llama.cpp/ggml-kompute.cpp

2020 lines
78 KiB
C++
Raw Normal View History

#include "ggml.h"
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-kompute.h"
// These are generated at build time by cmake custom command
#include "shaderop_scale.h"
#include "shaderop_scale_8.h"
#include "shaderop_add.h"
#include "shaderop_addrow.h"
#include "shaderop_mul.h"
#include "shaderop_silu.h"
#include "shaderop_relu.h"
#include "shaderop_gelu.h"
#include "shaderop_softmax.h"
#include "shaderop_norm.h"
#include "shaderop_rmsnorm.h"
#include "shaderop_diagmask.h"
#include "shaderop_mul_mat_f16.h"
#include "shaderop_mul_mat_q8_0.h"
#include "shaderop_mul_mat_q4_0.h"
#include "shaderop_mul_mat_q4_1.h"
#include "shaderop_mul_mat_q6_k.h"
#include "shaderop_mul_mat_mat_f32.h"
#include "shaderop_getrows_f16.h"
#include "shaderop_getrows_q4_0.h"
#include "shaderop_getrows_q4_1.h"
#include "shaderop_getrows_q6_k.h"
#include "shaderop_rope_f16.h"
#include "shaderop_rope_f32.h"
#include "shaderop_cpy_f16_f16.h"
#include "shaderop_cpy_f16_f32.h"
#include "shaderop_cpy_f32_f16.h"
#include "shaderop_cpy_f32_f32.h"
#include <algorithm>
#include <array>
#include <cassert>
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <iostream>
#include <memory>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include <kompute/Kompute.hpp>
#include <vulkan/vulkan.hpp>
#ifdef __linux__
#include <cstdlib> // for setenv
#endif
#define QK4_0 32
#define QR4_0 2
#define QK4_1 32
#define QK_NL 16
typedef ggml_fp16_t half;
static std::string ggml_kompute_format_name(int device) {
return "Kompute" + std::to_string(device);
}
struct ggml_kompute_context {
int device;
std::string name;
std::shared_ptr<vk::DescriptorPool> pool;
ggml_kompute_context(int device)
: device(device), name(ggml_kompute_format_name(device)) {}
};
// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
// and consolidate the init functions and simplify object lifetime management. As it currently stands,
// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
// is only created when a device is set and vulkan is explicitly turned on.
static ggml_kompute_context *s_kompute_context = nullptr;
class kompute_manager {
kp::Manager *s_mgr = nullptr;
public:
kp::Manager *operator()() {
if (s_mgr && !s_mgr->hasInstance()) {
destroy();
}
if (!s_mgr) {
s_mgr = new kp::Manager;
}
return s_mgr;
}
void destroy() {
delete s_mgr;
s_mgr = nullptr;
}
};
static kompute_manager komputeManager;
struct ggml_vk_memory {
void *data = nullptr;
size_t size = 0;
vk::DeviceMemory *primaryMemory = nullptr;
vk::Buffer *primaryBuffer = nullptr;
vk::DeviceMemory *stagingMemory = nullptr;
vk::Buffer *stagingBuffer = nullptr;
};
#ifdef __linux__
__attribute__((constructor))
static void enable_sam() {
setenv("RADV_PERFTEST", "sam", false);
}
#endif
static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
vk::PhysicalDeviceFeatures availableFeatures;
physical_device.getFeatures(&availableFeatures);
if (!availableFeatures.shaderInt16)
return false;
vk::PhysicalDeviceVulkan11Features availableFeatures11;
vk::PhysicalDeviceVulkan12Features availableFeatures12;
availableFeatures11.pNext = &availableFeatures12;
availableFeatures12.pNext = nullptr;
vk::PhysicalDeviceFeatures2 features2;
features2.pNext = &availableFeatures11;
physical_device.getFeatures2(&features2);
if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
!availableFeatures11.storageBuffer16BitAccess) {
return false;
}
if (!availableFeatures12.storageBuffer8BitAccess ||
!availableFeatures12.uniformAndStorageBuffer8BitAccess ||
!availableFeatures12.shaderFloat16 ||
!availableFeatures12.shaderInt8) {
return false;
}
return true;
}
static const char * ggml_vk_getVendorName(uint32_t vendorID) {
switch (vendorID) {
case 0x10DE:
return "nvidia";
case 0x1002:
return "amd";
case 0x8086:
return "intel";
default:
return "unknown";
}
}
static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
std::vector<ggml_vk_device> results;
if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
return results;
std::vector<vk::PhysicalDevice> physical_devices;
try {
physical_devices = komputeManager()->listDevices();
} catch (vk::SystemError & err) {
std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
return results;
}
uint32_t deviceCount = physical_devices.size();
if (deviceCount == 0)
return results;
std::unordered_map<std::string, size_t> count_by_name;
for (uint32_t i = 0; i < deviceCount; i++) {
const auto & physical_device = physical_devices[i];
VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
if (major < 1 || minor < 2)
continue;
if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
continue;
size_t heapSize = 0;
for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
heapSize = heap.size;
break;
}
}
if (heapSize < memoryRequired)
continue;
auto ext_props = physical_device.enumerateDeviceExtensionProperties();
bool has_maintenance4 = false;
// Check if maintenance4 is supported
for (const auto & properties : ext_props) {
if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
has_maintenance4 = true;
}
}
vk::PhysicalDeviceSubgroupProperties subgroup_props;
vk::PhysicalDeviceProperties2 dev_props2;
vk::PhysicalDeviceMaintenance3Properties dev_props3;
vk::PhysicalDeviceMaintenance4Properties dev_props4;
dev_props2.pNext = &dev_props3;
dev_props3.pNext = &subgroup_props;
if (has_maintenance4) {
subgroup_props.pNext = &dev_props4;
}
