reset on each node so that each tensor ends up in its own context
|
|
|
|
// and can be run and compared to its CPU equivalent separately
|
|
|
|
last_node = true;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
if (node->backend == GGML_BACKEND_CPU || last_node) {
|
|
|
|
ggml_vk_ctx_end(vk_ctx);
|
|
|
|
vk_ctx->exit_tensor = node;
|
|
|
|
vk_ctx = nullptr;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
bool ggml_vk_compute_forward(ggml_compute_params * params, ggml_tensor * tensor){
|
|
|
|
const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
|
|
|
|
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|
|
|
|
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
if (vk_disable || (!any_on_device && tensor->op != GGML_OP_MUL_MAT)) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = nullptr;
|
|
|
|
|
|
|
|
switch (tensor->op) {
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
case GGML_OP_GET_ROWS:
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DUP:
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_NONE:
|
|
|
|
extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
break;
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(tensor)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
if (!any_on_device && !ggml_vk_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (extra == nullptr) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (params->ith != 0) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", backend=" << tensor->backend << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
|
|
|
ggml_vk_check_results_0(params, tensor);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
GGML_ASSERT(extra->ready);
|
|
|
|
|
|
|
|
vk_context& ctx = vk_gc.contexts[extra->ctx_idx];
|
|
|
|
|
|
|
|
// Only run if ctx hasn't been submitted yet
|
|
|
|
if (!ctx.seqs.empty()) {
|
|
|
|
// Do staging buffer copies
|
|
|
|
for (auto& cpy : ctx.in_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_submit(&ctx, vk_fence);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (tensor == ctx.exit_tensor) {
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
// Do staging buffer copies
|
|
|
|
for (auto& cpy : ctx.out_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
ctx.in_memcpys.clear();
|
|
|
|
ctx.out_memcpys.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
extra->ready = false;
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_graph_cleanup() {
|
|
|
|
if (vk_disable) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_graph_cleanup()" << std::endl;
|
|
|
|
#endif
|
|
|
|
for (auto& buffer : vk_gc.temp_buffers) {
|
|
|
|
ggml_vk_pool_free(buffer);
|
|
|
|
}
|
|
|
|
vk_gc.temp_buffers.clear();
|
|
|
|
|
|
|
|
for (auto * pipeline : vk_gc.pipelines) {
|
|
|
|
ggml_vk_pipeline_cleanup(*pipeline);
|
|
|
|
}
|
|
|
|
vk_gc.pipelines.clear();
|
|
|
|
|
|
|
|
ggml_vk_queue_cleanup(vk_device.compute_queue);
|
|
|
|
ggml_vk_queue_cleanup(vk_device.transfer_queue);
|
|
|
|
|
|
|
|
for (size_t i = 0; i < vk_gc.semaphores.size(); i++) {
|
|
|
|
vk_device.device.destroySemaphore({ vk_gc.semaphores[i].s });
|
|
|
|
}
|
|
|
|
vk_gc.semaphores.clear();
|
|
|
|
|
|
|
|
for (size_t i = 0; i < vk_gc.tl_semaphores.size(); i++) {
|
|
|
|
vk_device.device.destroySemaphore({ vk_gc.tl_semaphores[i].s });
|
|
|
|
}
|
|
|
|
vk_gc.tl_semaphores.clear();
|
|
|
|
|
|
|
|
vk_event_idx = 0;
|
|
|
|
|
|
|
|
for (auto& event : vk_gc.events) {
|
|
|
|
vk_device.device.resetEvent(event);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_staging_offset = 0;
|
|
|
|
|
|
|
|
vk_ctx = nullptr;
|
|
|
|
vk_gc.contexts.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
static void ggml_vk_cleanup() {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_vk_cleanup()" << std::endl;
|
|
|
|
#endif
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_x);
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_y);
|
|
|
|
ggml_vk_destroy_buffer(vk_prealloc_split_k);
|
|
|
|
ggml_vk_destroy_buffer(vk_staging);
|
|
|
|
ggml_vk_destroy_buffer(vk_sync_staging);
|
|
|
|
|
|
|
|
vk_prealloc_size_x = 0;
|
|
|
|
vk_prealloc_size_y = 0;
|
|
|
|
vk_prealloc_size_split_k = 0;
|
|
|
|
vk_staging_size = 0;
|
|
|
|
|
|
|
|
for (auto& event : vk_gc.events) {
|
|
|
|
vk_device.device.destroyEvent(event);
|
|
|
|
}
|
|
|
|
vk_gc.events.clear();
|
|
|
|
}
|
|
|
|
|
|
|
|
// backend interface
|
|
|
|
|
|
|
|
#define UNUSED GGML_UNUSED
|
|
|
|
|
|
|
|
struct ggml_backend_vk_context {
|
|
|
|
std::string name;
|
|
|
|
};
|
|
|
|
|
|
|
|
// device backend
|
|
|
|
|
|
|
|
static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000; // NOLINT
|
|
|
|
|
|
|
|
struct ggml_backend_vk_buffer_context {
|
|
|
|
vk_buffer dev_buffer;
|
|
|
|
ggml_tensor_extra_gpu * temp_tensor_extras = nullptr;
|
|
|
|
size_t temp_tensor_extra_index = 0;
|
|
|
|
std::string name;
|
|
|
|
|
|
|
|
ggml_backend_vk_buffer_context(vk_buffer dev_buffer) :
|
|
|
|
dev_buffer(dev_buffer),
|
|
|
|
name(GGML_VK_NAME) {
|
|
|
|
}
|
|
|
|
|
|
|
|
~ggml_backend_vk_buffer_context() {
|
|
|
|
ggml_vk_destroy_buffer(dev_buffer);
|
|
|
|
delete[] temp_tensor_extras;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * ggml_vk_alloc_temp_tensor_extra() {
|
|
|
|
if (temp_tensor_extras == nullptr) {
|
|
|
|
temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_VK_MAX_NODES];
|
|
|
|
}
|
|
|
|
|
|
|
|
size_t alloc_index = temp_tensor_extra_index;
|
|
|
|
temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_VK_MAX_NODES;
|
|
|
|
ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index];
|
|
|
|
extra->reset();
|
|
|
|
|
|
|
|
return extra;
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
GGML_CALL static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
|
|
|
|
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
|
|
|
return ctx->name.c_str();
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
|
|
|
|
return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
|
|
|
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_destroy_buffer(ctx->dev_buffer);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
delete ctx;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
|
|
|
|
return vk_ptr_base;
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = ctx->ggml_vk_alloc_temp_tensor_extra();