physical_device.getProperties2(&dev_props2);
if (subgroup_props.subgroupSize < 32)
continue;
ggml_vk_device d;
d.index = i;
d.type = dev_props.deviceType;
d.heapSize = heapSize;
d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
d.subgroupSize = subgroup_props.subgroupSize;
d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
if (has_maintenance4) {
d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
} else {
d.maxAlloc = dev_props3.maxMemoryAllocationSize;
}
std::string name(dev_props.deviceName);
size_t n_idx = ++count_by_name[name];
if (n_idx > 1) {
name += " (" + std::to_string(n_idx) + ")";
}
d.name = strdup(name.c_str());
results.push_back(d);
}
std::stable_sort(results.begin(), results.end(),
[](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
if (lhs.type != rhs.type) {
if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
}
return lhs.heapSize < rhs.heapSize;
}
);
return results;
}
// public API returns a C-style array
ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count) {
auto devices = ggml_vk_available_devices_internal(memoryRequired);
*count = devices.size();
if (devices.empty()) {
return nullptr;
}
size_t nbytes = sizeof (ggml_vk_device) * (devices.size());
auto * arr = static_cast<ggml_vk_device *>(malloc(nbytes));
memcpy(arr, devices.data(), nbytes);
return arr;
}
static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
devices.erase(
std::remove_if(devices.begin(), devices.end(),
[&targetVendor](const ggml_vk_device& device) {
return device.vendor != targetVendor;
}),
devices.end()
);
}
static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
devices.erase(
std::remove_if(devices.begin(), devices.end(),
[&targetName](const ggml_vk_device& device) {
return device.name != targetName;
}),
devices.end()
);
}
static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
if (name.empty())
return false;
auto devices = ggml_vk_available_devices_internal(memoryRequired);
if (name == "amd" || name == "nvidia" || name == "intel") {
ggml_vk_filterByVendor(devices, name);
} else if (name != "gpu") {
ggml_vk_filterByName(devices, name);
}
if (devices.empty())
return false;
*device = devices.front();
return true;
}
bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
return ggml_vk_get_device(device, memoryRequired, std::string(name));
}
bool ggml_vk_has_vulkan() {
return komputeManager()->hasVulkan();
}
bool ggml_vk_has_device() {
return komputeManager()->hasDevice();
}
ggml_vk_device ggml_vk_current_device() {
if (!komputeManager()->hasDevice())
return ggml_vk_device();
auto devices = ggml_vk_available_devices_internal(0);
ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
GGML_ASSERT(!devices.empty());
return devices.front();
}
static
void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
vk::DescriptorPoolSize(
vk::DescriptorType::eStorageBuffer,
3 * size // Descriptor count is number of possible tensors to pass into an algorithm
)
};
vk::DescriptorPoolCreateInfo descriptorPoolInfo(
vk::DescriptorPoolCreateFlags(),
size, // Max sets
static_cast<uint32_t>(descriptorPoolSizes.size()),
descriptorPoolSizes.data());
ctx->pool = std::make_shared<vk::DescriptorPool>();
vk::Result r = komputeManager()->device()->createDescriptorPool(
&descriptorPoolInfo, nullptr, ctx->pool.get());
if (r != vk::Result::eSuccess)
std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
}
static
void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
if (ctx->pool) {
komputeManager()->device()->destroy(
*ctx->pool,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
ctx->pool = nullptr;
}
}
static
vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
vk::BufferCreateInfo bufferCreateInfo;
bufferCreateInfo.size = size;
bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
vk::BufferUsageFlagBits::eTransferSrc |
vk::BufferUsageFlagBits::eTransferDst;
bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
vk::Buffer *vkBuffer = new vk::Buffer;
vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
if (r != vk::Result::eSuccess)
std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
return vkBuffer;
}
static
vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
uint32_t memoryTypeIndex = -1;
bool memoryTypeIndexFound = false;
vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
if (memoryHeap.size < size) {
continue;
}
if (requirements.memoryTypeBits & (1 << i)) {
if (((memoryProperties.memoryTypes[i]).propertyFlags &
flags) == flags) {
memoryTypeIndex = i;
memoryTypeIndexFound = true;
if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
*isHostVisible = true;
}
break;
}
}
}
if (!memoryTypeIndexFound) {
throw std::runtime_error(
"Memory type index for buffer creation not found");
}
vk::MemoryAllocateInfo allocInfo;
allocInfo.allocationSize = size;
allocInfo.memoryTypeIndex = memoryTypeIndex;
vk::DeviceMemory *vkDeviceMemory = new vk::DeviceMemory;
vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
if (r != vk::Result::eSuccess) {
std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
throw std::runtime_error("Error allocating vulkan memory.");
}
return vkDeviceMemory;
}
static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
// If offset is already aligned, return it directly
if (offset % minStorageBufferOffsetAlignment == 0) {
return offset;
}
// Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
}
static ggml_vk_memory ggml_vk_allocate(size_t size) {
ggml_vk_memory memory;
bool isHostVisible = false;
{
memory.primaryBuffer = ggml_vk_allocate_buffer(size);
vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
if (isHostVisible) {
vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
if (r != vk::Result::eSuccess)
std::cerr << "Error mapping memory" << vk::to_string(r);
}
}
if (!isHostVisible) {
memory.stagingBuffer = ggml_vk_allocate_buffer(size);
vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
vk::MemoryPropertyFlagBits::eHostCoherent |
vk::MemoryPropertyFlagBits::eHostCached;
memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
if (r != vk::Result::eSuccess)
std::cerr << "Error mapping memory" << vk::to_string(r);
}
memory.size = size;
return memory;
}
static void ggml_vk_free_memory(ggml_vk_memory &memory)
{
komputeManager()->device()->destroy(
*memory.primaryBuffer,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
if (memory.stagingBuffer) {
komputeManager()->device()->destroy(
*memory.stagingBuffer,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
}
komputeManager()->device()->freeMemory(
*memory.primaryMemory,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
if (memory.stagingMemory) {
komputeManager()->device()->freeMemory(
*memory.stagingMemory,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
}
}
static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
static
ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
// compatibility with ggml-backend
GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
offset = uint64_t(ioffs);
return buf_ctx;
}
static
const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
uint64_t originalOffset = 0;
auto * res = ggml_vk_find_tensor(t, originalOffset);
if (!res) {
static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
return nullTensor;
}
// Create a tensor whose memory will be composed of our buffers at the correct offset
const size_t nelements = ggml_nelements(t);
size_t nbytes = ggml_nbytes(t);
size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
if (alignedOffset) {
*alignedOffset = originalOffset - vulkanOffset;
nbytes += *alignedOffset;
}
return komputeManager()->tensor(
t->data,
nelements,
nbytes, kp::Tensor::TensorDataTypes::eFloat,
res->primaryMemory, res->primaryBuffer,
res->stagingMemory, res->stagingBuffer,
vulkanOffset);
}
static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
if (size % sizeof(uint32_t) != 0) {
throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
}
const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
size_t count = size / sizeof(uint32_t);
return std::vector<uint32_t>(data_ptr, data_ptr + count);
}
inline static
uint32_t safe_divide(uint32_t a, uint32_t b) {
if (b <= 1) {
return a;
}
if ((a % b) != 0) {
fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
GGML_ASSERT(!"safe_divide result would've had remainder");
}
return a / b;
}
static void ggml_vk_add(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
int32_t ne0,
int32_t nb0, int32_t nb1, int32_t nb2, int32_t nb3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
kp::shader_data::op_add_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00;
int32_t nb00, nb01, nb02, nb03;
int32_t ne10, ne11, ne12, ne13;
int32_t nb10, nb11, nb12, nb13;
int32_t ne0;
int32_t nb0, nb1, nb2, nb3;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_addrow(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
uint32_t size, uint32_t row = 0) {
const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
kp::shader_data::op_addrow_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
uint32_t row;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
row
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__))
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
int32_t ne0,
int32_t nb0, int32_t nb1, int32_t nb2, int32_t nb3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
kp::shader_data::op_mul_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00;
int32_t nb00, nb01, nb02, nb03;
int32_t ne10, ne11, ne12, ne13;
int32_t nb10, nb11, nb12, nb13;
int32_t ne0;
int32_t nb0, nb1, nb2, nb3;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_scale(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t size, float scale) {
const static auto spirv_1 = getSpirvShader(
kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
);
const static auto spirv_8 = getSpirvShader(
kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
);
struct PushConstants {
uint32_t inOff, outOff;
float scale;
} const pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
scale
};
const auto * spirv = &spirv_1;
std::string name(__func__);
if (size % 8 == 0) {
size /= 8;
name += "_8";
spirv = &spirv_8;
}
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_xxlu(
const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t size
) {
struct PushConstants {
uint32_t inOff, outOff;
} const pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_silu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
kp::shader_data::op_silu_comp_spv_len);
ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_relu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
kp::shader_data::op_relu_comp_spv_len);
ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_gelu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
kp::shader_data::op_gelu_comp_spv_len);
ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
}
static void ggml_vk_soft_max(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