|
|
|
|
if (tensor->view_src != nullptr && tensor->view_src->extra != nullptr) {
|
|
|
|
ggml_tensor_extra_gpu * extra_view = (ggml_tensor_extra_gpu *) tensor->view_src->extra;
|
|
|
|
extra->buffer_gpu = extra_view->buffer_gpu;
|
|
|
|
extra->offset = extra_view->offset + tensor->view_offs;
|
|
|
|
} else {
|
|
|
|
extra->buffer_gpu = ctx->dev_buffer;
|
|
|
|
extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
|
|
|
|
}
|
|
|
|
|
|
|
|
tensor->backend = GGML_BACKEND_GPU;
|
|
|
|
tensor->extra = extra;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
ggml_vk_buffer_write(&extra->buffer_gpu, extra->offset + offset, data, size);
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, extra->offset + offset, data, size);
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
|
|
|
|
if (ggml_backend_buffer_is_vk(src->buffer)) {
|
|
|
|
ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra;
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
|
|
|
|
ggml_vk_buffer_copy(&src_extra->buffer_gpu, src_extra->offset, &dst_extra->buffer_gpu, dst_extra->offset, ggml_nbytes(src));
|
|
|
|
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
|
|
|
|
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
|
|
|
|
|
|
|
ggml_vk_buffer_memset(&ctx->dev_buffer, 0, value, buffer->size);
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_vk_buffer_get_name,
|
|
|
|
/* .free_buffer = */ ggml_backend_vk_buffer_free_buffer,
|
|
|
|
/* .get_base = */ ggml_backend_vk_buffer_get_base,
|
|
|
|
/* .init_tensor = */ ggml_backend_vk_buffer_init_tensor,
|
|
|
|
/* .set_tensor = */ ggml_backend_vk_buffer_set_tensor,
|
|
|
|
/* .get_tensor = */ ggml_backend_vk_buffer_get_tensor,
|
|
|
|
/* .cpy_tensor = */ ggml_backend_vk_buffer_cpy_tensor,
|
|
|
|
/* .clear = */ ggml_backend_vk_buffer_clear,
|
|
|
|
/* .reset = */ NULL,
|
|
|
|
};
|
|
|
|
|
|
|
|
// vk buffer type
|
|
|
|
struct ggml_backend_vk_buffer_type_context {
|
|
|
|
std::string name;
|
|
|
|
};
|
|
|
|
|
|
|
|
GGML_CALL static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
|
|
|
ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
|
|
|
|
|
|
|
|
return ctx->name.c_str();
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_buffer_type_alloc_buffer(" << size << ")" << std::endl;
|
|
|
|
#endif
|
2024-01-30 12:59:30 +00:00
|
|
|
vk_buffer dev_buffer = ggml_vk_create_buffer_device(size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
|
|
|
|
ggml_backend_vk_buffer_context * ctx = new ggml_backend_vk_buffer_context(dev_buffer);
|
|
|
|
|
|
|
|
return ggml_backend_buffer_init(buft, ggml_backend_vk_buffer_interface, ctx, size);
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
|
|
|
return vk_device.properties.limits.minStorageBufferOffsetAlignment;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
|
|
|
|
return vk_device.max_memory_allocation_size;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
|
|
|
|
return ggml_nbytes(tensor);
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_vk_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
|
|
|
|
return ggml_backend_is_vk(backend);
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_vk_buffer_type_name,
|
|
|
|
/* .alloc_buffer = */ ggml_backend_vk_buffer_type_alloc_buffer,
|
|
|
|
/* .get_alignment = */ ggml_backend_vk_buffer_type_get_alignment,
|
|
|
|
/* .get_max_size = */ ggml_backend_vk_buffer_type_get_max_size,
|
|
|
|
/* .get_alloc_size = */ ggml_backend_vk_buffer_type_get_alloc_size,
|
|
|
|
/* .supports_backend = */ ggml_backend_vk_buffer_type_supports_backend,
|
|
|
|
/* .is_host = */ NULL,
|
|
|
|
};
|
|
|
|
|
|
|
|
GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type() {
|
|
|
|
static ggml_backend_buffer_type ggml_backend_vk_buffer_type;
|
|
|
|
|
|
|
|
static bool ggml_backend_vk_buffer_type_initialized = false;
|
|
|
|
|
|
|
|
if (!ggml_backend_vk_buffer_type_initialized) {
|
|
|
|
ggml_backend_vk_buffer_type = {
|
|
|
|
/* .iface = */ ggml_backend_vk_buffer_type_interface,
|
|
|
|
/* .context = */ new ggml_backend_vk_buffer_type_context{GGML_VK_NAME},
|
|
|
|
};
|
|
|
|
ggml_backend_vk_buffer_type_initialized = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return &ggml_backend_vk_buffer_type;
|
|
|
|
}
|
|
|
|
|
|
|
|
// host buffer type
|
|
|
|
|
|
|
|
GGML_CALL static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
|
|
|
|
return GGML_VK_NAME "_Host";
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) {
|
|
|
|
return GGML_VK_NAME "_Host";
|
|
|
|
|
|
|
|
UNUSED(buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
|
|
|
|
ggml_vk_host_free(buffer->context);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
|
2024-01-30 12:59:30 +00:00
|
|
|
void * ptr = nullptr;
|
|
|
|
try {
|
|
|
|
ptr = ggml_vk_host_malloc(size);
|
|
|
|
} catch (vk::SystemError& e) {
|
|
|
|
std::cerr << "ggml_vulkan: Failed to allocate pinned memory." << std::endl;
|
|
|
|
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
// fallback to cpu buffer
|
|
|
|
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
|
|
|
|
buffer->buft = buft;
|
|
|
|
buffer->iface.get_name = ggml_backend_vk_host_buffer_name;
|
|
|
|
buffer->iface.free_buffer = ggml_backend_vk_host_buffer_free_buffer;
|
|
|
|
|
|
|
|
return buffer;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
|
|
|
|
return vk_device.properties.limits.minMemoryMapAlignment;
|
|
|
|
|
|
|
|
UNUSED(buft);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
|
|
|
|
static struct ggml_backend_buffer_type ggml_backend_vk_buffer_type_host = {
|
|
|
|
/* .iface = */ {
|
|
|
|
/* .get_name = */ ggml_backend_vk_host_buffer_type_name,
|
|
|
|
/* .alloc_buffer = */ ggml_backend_vk_host_buffer_type_alloc_buffer,
|
|
|
|
/* .get_alignment = */ ggml_backend_vk_host_buffer_type_get_alignment,
|
|
|
|
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
|
|
|
|
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
|
|
|
|
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
|
|
|
|
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
|
|
|
|
},
|
|
|
|
/* .context = */ nullptr,
|
|
|
|
};
|
|
|
|
|
|
|
|
return &ggml_backend_vk_buffer_type_host;
|
|
|
|
}
|
|
|
|
|
|
|
|
// backend
|
|
|
|
|
|
|
|