float scale
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
kp::shader_data::op_softmax_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
float scale;
int32_t mask;
} pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
scale,
bool(inB)
};
auto & inB_ = inB ? inB : inA;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
// FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
const uint32_t local_x = 32;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB_, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_norm_(
const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
int32_t ne00, int32_t nb01,
int32_t nrows, float epsilon
) {
GGML_ASSERT(nb01%sizeof(float) == 0);
GGML_ASSERT(ne00%sizeof(float) == 0);
struct PushConstants {
uint32_t inOff, outOff;
uint32_t ne00, nb01;
float eps;
} pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
(uint32_t)ne00, (uint32_t)nb01, epsilon
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({(uint32_t)nrows});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_norm(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
kp::shader_data::op_norm_comp_spv_len);
ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_rms_norm(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
kp::shader_data::op_rmsnorm_comp_spv_len);
ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
}
static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t n_past,
int32_t ne00, int32_t ne01, int32_t ne02) {
const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
kp::shader_data::op_diagmask_comp_spv_len);
struct PushConstants {
uint32_t inOff, outOff;
uint32_t n_past;
int32_t ne00, ne01;
} pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
n_past,
ne00, ne01
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__))
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_f16(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
uint32_t nb00, uint32_t nb01, uint32_t nb02,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
uint32_t nb10, uint32_t nb11, uint32_t nb12,
int32_t ne0, int32_t ne1,
uint32_t r2, uint32_t r3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
kp::shader_data::op_mul_mat_f16_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
uint32_t nb00, nb01, nb02;
int32_t ne10, ne11, ne12;
uint32_t nb10, nb11, nb12;
int32_t ne0, ne1;
uint32_t r2, r3;
} pushConsts {
safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
nb00, nb01, nb02,
ne10, ne11, ne12,
nb10, nb11, nb12,
ne0, ne1,
r2, r3
};
const unsigned ny = unsigned((ne11 + 4 - 1)/4);
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
uint32_t nb01, uint32_t nb02,
int32_t ne11, int32_t ne12,
uint32_t nb11, uint32_t nb12,
uint32_t nb1, uint32_t nb2) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02, ne11, ne12;
uint32_t nb01, nb02;
uint32_t nb11, nb12;
uint32_t nb1, nb2;
} pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02, ne11, ne12,
nb01, nb02, nb11, nb12,
nb1, nb2
};
const uint32_t local_x = ggml_vk_current_device().subgroupSize;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
{inA, inB, out}, spirv,
{unsigned(ne01),
unsigned(ne11),
unsigned(std::max(ne12, ne02))
},
{local_x},
{pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01),
unsigned(ne11),
unsigned(std::max(ne12, ne02)),
});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_impl(
const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t ne0, int32_t ne1,
uint32_t r2, uint32_t r3
) {
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
int32_t ne10, ne12;
int32_t ne0, ne1;
uint32_t r2, r3;
} pushConsts {
safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
ne10, ne12,
ne0, ne1,
r2, r3
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_mul_mat_q4_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_mul_mat_q4_1(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_mul_mat_q8_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
static void ggml_vk_mul_mat_q6_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1,
int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne10, ne0, ne1, ne01, gqa;
} pushConsts {
inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne10, ne0, ne1, ne01, ne12/ne02
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_get_rows(
const std::vector<uint32_t>& spirv,
const char * suffix,
unsigned element_size, unsigned qk,
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t nb01, int32_t nb1,
uint32_t size
) {
GGML_ASSERT(nb01%element_size == 0);
GGML_ASSERT(nb1%sizeof(float) == 0);
if (qk) GGML_ASSERT(ne00%qk == 0);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, nb01, nb1;
} pushConsts {
safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, nb01, nb1
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_get_rows_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
kp::shader_data::op_getrows_f16_comp_spv_len);
ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q4_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
kp::shader_data::op_getrows_q4_0_comp_spv_len);
ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q4_1(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
kp::shader_data::op_getrows_q4_1_comp_spv_len);
ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q6_k(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
kp::shader_data::op_getrows_q6_k_comp_spv_len);
ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
}
static void ggml_vk_rope(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_orig_ctx,
float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
int32_t ne01, int32_t ne02, int32_t ne03,
uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
int32_t ne0,
uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
) {
GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
static const auto spirv_f16 = getSpirvShader(
kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len
);
static const auto spirv_f32 = getSpirvShader(
kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len
);
int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
GGML_ASSERT(nb03 % type_size == 0);
GGML_ASSERT(nb02 % type_size == 0);
GGML_ASSERT(nb01 % type_size == 0);
GGML_ASSERT(nb00 % type_size == 0);
GGML_ASSERT(nb3 % type_size == 0);
GGML_ASSERT(nb2 % type_size == 0);
GGML_ASSERT(nb1 % type_size == 0);
GGML_ASSERT(nb0 % type_size == 0);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t n_dims, mode, n_orig_ctx;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
uint32_t nb00, nb01, nb02, nb03;
int32_t ne0;
uint32_t nb0, nb1, nb2, nb3;
} pushConsts {
safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size),
n_dims, mode, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
nb00, nb01, nb02, nb03,
ne0,
nb0, nb1, nb2, nb3
};
auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(
name, s_kompute_context->pool.get(), {inA, inB, out},
src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32,
{unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
);
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_cpy(
const std::vector<uint32_t>& spirv,
uint32_t in_element_size, uint32_t out_element_size,
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
int32_t ne0, int32_t ne1, int32_t ne2,
uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
) {
struct PushConstants {
uint32_t inOff, outOff;
int32_t ne00, ne01, ne02;
uint32_t nb00, nb01, nb02, nb03;
int32_t ne0, ne1, ne2;
uint32_t nb0, nb1, nb2, nb3;
} pushConsts {
safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
ne00, ne01, ne02,
nb00, nb01, nb02, nb03,
ne0, ne1, ne2,
nb0, nb1, nb2, nb3
};
std::string name = std::string(__func__)
+ "_i_" + std::to_string(in_element_size)
+ "_o_" + std::to_string(out_element_size);
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name))
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_cpy_f32_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
kp::shader_data::op_cpy_f32_f16_comp_spv_len);
ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f32_f32(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
kp::shader_data::op_cpy_f32_f32_comp_spv_len);
ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f16_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
kp::shader_data::op_cpy_f16_f16_comp_spv_len);
ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f16_f32(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
kp::shader_data::op_cpy_f16_f32_comp_spv_len);
ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
}
static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
break;
default:
return false;
}
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_SILU:
return true;
default:
;
}
break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_SCALE:
case GGML_OP_SOFT_MAX:
case GGML_OP_RMS_NORM:
case GGML_OP_NORM:
case GGML_OP_ROPE:
return true;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
switch (op->src[0]->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
break;
default:
return false;
}
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
break;
default:
return false;
}
return true;
case GGML_OP_DIAG_MASK_INF:
return op->ne[3] == 1;
case GGML_OP_GET_ROWS:
switch (op->src[0]->type) {
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q6_K:
return op->ne[2] == 1 && op->ne[3] == 1;
default:
;
}
return false;
case GGML_OP_MUL_MAT:
if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
return false;
switch (op->src[0]->type) {
case GGML_TYPE_F32:
case GGML_TYPE_Q6_K:
return op->ne[3] == 1;
case GGML_TYPE_F16:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return true;
default:
;
}
default:
;
}
return false;
}
static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
const int n_seq = 8;
// FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
// it to the size of the graph, but I think it can be made smaller?
ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
for (auto& sequence : sequences) {
sequence = komputeManager()->sequence();
}
for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
auto& seq = *sequences[seq_idx];
const int node_start = (seq_idx + 0) * n_nodes_per_seq;
const int node_end = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
bool any_commands_recorded = false;
for (int i = node_start; i < node_end; ++i) {
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 09:16:08 +00:00
struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
struct ggml_tensor * dst = gf->nodes[i];
GGML_ASSERT(dst->data != nullptr);