GGML_CALL static const char * ggml_backend_vk_name(ggml_backend_t backend) {
|
|
|
|
ggml_backend_vk_context * vk_ctx = (ggml_backend_vk_context *)backend->context;
|
|
|
|
|
|
|
|
return vk_ctx->name.c_str();
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_free(ggml_backend_t backend) {
|
|
|
|
ggml_backend_vk_context * vk_ctx = (ggml_backend_vk_context *)backend->context;
|
|
|
|
|
|
|
|
delete vk_ctx;
|
|
|
|
delete backend;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) {
|
|
|
|
return ggml_backend_vk_buffer_type();
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_set_tensor_async(" << size << ")" << std::endl;
|
|
|
|
#endif
|
2024-01-30 12:59:30 +00:00
|
|
|
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type() || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
if (vk_ctx == nullptr) {
|
|
|
|
// Initialize new transfer context
|
|
|
|
vk_ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(vk_ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_buffer_write_async(vk_ctx, &extra->buffer_gpu, extra->offset + offset, data, size);
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_get_tensor_async(" << size << ")" << std::endl;
|
|
|
|
#endif
|
2024-01-30 12:59:30 +00:00
|
|
|
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type() || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
if (vk_ctx == nullptr) {
|
|
|
|
// Initialize new transfer context
|
|
|
|
vk_ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(vk_ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_buffer_read_async(vk_ctx, &extra->buffer_gpu, extra->offset + offset, data, size);
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_cpy_tensor_async()" << std::endl;
|
|
|
|
#endif
|
2024-01-30 12:59:30 +00:00
|
|
|
if ((dst->buffer->buft == ggml_backend_vk_buffer_type() || dst->buffer->buft == ggml_backend_vk_host_buffer_type()) && ggml_backend_buffer_is_vk(src->buffer)) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra;
|
|
|
|
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
|
|
|
|
|
|
|
|
if (vk_ctx == nullptr) {
|
|
|
|
// Initialize new transfer context
|
|
|
|
vk_ctx = ggml_vk_create_context(vk_device.transfer_queue);
|
|
|
|
ggml_vk_ctx_begin(vk_ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_buffer_copy_async(vk_ctx, &src_extra->buffer_gpu, src_extra->offset, &dst_extra->buffer_gpu, dst_extra->offset, ggml_nbytes(src));
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
return false;
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
|
|
|
|
#ifdef VK_DEBUG
|
|
|
|
std::cerr << "ggml_backend_vk_synchronize()" << std::endl;
|
|
|
|
#endif
|
|
|
|
if(vk_ctx == nullptr) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_ctx_end(vk_ctx);
|
|
|
|
|
|
|
|
for (auto& cpy : vk_ctx->in_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_submit(vk_ctx, vk_fence);
|
|
|
|
VK_CHECK(vk_device.device.waitForFences({ vk_fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences");
|
|
|
|
vk_device.device.resetFences({ vk_fence });
|
|
|
|
|
|
|
|
for (auto& cpy : vk_ctx->out_memcpys) {
|
|
|
|
memcpy(cpy.dst, cpy.src, cpy.n);
|
|
|
|
}
|
|
|
|
|
|
|
|
vk_ctx = nullptr;
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
|
|
|
|
// ggml_backend_vk_context * vk_ctx = (ggml_backend_vk_context *)backend->context;
|
|
|
|
|
|
|
|
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
|
|
ggml_vk_preallocate_buffers_graph(cgraph->nodes[i]);
|
|
|
|
}
|
|
|
|
ggml_vk_preallocate_buffers();
|
|
|
|
|
2024-01-31 10:44:19 +00:00
|
|
|
int last_node = cgraph->n_nodes - 1;
|
|
|
|
|
|
|
|
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
|
|
|
|
while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_GPU) {
|
|
|
|
last_node -= 1;
|
|
|
|
}
|
|
|
|
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
for (int i = 0; i < cgraph->n_nodes; i++) {
|
2024-01-31 10:44:19 +00:00
|
|
|
ggml_vk_build_graph(cgraph->nodes[i], i == last_node);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
ggml_compute_params params = {};
|
|
|
|
params.type = GGML_TASK_COMPUTE;
|
|
|
|
params.ith = 0;
|
|
|
|
for (int i = 0; i < cgraph->n_nodes; i++) {
|
|
|
|
ggml_tensor * node = cgraph->nodes[i];
|
|
|
|
|
|
|
|
if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
bool ok = ggml_vk_compute_forward(¶ms, node);
|
|
|
|
if (!ok) {
|
|
|
|
fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
|
|
|
|
}
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
|
|
|
else {
|
|
|
|
ggml_vk_check_results_1(¶ms, node);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
GGML_ASSERT(ok);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_graph_cleanup();
|
|
|
|
|
|
|
|
return true;
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
|
|
|
|
switch (op->op) {
|
|
|
|
case GGML_OP_UNARY:
|
|
|
|
switch (ggml_get_unary_op(op)) {
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case GGML_OP_MUL_MAT:
|
|
|
|
{
|
|
|
|
struct ggml_tensor * a;
|
|
|
|
struct ggml_tensor * b;
|
|
|
|
if (op->op == GGML_OP_MUL_MAT) {
|
|
|
|
a = op->src[0];
|
|
|
|
b = op->src[1];
|
|
|
|
} else {
|
|
|
|
a = op->src[2];
|
|
|
|
b = op->src[1];
|
|
|
|
}
|
|
|
|
if (a->ne[3] != b->ne[3]) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
} break;
|
|
|
|
// case GGML_OP_GET_ROWS:
|
|
|
|
// {
|
|
|
|
// switch (op->src[0]->type) {
|
|
|
|
// case GGML_TYPE_F16:
|
|
|
|
// case GGML_TYPE_F32:
|
|
|
|
// case GGML_TYPE_Q4_0:
|
|
|
|
// case GGML_TYPE_Q4_1:
|
|
|
|
// case GGML_TYPE_Q5_0:
|
|
|
|
// case GGML_TYPE_Q5_1:
|
|
|
|
// case GGML_TYPE_Q8_0:
|
|
|
|
// return true;
|
|
|
|
// default:
|
|
|
|
// return false;
|
|
|
|
// }
|
|
|
|
// } break;
|
|
|
|
case GGML_OP_CPY:
|
|
|
|
{
|
|
|
|
ggml_type src0_type = op->src[0]->type;
|
|
|
|
ggml_type src1_type = op->src[1]->type;
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
} break;
|
|
|
|
// case GGML_OP_DUP:
|
|
|
|
// case GGML_OP_REPEAT:
|
|
|
|
// {
|
|
|
|
// ggml_type src0_type = op->src[0]->type;
|
|
|
|
// return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
|
|
|
|
// } break;
|
|
|
|
case GGML_OP_ROPE:
|
|
|
|
{
|
|
|
|
const int mode = ((const int32_t *) op->op_params)[2];
|
|
|
|
const bool is_glm = mode & 4;
|
|
|
|
|
|
|
|
return !is_glm;
|
|
|
|
} break;
|
|
|
|
case GGML_OP_NONE:
|
|
|
|
case GGML_OP_RESHAPE:
|
|
|
|
case GGML_OP_VIEW:
|
|
|
|
case GGML_OP_PERMUTE:
|
|
|
|
case GGML_OP_TRANSPOSE:
|
|
|
|
case GGML_OP_NORM:
|
|
|
|
case GGML_OP_ADD:
|
|
|
|
case GGML_OP_MUL:
|
|
|
|
case GGML_OP_RMS_NORM:
|
|
|
|
case GGML_OP_SCALE:
|
|
|
|
case GGML_OP_SQR:
|
|
|
|
case GGML_OP_CLAMP:
|
|
|
|
case GGML_OP_CONT:
|
|
|
|
case GGML_OP_DIAG_MASK_INF:
|
|
|
|
case GGML_OP_SOFT_MAX:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
UNUSED(backend);
|
|
|
|
}
|
|
|
|
|
|
|
|
// TODO: enable async and synchronize
|
|
|
|
static ggml_backend_i ggml_backend_vk_interface = {
|
|
|
|
/* .get_name = */ ggml_backend_vk_name,
|
|
|
|
/* .free = */ ggml_backend_vk_free,
|
|
|
|
/* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
|
|
|
|
/* .set_tensor_async = */ NULL, // ggml_backend_vk_set_tensor_async,
|
|
|
|
/* .get_tensor_async = */ NULL, // ggml_backend_vk_get_tensor_async,
|
|
|
|
/* .cpy_tensor_async = */ NULL, // ggml_backend_vk_cpy_tensor_async,
|
|
|
|
/* .synchronize = */ NULL, // ggml_backend_vk_synchronize,
|
|
|
|
/* .graph_plan_create = */ NULL,
|
|
|
|
/* .graph_plan_free = */ NULL,
|
|
|
|
/* .graph_plan_compute = */ NULL,
|
|
|
|
/* .graph_compute = */ ggml_backend_vk_graph_compute,
|
|
|
|
/* .supports_op = */ ggml_backend_vk_supports_op,
|
|
|
|
};
|
|
|
|
|
|
|
|
GGML_CALL ggml_backend_t ggml_backend_vk_init() {
|
|
|
|
ggml_vk_init(); // TODO: remove from ggml.c
|
|
|
|
|
|
|
|
ggml_backend_vk_context * ctx = new ggml_backend_vk_context {
|
|
|
|
/* .name = */ GGML_VK_NAME,
|
|
|
|
};
|
|
|
|
|
|
|
|
ggml_backend_t vk_backend = new ggml_backend {
|
|
|
|
/* .interface = */ ggml_backend_vk_interface,
|
|
|
|
/* .context = */ ctx
|
|
|
|
};
|
|
|
|
|
|
|
|
return vk_backend;
|
|
|
|
}
|
|
|
|
|
|
|
|
GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend) {
|
|
|
|
return backend && backend->iface.get_name == ggml_backend_vk_name;
|
|
|
|
}
|
|
|
|
|
|
|
|
// backend registry
|
|
|
|
GGML_CALL static ggml_backend_t ggml_backend_reg_vk_init(const char * params, void * user_data) {
|
|
|
|
ggml_backend_t vk_backend = ggml_backend_vk_init();
|
|
|
|
return vk_backend;
|
|
|
|
|
|
|
|
UNUSED(params);
|
|
|
|
UNUSED(user_data);
|
|
|
|
}
|
|
|
|
|
|
|
|
extern "C" GGML_CALL int ggml_backend_vk_reg_devices();
|
|
|
|
|
|
|
|
GGML_CALL int ggml_backend_vk_reg_devices() {
|
|
|
|
ggml_backend_register(GGML_VK_NAME, ggml_backend_reg_vk_init, ggml_backend_vk_buffer_type(), nullptr);
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
// checks
|
|
|
|
|
|
|
|
#ifdef GGML_VULKAN_CHECK_RESULTS
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_print_graph_origin(const ggml_tensor * tensor, std::vector<const ggml_tensor *>& done, int level = 0) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
if (std::find(done.begin(), done.end(), tensor) != done.end() || level > 10) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
for (int j = 0; j < level; j++) {
|
|
|
|
std::cerr << " ";
|
|
|
|
}
|
|
|
|
std::cerr << ggml_op_name(tensor->op) << " gpu=" << (tensor->extra != nullptr) << " backend=" << tensor->backend << std::endl;
|
|
|
|
|
|
|
|
done.push_back(tensor);
|
|
|
|
|
|
|
|
for (int i = 0; i < GGML_MAX_SRC; i++) {
|
|
|
|
if (tensor->src[i] != nullptr) {
|
|
|
|
ggml_vk_print_graph_origin(tensor->src[i], done, level + 1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * data, int i0, int i1, int i2, int i3) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
|
|
|
|
return;
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
i0 = std::max(i0, 5);
|
|
|
|
i1 = std::max(i1, 5);
|
|
|
|
i2 = std::max(i2, 0);
|
|
|
|
i3 = std::max(i3, 0);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
fprintf(stderr, " ");
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
fprintf(stderr, "%7d ", idx1);
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
|
|
|
|
fprintf(stderr, "%7d: ", idx0);
|
|
|
|
for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
|
|
|
|
if (idx0 >= 0 && idx0 < tensor->ne[0] && idx1 >= 0 && idx1 < tensor->ne[1] && i2 >= 0 && i2 < tensor->ne[2] && i3 >= 0 && i3 < tensor->ne[3]) {
|
|
|
|
float val;
|
|
|
|
if (tensor->type == GGML_TYPE_F32) {
|
2024-01-30 12:59:30 +00:00
|
|
|
val = *(const float *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
} else if (tensor->type == GGML_TYPE_F16) {
|
2024-01-30 12:59:30 +00:00
|
|
|
val = ggml_fp16_to_fp32(*(const ggml_fp16_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
fprintf(stderr, "% 7.2f ", val);
|
|
|
|
} else {
|
|
|
|
fprintf(stderr, " ");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
fprintf(stderr, "\n");
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
void * tensor_data = tensor->data;
|
|
|
|
|
|
|
|
if (tensor->backend == GGML_BACKEND_GPU) {
|
|
|
|
const size_t tensor_size = ggml_nbytes(tensor);
|
|
|
|
tensor_data = malloc(tensor_size);
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, extra->offset, tensor_data, tensor_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl;
|
|
|
|
std::cerr << "tensor=" << tensor << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
|
|
|
|
if (tensor->src[0] != nullptr) {
|
|
|
|
std::cerr << "tensor->src[0]=" << tensor->src[0] << " name=" << tensor->src[0]->name << " op=" << ggml_op_name(tensor->src[0]->op) << " type=" << ggml_type_name(tensor->src[0]->type) << " backend=" << tensor->src[0]->backend << " ne0=" << tensor->src[0]->ne[0] << " nb0=" << tensor->src[0]->nb[0] << " ne1=" << tensor->src[0]->ne[1] << " nb1=" << tensor->src[0]->nb[1] << " ne2=" << tensor->src[0]->ne[2] << " nb2=" << tensor->src[0]->nb[2] << " ne3=" << tensor->src[0]->ne[3] << " nb3=" << tensor->src[0]->nb[3] << std::endl;
|
|
|
|
}
|
|
|
|
if (tensor->src[1] != nullptr) {
|
|
|
|
std::cerr << "tensor->src[1]=" << tensor->src[1] << " name=" << tensor->src[1]->name << " op=" << ggml_op_name(tensor->src[1]->op) << " type=" << ggml_type_name(tensor->src[1]->type) << " backend=" << tensor->src[1]->backend << " ne0=" << tensor->src[1]->ne[0] << " nb0=" << tensor->src[1]->nb[0] << " ne1=" << tensor->src[1]->ne[1] << " nb1=" << tensor->src[1]->nb[1] << " ne2=" << tensor->src[1]->ne[2] << " nb2=" << tensor->src[1]->nb[2] << " ne3=" << tensor->src[1]->ne[3] << " nb3=" << tensor->src[1]->nb[3] << std::endl;