llama : greatly reduce output buffer memory usage (#6122) * llama : greatly reduce logits memory usage * llama : more compact state saving and reloading * llama : fix lctx.n_outputs not being set before building graph * perplexity : adapt to the logits API changes * perplexity : fix Winogrande, use correct logits for second choice start The first logits used to evaluate the second choice were not from the end of the common prefix; instead, they were the logits from the end of the first choice. This has been corrected. The previous implementation sometimes had outliers in the scores of choices for some tasks, and the logic to skip choices words in the log-likelihood evaluation probably was an attempt to reduce those, but it was complex and didn't quite seem to be the right thing. This is simpler now, and the outlier scores aren't there anymore. * perplexity : normalize spaces and punctuation in Winogrande sentences * llama : fix embedding conditions * llama : fix llama_get_embeddings_ith when the resulting id is 0 * llama : fix wrong n_outputs in llama_set_inputs A mismatch happened when using a smaller n_ubatch than n_batch and then using llama_batch_get_one(). The decision of what n_outputs should be now almost fully depends on how lctx.n_outputs is set in llama_decode_internal. The conditions are simpler this way. * llama : when saving the state, recalculate n_outputs This ensures the correct number of outputs for the entire previous batch is stored in the session file, even when n_ubatch is smaller than n_batch. * llama : fix not-skipping outputs of non-causal models * llama : fix running a batch with n_outputs == 0 It previously worked because lctx.inp_out_ids was not initialized, so it pointed to some garbage address which was somehow still valid when I ran my tests. * llama : keep same graph topology even when n_outputs == 0 * ggml : saner ggml_can_repeat with empty tensors * ggml : future-proof ggml_is_empty by using GGML_MAX_DIMS - 1 * ggml : do not multi-thread ops returning empty tensors * ggml : make ggml_is_empty public and work with views * llama : use a vector for ctx->output_ids * llama : rework reallocation logic for llama_output_reserve Now comparing the actual size with the new total size of the output buffer to allow more efficient enabling and disabling of the embeddings and/or logits output in the future. * ggml : skip empty tensors in all backends * llama : fix llama_output_reserve nullptr deref when new_size is 0 * perplexity : make Winogrande work as it does on master The problems with the Winogrande implementation will need to be fixed in a separate PR to ease review. * llama : clearer error messages for invalid logits or embeddings ids * llama : assert all models that can have inp_out_ids Since the graph topology is now constant, this presence check can be done even when there are no outputs. * llama : assert logits and embd buffers exist before writing to them * llama : handle errors from llama_output_reserve at call sites * perplexity : make hellaswag and multiple-choice outputs identical to master Due to how the KV cache is updated, the logprobs for tokens in a batch are very slightly affected by the other tokens present in the batch, so to make hellaswag and multiple-choice return exactly the same results as on master, the last token of each sequence needs to be evaluated even though its output is not used at all. This will probably be changed back in the future to make these benchmarks a tiny bit faster. * perplexity : fix division by zero when using less than 100 multiple-choice tasks * llama : allow loading state saved with a different ctx size When loading a session file, the context size is now only required to be at least enough to load the KV cells contained in that session file, instead of requiring to use exactly the same context size as when saving. Doing this enables the use-case of extending or shrinking the context size of a saved session. This breaks existing session files because the meaning of kv_buf_size is slightly changed (previously it was the size of the whole KV cache, now it's only the size of the saved part of it). This allows for finer-grained sanity checks when loading in an effort to keep kv_buf_size useful even when the kv_size is changed. * llama : minor ggml-ci * readme : update recent API changes, and warn about Vulkan --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2024-03-26 14:46:41 +00:00
if (ggml_is_empty(dst)) {
continue;
}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
continue; // noop -> next node
default:
break;
}
any_commands_recorded = true;
if (!ggml_vk_supports_op(dst)) {
fprintf(stderr, "%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
GGML_ASSERT(!"unsupported op");
}
const int32_t ne00 = src0 ? src0->ne[0] : 0;
const int32_t ne01 = src0 ? src0->ne[1] : 0;
const int32_t ne02 = src0 ? src0->ne[2] : 0;
const int32_t ne03 = src0 ? src0->ne[3] : 0;
const uint32_t nb00 = src0 ? src0->nb[0] : 0;
const uint32_t nb01 = src0 ? src0->nb[1] : 0;
const uint32_t nb02 = src0 ? src0->nb[2] : 0;
const uint32_t nb03 = src0 ? src0->nb[3] : 0;
const int32_t ne10 = src1 ? src1->ne[0] : 0;
const int32_t ne11 = src1 ? src1->ne[1] : 0;
const int32_t ne12 = src1 ? src1->ne[2] : 0;
const int32_t ne13 = src1 ? src1->ne[3] : 0;
const uint32_t nb10 = src1 ? src1->nb[0] : 0;
const uint32_t nb11 = src1 ? src1->nb[1] : 0;
const uint32_t nb12 = src1 ? src1->nb[2] : 0;
const uint32_t nb13 = src1 ? src1->nb[3] : 0;
const int32_t ne0 = dst ? dst->ne[0] : 0;
const int32_t ne1 = dst ? dst->ne[1] : 0;
const int32_t ne2 = dst ? dst->ne[2] : 0;
// const int32_t ne3 = dst ? dst->ne[3] : 0;
const uint32_t nb0 = dst ? dst->nb[0] : 0;
const uint32_t nb1 = dst ? dst->nb[1] : 0;
const uint32_t nb2 = dst ? dst->nb[2] : 0;
const uint32_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
uint32_t off_src0 = 0;
uint32_t off_src1 = 0;
uint32_t off_dst = 0;
const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
const std::shared_ptr<kp::Tensor>& id_dst = dst ? ggml_vk_get_tensor(dst, &off_dst) : nullTensor;
switch (dst->op) {
case GGML_OP_ADD:
{
if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
// src1 is a row
ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
} else {
ggml_vk_add(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne03,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
);
}
} break;
case GGML_OP_MUL:
{
ggml_vk_mul(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne03,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
);
} break;
case GGML_OP_SCALE:
{
float scale; memcpy(&scale, dst->op_params, sizeof(float));
ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
} break;
case GGML_OP_UNARY:
{
int64_t n = ggml_nelements(dst);
GGML_ASSERT(n % 4 == 0);
switch (ggml_get_unary_op(gf->nodes[i])) {
case GGML_UNARY_OP_SILU:
{
ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
} break;
case GGML_UNARY_OP_RELU:
{
ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
} break;
case GGML_UNARY_OP_GELU:
{
GGML_ASSERT(n % 8 == 0);
ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
} break;
case GGML_OP_SOFT_MAX:
{
float scale;
float max_bias;
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 09:16:08 +00:00
memcpy(&scale, (float *)dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 09:16:08 +00:00
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
#pragma message("TODO: add ALiBi support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192")
GGML_ASSERT(max_bias == 0.0f);
ggml : add Flash Attention (#5021) * ggml : add ggml_flash_attn_ext API * ggml : fix GQA support in ggml_flash_attn_ext * ggml : online attention (CPU) * metal : initial implementation * metal : f16 precision * metal : reduce branches * metal : specialize for head size * wip : 8 rows per simd group * wip : 4 rows per simd group * wip : template for rows per warp * metal : parallelize across KV size * metal : parallel reduce across heads * metal : efficient flash_attn_f16 implementation * metal : avoid redundant loads of the attention * metal : scale and mask in matrix form * metal : fix comment * llama : avoid ggml_cast, use F32 query * metal : add parallel reduce version (disabled) * metal : move output into local memory + optimize - the result from each simdgroup now stays in the registers - significantly reduced SRAM usage - more efficient skipping of -INF blocks - avoid simdgroup barrier in hot loop - add comments * metal : add tests, fix scaling, support C > 32 * metal : improve precision * ggml : fix f16 mad * metal : minor * metal : support Q > 8 * tests : add ATTN tests * metal : disable buffer allocation logs * tests : more * metal : faster inner loop for C == 32 * metal : fix array initialization * tests : ifdef * ggml : switch to padded F16 mask for ggml_soft_max, ggml_flash_attn_ext * ggml : fix ggml_soft_max mask requirement * cuda : fix soft_max to use correct mask size * cuda : add flash_attn kernel (wip) * metal : optimize softmax for C > 32 * metal : optimize softmax * tests : minor fix * cuda : avoid zeroing fragments * tests : update dims * cuda : fix __hisinf() result check * cuda : avoid warp_reduce for smax * cuda : use int instead of int64_t Noticeably improves performance (thanks to Johannes) * cuda : make loops use the same loop values Thanks Johannes again for the tip * cuda : unroll some of the loops * cuda : avoid __hisinf branches * cuda : use half2 in softmax * cuda : switch to 1 warp for bs > 16 * cuda : speed-up reduce part of the kernel * cuda : unroll Q*K^T loop * cuda : fix -INF block check * cuda : simplify softmax * cuda : fix matrix names * cuda : minor * llama : adapt to F16 KQ_pos * llama : adapt new models to F16 KQ_mask * ggml : fix F16 store (ARM NEON) * llama : fix type of KQ_mask and KQ_pos * ggml : fix CPU soft_max * tests : add hs=256 * cuda : fix build * metal : improve perf via smaller int registers * cuda : adapt soft_max to F16 mask and pos * CUDA: faster FlashAttention, kernel for bs == 1 * 16 cols for Phi-2 * no vec for hs, no hs==256 ncols==32 for Volta * adjust kernel selection logic * 4 warps, 256 stride for all D * no ncols == 64 * Multiple parallel blocks for batch size 1 * fix compile warnings * fix excessive KQ_b loads * fix cmake build * fix KV cache padding, NaN from INFINITY (#6438) * llama : flash_attn cparam + fix defrag * server: support flash_attn param * server: bench: enable flash_attn param * CUDA: refactor host code, dyn. par. blocks * fix flash_attn_vec_f16 race condition * flush softmax exp below threshold to 0 * store temp KQ in registers * Calculate KQ as FP32 if KQV has GGML_PREC_F32 * Add __hgt2_mask implementation for CUDA 11 * fix KQ FP32 precision fpr parallel_blocks > 1 * llama-bench : add -fa,--flash-attn arg * metal : add BS=1 kernel for flash attention (#6508) * metal : add BS=1 kernel for flash attention (wip) * metal : support more than 1 warps * metal : opts * metal : opt * metal : switch to parallel reduce * metal : reduce registers * metal : simplify * metal : initial FA vec kernel * metal : use F32 attention accumulators * batched-bench : add fattn arg * llama : simplify llama_build_kv_store ggml-ci * llama : adapt