|
|
|
|
}
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
std::cerr << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(tensor, done);
|
|
|
|
|
|
|
|
if (tensor->backend == GGML_BACKEND_GPU) {
|
|
|
|
free(tensor_data);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_check_tensor(const std::string& name, const ggml_tensor * tensor) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
return;
|
|
|
|
GGML_ASSERT(tensor->backend == GGML_BACKEND_CPU);
|
|
|
|
if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
|
|
|
|
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
|
|
|
|
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
|
|
|
|
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
|
|
|
|
float val = 0.0f;
|
|
|
|
if (tensor->type == GGML_TYPE_F32) {
|
|
|
|
val = *(float *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
|
|
|
|
} else if (tensor->type == GGML_TYPE_F16) {
|
|
|
|
val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
|
|
|
|
}
|
|
|
|
if (std::isnan(val)) {
|
|
|
|
std::cerr << "ERROR: TENSOR CHECK " << name << ": Invalid value in " << ggml_op_name(tensor->op) << " i3=" << i3 << " i2=" << i2 << " i1=" << i1 << " i0=" << i0 << " val=" << val << std::endl;
|
|
|
|
std::cerr << "tensor=" << tensor << " tensor->type=" << ggml_type_name(tensor->type) << " tensor->backend: " << tensor->backend << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
|
|
|
|
std::cerr << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, tensor->data, i0, i1, i2, i3);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(tensor, done);
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void * comp_result;
|
|
|
|
size_t comp_size;
|
|
|
|
size_t comp_nb[GGML_MAX_DIMS];
|
|
|
|
size_t check_counter = 0;
|
2024-01-30 12:59:30 +00:00
|
|
|
static void ggml_vk_check_results_0(ggml_compute_params * params, ggml_tensor * tensor) {
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
if (params->ith != 0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
check_counter++;
|
|
|
|
if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor * src0 = tensor->src[0];
|
|
|
|
ggml_tensor * src1 = tensor->src[1];
|
|
|
|
|
|
|
|
struct ggml_init_params iparams = {
|
2024-01-30 12:59:30 +00:00
|
|
|
/*.mem_size =*/ 1024*1024*1024,
|
|
|
|
/*.mem_buffer =*/ NULL,
|
|
|
|
/*.no_alloc =*/ false,
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
struct ggml_context * ctx = ggml_init(iparams);
|
|
|
|
|
|
|
|
struct ggml_tensor * src0_clone = nullptr;
|
|
|
|
struct ggml_tensor * src1_clone = nullptr;
|
|
|
|
struct ggml_tensor * tensor_clone = nullptr;
|
|
|
|
|
|
|
|
size_t src0_size;
|
|
|
|
size_t src1_size;
|
|
|
|
|
|
|
|
void * src0_buffer;
|
|
|
|
void * src1_buffer;
|
|
|
|
|
|
|
|
if (src0 != nullptr) {
|
|
|
|
src0_clone = ggml_dup_tensor(ctx, src0);
|
|
|
|
|
|
|
|
src0_size = ggml_nbytes(src0);
|
|
|
|
|
|
|
|
src0_buffer = malloc(src0_size);
|
|
|
|
src0_clone->data = src0_buffer;
|
|
|
|
if (src0->backend == GGML_BACKEND_CPU) {
|
|
|
|
memcpy(src0_clone->data, src0->data, src0_size);
|
|
|
|
memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
|
|
|
|
} else if (src0->backend == GGML_BACKEND_GPU) {
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra;
|
|
|
|
uint64_t offset = extra->offset;
|
|
|
|
if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) {
|
|
|
|
for (int i3 = 0; i3 < src0->ne[3]; i3++) {
|
|
|
|
for (int i2 = 0; i2 < src0->ne[2]; i2++) {
|
|
|
|
const int idx = i3*src0->ne[2] + i2;
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
src0_clone->nb[0] = src0->nb[0];
|
|
|
|
src0_clone->nb[1] = src0->nb[1];
|
|
|
|
for (int i = 2; i < GGML_MAX_DIMS; i++) {
|
|
|
|
src0_clone->nb[i] = src0_clone->nb[i - 1]*src0_clone->ne[i - 1];
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (offset + src0_size >= extra->buffer_gpu.size) {
|
|
|
|
src0_size = extra->buffer_gpu.size - offset;
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, offset, src0_clone->data, src0_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
|
|
|
|
ggml_vk_print_tensor(src0, "src0");
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src0", src0_clone);
|
|
|
|
}
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
src1_clone = ggml_dup_tensor(ctx, src1);
|
|
|
|
|
|
|
|
src1_size = ggml_nbytes(src1);
|
|
|
|
|
|
|
|
src1_buffer = malloc(src1_size);
|
|
|
|
src1_clone->data = src1_buffer;
|
|
|
|
if (src1->backend == GGML_BACKEND_CPU) {
|
|
|
|
memcpy(src1_clone->data, src1->data, src1_size);
|
|
|
|
memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
|
|
|
|
} else if (src1->backend == GGML_BACKEND_GPU) {
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra;
|
|
|
|
uint64_t offset = extra->offset;
|
|
|
|
if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) {
|
|
|
|
for (int i3 = 0; i3 < src1->ne[3]; i3++) {
|
|
|
|
for (int i2 = 0; i2 < src1->ne[2]; i2++) {
|
|
|
|
const int idx = i3*src1->ne[2] + i2;
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
src1_clone->nb[0] = src1->nb[0];
|
|
|
|
src1_clone->nb[1] = src1->nb[1];
|
|
|
|
for (int i = 2; i < GGML_MAX_DIMS; i++) {
|
|
|
|
src1_clone->nb[i] = src1_clone->nb[i - 1]*src1_clone->ne[i - 1];
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
if (offset + src1_size >= extra->buffer_gpu.size) {
|
|
|
|
src1_size = extra->buffer_gpu.size - offset;
|
|
|
|
}
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, offset, src1_clone->data, src1_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
|
|
|
|
ggml_vk_print_tensor(src1, "src1");
|
|
|
|
std::cerr << "TENSOR CHECK: " << ggml_op_name(src1_clone->op) << " (check " << check_counter << ")" << std::endl;
|
|
|
|
std::cerr << "src1_clone=" << tensor << " src1_clone->backend: " << src1_clone->backend << " src1_clone->type: " << ggml_type_name(src1_clone->type) << " ne0=" << src1_clone->ne[0] << " nb0=" << src1_clone->nb[0] << " ne1=" << src1_clone->ne[1] << " nb1=" << src1_clone->nb[1] << " ne2=" << src1_clone->ne[2] << " nb2=" << src1_clone->nb[2] << " ne3=" << src1_clone->ne[3] << " nb3=" << src1_clone->nb[3] << std::endl;
|
|
|
|
if (src1->src[0] != nullptr) {
|
|
|
|
std::cerr << "src1->src[0]=" << src1->src[0] << " op=" << ggml_op_name(src1->src[0]->op) << " type=" << ggml_type_name(src1->src[0]->type) << " backend=" << src1->src[0]->backend << " ne0=" << src1->src[0]->ne[0] << " nb0=" << src1->src[0]->nb[0] << " ne1=" << src1->src[0]->ne[1] << " nb1=" << src1->src[0]->nb[1] << " ne2=" << src1->src[0]->ne[2] << " nb2=" << src1->src[0]->nb[2] << " ne3=" << src1->src[0]->ne[3] << " nb3=" << src1->src[0]->nb[3] << std::endl;
|
|
|
|
}
|
|
|
|
if (src1->src[1] != nullptr) {
|
|
|
|
std::cerr << "src1->src[1]=" << src1->src[1] << " op=" << ggml_op_name(src1->src[1]->op) << " type=" << ggml_type_name(src1->src[1]->type) << " backend=" << src1->src[1]->backend << " ne0=" << src1->src[1]->ne[0] << " nb0=" << src1->src[1]->nb[0] << " ne1=" << src1->src[1]->ne[1] << " nb1=" << src1->src[1]->nb[1] << " ne2=" << src1->src[1]->ne[2] << " nb2=" << src1->src[1]->nb[2] << " ne3=" << src1->src[1]->ne[3] << " nb3=" << src1->src[1]->nb[3] << std::endl;
|
|
|
|
}
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 0, 0);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 1, 0);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(src1_clone, done);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src1", src1_clone);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (tensor->op == GGML_OP_MUL_MAT) {
|
|
|
|
tensor_clone = ggml_mul_mat(ctx, src0_clone, src1_clone);
|
|
|
|
} else if (tensor->op == GGML_OP_MUL) {
|
|
|
|
tensor_clone = ggml_mul(ctx, src0_clone, src1_clone);
|
|
|
|
} else if (tensor->op == GGML_OP_SCALE) {
|
|
|
|
tensor_clone = ggml_scale(ctx, src0_clone, ((float *)tensor->op_params)[0]);
|
|
|
|
} else if (tensor->op == GGML_OP_SQR) {
|
|
|
|
tensor_clone = ggml_sqr(ctx, src0_clone);
|
|
|
|
} else if (tensor->op == GGML_OP_CLAMP) {
|
|
|
|
tensor_clone = ggml_clamp(ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
|
|
|
|
} else if (tensor->op == GGML_OP_ADD) {
|
|
|
|
tensor_clone = ggml_add(ctx, src0_clone, src1_clone);
|
|
|
|
} else if (tensor->op == GGML_OP_NORM) {
|
|
|
|
tensor_clone = ggml_norm(ctx, src0_clone, *(float *)tensor->op_params);
|
|
|
|
} else if (tensor->op == GGML_OP_RMS_NORM) {
|
|
|
|
tensor_clone = ggml_rms_norm(ctx, src0_clone, *(float *)tensor->op_params);
|
|
|
|
} else if (tensor->op == GGML_OP_SOFT_MAX) {
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
tensor_clone = ggml_soft_max_ext(ctx, src0_clone, src1_clone, *(float *)tensor->op_params);
|
|
|
|
} else {
|
|
|
|
tensor_clone = ggml_soft_max(ctx, src0_clone);
|
|
|
|
}
|
|
|
|
} else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
|
|
|
|
tensor_clone = ggml_diag_mask_inf(ctx, src0_clone, *(float *)tensor->op_params);
|
|
|
|
} else if (tensor->op == GGML_OP_ROPE) {
|
|
|
|
const int n_dims = ((int32_t *) tensor->op_params)[1];
|
|
|
|
const int mode = ((int32_t *) tensor->op_params)[2];
|
|
|
|
const int n_ctx = ((int32_t *) tensor->op_params)[3];
|
|
|
|
const int n_orig_ctx = ((int32_t *) tensor->op_params)[4];
|
|
|
|
float freq_base = ((float *) tensor->op_params)[5];
|
|
|
|
float freq_scale = ((float *) tensor->op_params)[6];
|
|
|
|
float ext_factor = ((float *) tensor->op_params)[7];
|
|
|
|
float attn_factor = ((float *) tensor->op_params)[8];
|
|
|
|
float beta_fast = ((float *) tensor->op_params)[9];
|
|
|
|
float beta_slow = ((float *) tensor->op_params)[10];
|
|
|
|
tensor_clone = ggml_rope_custom(ctx, src0_clone, src1_clone, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
|
|
|
|
} else if (tensor->op == GGML_OP_UNARY) {
|
|
|
|
switch (ggml_get_unary_op(tensor)) {
|
|
|
|
case GGML_UNARY_OP_SILU:
|
|
|
|
tensor_clone = ggml_silu(ctx, src0_clone);
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_GELU:
|
|
|
|
tensor_clone = ggml_gelu(ctx, src0_clone);
|
|
|
|
break;
|
|
|
|
case GGML_UNARY_OP_RELU:
|
|
|
|
tensor_clone = ggml_relu(ctx, src0_clone);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
} else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
|
|
|
|
if (src1 == nullptr) {
|
|
|
|
tensor_clone = ggml_dup(ctx, src0_clone);
|
2024-01-30 12:59:30 +00:00
|
|
|
tensor_clone->type = tensor->type;
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
} else {
|
|
|
|
tensor_clone = ggml_cpy(ctx, src0_clone, src1_clone);
|
|
|
|
}
|
|
|
|
} else if (tensor->op == GGML_OP_CONT) {
|
|
|
|
tensor_clone = ggml_cont_4d(ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
|
|
|
|
} else if (tensor->op == GGML_OP_RESHAPE) {
|
|
|
|
tensor_clone = ggml_reshape_4d(ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
|
|
|
|
} else if (tensor->op == GGML_OP_VIEW) {
|
|
|
|
tensor_clone = ggml_view_4d(ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]);
|
|
|
|
} else if (tensor->op == GGML_OP_PERMUTE) {
|
|
|
|
int32_t * params = (int32_t *)tensor->op_params;
|
|
|
|
tensor_clone = ggml_permute(ctx, src0_clone, params[0], params[1], params[2], params[3]);
|
|
|
|
} else if (tensor->op == GGML_OP_TRANSPOSE) {
|
|
|
|
tensor_clone = ggml_transpose(ctx, src0_clone);
|
|
|
|
} else {
|
|
|
|
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Disable vulkan here to avoid the hooks in ggml.c
|
|
|
|
vk_disable = true;
|
|
|
|
|
|
|
|
ggml_cgraph * cgraph = ggml_new_graph(ctx);
|
|
|
|
ggml_build_forward_expand(cgraph, tensor_clone);
|
|
|
|
|
|
|
|
ggml_graph_compute_with_ctx(ctx, cgraph, 8);
|
|
|
|
|
|
|
|
vk_disable = false;
|
|
|
|
|
|
|
|
ggml_vk_check_tensor(ggml_op_name(tensor->op), tensor_clone);
|
|
|
|
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
|
|
|
|
ggml_vk_print_tensor(tensor_clone, "tensor_clone");
|
|
|
|
}
|
|
|
|
|
|
|
|
comp_size = ggml_nbytes(tensor_clone);
|
|
|
|
|
|
|
|
comp_result = malloc(comp_size);
|
|
|
|
memcpy(comp_result, tensor_clone->data, comp_size);
|
|
|
|
memcpy(comp_nb, tensor_clone->nb, sizeof(size_t) * GGML_MAX_DIMS);
|
|
|
|
|
|
|
|
if (src0 != nullptr) {
|
|
|
|
free(src0_buffer);
|
|
|
|
}
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
free(src1_buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_free(ctx);
|
|
|
|
}
|
|
|
|
|
|
|
|
void ggml_vk_check_results_1(ggml_compute_params * params, ggml_tensor * tensor) {
|
|
|
|
if (params->ith != 0) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
ggml_tensor * src0 = tensor->src[0];
|
|
|
|
ggml_tensor * src1 = tensor->src[1];
|
|
|
|
|
|
|
|
void * tensor_data = tensor->data;
|
|
|
|
|
|
|
|
if (tensor->backend == GGML_BACKEND_GPU) {
|
|
|
|
size_t tensor_size = ggml_nbytes(tensor);
|
|
|
|
tensor_data = malloc(tensor_size);
|
|
|
|
|
|
|
|
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
|
|
|
|
|
|
|
if (extra->offset + tensor_size >= extra->buffer_gpu.size) {
|
|
|
|
tensor_size = extra->buffer_gpu.size - (extra->offset);
|
|
|
|
}
|
|
|
|
|
2024-01-30 12:59:30 +00:00
|
|
|
ggml_vk_buffer_read(&extra->buffer_gpu, extra->offset, tensor_data, tensor_size);
|
ggml : add Vulkan backend (#2059)
* Vulkan loader code
* Fix matmul kernel, continue implementation
* Continue implementation
* Vulkan memory management
* Vulkan development
* Matmul call
* Add aligned malloc and free for VMA
* Continue implementation
* First matmul success
* GEMM Kernel optimization
* 1D Blocktiling
* 2D Blocktiling
* Write coalescing
* Continue vulkan implementation and optimization
* First FP16 attempt, disabled for now
* Code abstraction, FP16 implementation, fix kernel, add FP16 to FP32 kernel
* Enable device extensions properly, restore fp16 matmul op
* Fix mulmat_f16
* Output FP32 in fp16 matmul shader
* Fix f16_to_f32 kernel
* dequant_q4_0 kernel
* Add VMA library
* Avoid requesting dedicated memory, VMA can decide that by itself
* Add bounds checking to matmul kernels, improve implementation, fix command buffers not freed properly
* add cmake commands
* Add 2d write operation, profiling code
* Fix 2d write
* Fix queue selection for AMD RADV
* Fix trailing whitespace in vk_mem_alloc.h
* Add WIP warp tile mat mul shaders
* Disable glslc optimization
* Disable glslc optimization for CMake
* Optimize warptile matmul shader, replace blocktile with it
* Add split-k optimization for small matrix multiplication
Use semaphores for synchronization instead of fences or waitidle
Rework async write/read for synchronization
* Fix validation errors, improve compatibility with AMD GPUs
* Rework command buffer handling
* Variable matmul kernel using specialization constants
* Fix synchronization on AMD, add barriers for buffer ownership transfer, add debug flag and prints
* Reuse semaphores
* Handle stage flags during command buffer submission properly
* Increase matmul test runs for consistent results
* Fix F32 matmul
* Add vectorized loading and zeropadding for matrix multiplication
* Use pinned memory for f16 preprocessing
* Don't force aligned matmul
* Don't free before queue done
* Replace VMA library with native Vulkan buffer management
* Basic offloading support with mul_f32 and dmmv for q4_0
* Run glslc commands in parallel
* Unroll loops in dmmv shader
* Reduce usage of waitIdle
* Reuse pinned allocation for f16 conversion
* Handle devices with only a single queue
* Fix trailing whitespace in CMakeLists.txt
* Allow parallel execution of kernels, parallelize third and fourth dimension calls
* Add fallback for devices only supporting one DescriptorSet per DescriptorPool
* Move to graph function similar to CUDA implementation
* Use F16 kernel for most things, replace q_f32 with mul_mat_q_f16 function
* Add F32 dmmv shaders
* Batch submissions
* Add .spv to gitignore
* Split off matrix vector multiplication for separate optimization
* Use single command buffer for matrix vector multiplication ops
* Reduce overhead of mul_f32 calls by using a single command buffer
* Add submission batching to mul_f32
* Fix tests
* Add missing barrier
* Add further missing barrier
* Add further ops
* Replace vk::QueueFamilyIgnored with VK_QUEUE_FAMILY_IGNORED to support more Vulkan header versions
* Remove unnecessary cblas link
* Fix descriptor set pre-allocation assert
* Add runtime shader compilation, start transferring shaders to this approach
* Transfer remaining shaders to header and compile on runtime
* Fix fp32 fallback if device doesn't support fp16, add force disable env var GGML_VULKAN_DISABLE_F16
* Add support for q4_1, q5_0, q5_1 and q8_0
* Remove unnecessary scalar layout extension
* Parse graph early to pre-record command buffers
* Add q6_k support
* Add multi-submit for command buffers
* Fix q6_k dequant shader for AMD
* Fix q6_k for GPUs without fp16 support
* Simplify q6_k fp16 fix
* Minor fixes
* Fix wg_denom of m-mulmat shaders
* Add Python-based Vulkan shader generator
* Replace shaderc dependency with precompiled shaders
Fix python script to generate shaders
* Clean up code
* Fix shader generator script Windows compatibility
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
* Close file before deletion
* Fix vulkan shader fp32 name
* Add q2_k and q3_k support
Add validation check to compare shader results to cpu results
* Add q4_k support
* Add q5_k support
* Bake SPIR-V bytecode into the library instead of loading shaders from file
* Switch to signal semaphores for flexibility
Prepare broadcasting support for mul mat
* Finish broadcasting mul mat support for GQA
* Clean up unused functions
Add repeat op
* Add further ops, not yet enabled. Improve semaphore code
* Reduce number of used semaphores by utilizing timelines more properly
* Remove queue information
* Reuse timeline semaphores, allow parallel operation with binary semaphores to work around nvidia driver limitations
* Add Vulkan to llama-bench
* Remove cblas dependency
* Fix matmul k-split bug
* Fix q4_k dmmv K_QUANTS_PER_ITERATION 1 shader
* Add RMS Norm shader, rework op_f32 shader setup, fix matmul bug
* Fix issues with float16 overflows in shaders
* Fix issues with older Vulkan headers on Ubuntu 22.04
* Allow multi-op partial offloading by parsing the graph to preallocate enough between-op buffers
* Implement further ops, rework op_f32 calls, fix bugs
* Finish full offloading support, add last remaining ops, fix bugs, remove redundant code
* Upload generated file ggml-vulkan-shaders.hpp, remove redundant shaders
* Merge upstream changes, fix conflicts, adapt soft_max op
* Fix Python and shader header format
* Free model gpu buffers on exit
* Use single queue per device to simplify code
* Add matmul shader support for running multiple calculations in parallel
* Switch from semaphore-synchronized multiple command buffers per op to single command buffer for multiple ops, whole graph if possible
* Fix missing event cast
* Replace uint64_t(-1) with UINT64_MAX, rename function for clarity
* Fix warning about empty C function parameters
* Fix compiler warnings
* Properly implement Vulkan backend buffer handling
* Fix oversized host staging buffers
* Simplify barrier synchronization calls
* Fix gcc warnings
* Implement max_size for backend buffer types to limit the size of a single allocation
* Use min of maxMemoryAllocationSize and maxBufferSize for device max allocation size
* refactor multi buf
* Disable unsupported ops to fix tests
* Check for maintenance4 support before using it
* Handle devices with only a single queue
* Fix single queue logic
* propagate buffer usage in multi buffers
* Implement rope_neox op
* Cleanup header and other files
* Simplify gpu_extras by removing events and putting staging memcpys into contexts
* Move queue into context
Add not-yet-enabled async backend ops
* Simplify context use, optimize matmul shader for warp size 64 (AMD GCN), fix split_k matmul shader optimization
* Add get_max_size to SYCL backend.
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : fix trailing whitespace
---------
Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>
Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-01-28 17:03:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
float first_error_result = -1.0f;
|
|
|
|
float first_error_correct = -1.0f;
|
|
|
|
std::array<int, 4> first_error = { -1, -1, -1, -1 };
|
|
|
|
double avg_err = 0.0;
|
|
|
|
size_t counter = 0;
|
|
|
|
|
|
|
|
for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
|
|
|
|
for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
|
|
|
|
for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
|
|
|
|
for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
|
|
|
|
const bool buffer_size_fit = i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0] < comp_size;
|
|
|
|
float correct = 0.0f;
|
|
|
|
float result = 0.0f;
|
|
|
|
|
|
|
|
if (buffer_size_fit) {
|
|
|
|
if (tensor->type == GGML_TYPE_F32) {
|
|
|
|
correct = *(float *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
|
|
|
|
result = *(float *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
|
|
|
|
} else if (tensor->type == GGML_TYPE_F16) {
|
|
|
|
correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
|
|
|
|
result = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
|
|
|
|
} else {
|
|
|
|
std::cerr << "comp_size=" << comp_size << " but required is " << (i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]) << std::endl;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl;
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
|
|
|
|
if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) {
|
|
|
|
std::cerr << "ERROR: Invalid value in " << ggml_op_name(tensor->op) << " i3=" << i3 << " i2=" << i2 << " i1=" << i1 << " i0=" << i0 << " result=" << result << " correct=" << correct << " avg_err=" << (avg_err / counter) << std::endl;
|
|
|
|
std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
|
|
|
|
if (src0 != nullptr) {
|
|
|
|
std::cerr << "src0=" << src0 << " src0->name=" << src0->name << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
std::cerr << "src1=" << src1 << " src1->name=" << src1->name << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3);
|
|
|
|
std::cerr << std::endl << "Correct:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, comp_result, i0, i1, i2, i3);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(tensor, done);
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
}
|
|
|
|
if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) {
|
|
|
|
first_error[0] = i0;
|
|
|
|
first_error[1] = i1;
|
|
|
|
first_error[2] = i2;
|
|
|
|
first_error[3] = i3;
|
|
|
|
first_error_result = result;
|
|
|
|
first_error_correct = correct;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Special case, value is infinite, avoid NaN result in avg_err
|
|
|
|
// NaN also appears in results, if both are nan error is 0
|
|
|
|
if (!std::isinf(correct) && !std::isinf(result) && !std::isnan(correct) && !std::isnan(result)) {
|
|
|
|
avg_err += std::fabs(correct - result);
|
|
|
|
}
|
|
|
|
counter++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
avg_err /= counter;
|
|
|
|
|
|
|
|
if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
|
|
|
|
std::cerr << "TENSOR CHECK: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
|
|
|
|
std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
|
|
|
|
if (src0 != nullptr) {
|
|
|
|
std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
|
|
|
|
std::cerr << std::endl << "Correct:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 0, 0);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 1, 0);
|
|
|
|
std::cerr << std::endl << "Correct:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 1, 0);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(tensor, done);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (avg_err > 0.05 || std::isnan(avg_err)) {
|
|
|
|
std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
|
|
|
|
std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
|
|
|
|
if (src0 != nullptr) {
|
|
|
|
std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
if (src1 != nullptr) {
|
|
|
|
std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
|
|
|
|
}
|
|
|
|
std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
|
|
|
|
std::cerr << std::endl << "Result:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]);
|
|
|
|
std::cerr << std::endl << "Correct:" << std::endl;
|
|
|
|
ggml_vk_print_tensor_area(tensor, comp_result, first_error[0], first_error[1], first_error[2], first_error[3]);
|
|
|
|
std::cerr << std::endl;
|
|
|
|
std::vector<const ggml_tensor *> done;
|
|
|
|
ggml_vk_print_graph_origin(tensor, done);
|
|
|
|
GGML_ASSERT(false);
|
|
|
|
} else {
|
|
|
|
std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " backend=" << tensor->backend << " avg_err=" << avg_err << std::endl;
|
|
|
|
}
|
|
|
|
|
|
|
|
free(comp_result);
|
|
|
|
comp_result = nullptr;
|
|
|
|
comp_size = 0;
|
|
|
|
|
|
|
|
if (tensor->backend == GGML_BACKEND_GPU) {
|
|
|
|
free(tensor_data);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|