build_olmo to changes * ggml : fix arm fp16 store on windows * metal : clean-up * metal : clean-up kernel code * metal : minor * tests : remove benchmarks ggml-ci * ggml : fix avx512 const correctness ggml-ci * ggml : fix soft_max with bias on CPU ggml-ci * common : print --flash-attn in help * ggml : fix num dimensions in ggml_flash_attn_ext * llama : force disable flash attention for incompatible models * ggml : ggml_soft_max support F16/F32 mask/pos ggml-ci * cuda : uint -> uint32_t * cuda : "constexpr dim3" -> "const dim3" ggml-ci * cuda : try to fix __hgt2_mask ggml-ci * ggml : add TODO's for F16/F32 mask/pos support in other backends * llama : replace bool need_kq_pos with use_alibi * llama : prep ALiBi support for BERT models ggml-ci * llama : fix n_batch requirements ggml-ci * cont * server : add help for --flash-attn arg * llama : disable FA for AMD * tests : remove TMP_ATTN_BENCH ggml-ci * llama : support save/load state with FA enabled ggml-ci * ci : add CUDA save-load-state tests ggml-ci * llama : llama_kv_cache_clear zeroes data + fix save-load seq ggml-ci * llama : fix copy-paste errors, add TODO * llama : disallow incompatible states * llama : update llama_state_get_size after v_trans field * metal : remove tmp log * llama : add static reminder for llama_state_get_size * metal : fix max nsg ggml-ci * ci : fix arg order ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de> Co-authored-by: Pierrick HYMBERT <pierrick.hymbert@gmail.com>
2024-04-30 09:16:08 +00:00
ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
} break;
case GGML_OP_DIAG_MASK_INF:
{
const int n_past = ((int32_t *)(dst->op_params))[0];
ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
} break;
case GGML_OP_NORM:
{
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
} break;
case GGML_OP_RMS_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
} break;
case GGML_OP_MUL_MAT:
{
GGML_ASSERT(ne00 == ne10);
// TODO: assert that dim2 and dim3 are contiguous
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
const uint32_t r2 = ne12/ne02;
const uint32_t r3 = ne13/ne03;
if (src1t != GGML_TYPE_F32) {
fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
if (ggml_is_transposed(src0) ||
ggml_is_transposed(src1)) {
fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
switch (src0t) {
case GGML_TYPE_F32:
ggml_vk_mul_mat_mat_f32(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
);
break;
case GGML_TYPE_F16:
ggml_vk_mul_mat_f16(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12,
ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q8_0:
ggml_vk_mul_mat_q8_0(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q4_0:
ggml_vk_mul_mat_q4_0(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q4_1:
ggml_vk_mul_mat_q4_1(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q6_K:
ggml_vk_mul_mat_q6_k(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02
);
break;
default: {
fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
}
} break;
case GGML_OP_GET_ROWS:
{
if (src0t == GGML_TYPE_F16) {
ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q4_0) {
ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q4_1) {
ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q6_K) {
ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else {
fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
goto not_implemented;
}
} break;
case GGML_OP_ROPE:
{
GGML_ASSERT(ne10 == ne02);
GGML_ASSERT(src0t == dstt);
// const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
// skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
ggml_vk_rope(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
);
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
switch (src0t) {
case GGML_TYPE_F32:
{
switch (dstt) {
case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
default: goto not_implemented;
}
} break;
case GGML_TYPE_F16:
{
switch (dstt) {
case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
default: goto not_implemented;
} break;
default: goto not_implemented;
}
}
} break;
default: goto not_implemented;
}
continue;
not_implemented: {}
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
//GGML_ASSERT(false);
}
// Evaluate sequence
if (any_commands_recorded) {
seq.evalAsync();
}
}
// Wait for all sequences to finish
for (auto& sequence : sequences) {
if (sequence->isRunning())
sequence->evalAwait();
}
ggml_vk_free_descriptor_pool(ctx);
}
template<>
kp::Tensor::TensorDataTypes
kp::TensorT<half>::dataType()
{
return TensorDataTypes::eFloat;
}
template<>
kp::Tensor::TensorDataTypes
kp::TensorT<uint8_t>::dataType()
{
return TensorDataTypes::eUnsignedInt;
}
////////////////////////////////////////////////////////////////////////////////
// backend interface
struct ggml_backend_kompute_buffer_type_context {
int device;
int device_ref = 0;
uint64_t buffer_alignment;
uint64_t max_alloc;
std::string name;
ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
: device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
};
static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
if (!ctx->device_ref) {
komputeManager()->initializeDevice(
ctx->device, {}, {
"VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
"VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
}
);
}
assert(ggml_vk_has_device());
ctx->device_ref++;
}
static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
assert(ctx->device_ref > 0);
ctx->device_ref--;
if (!ctx->device_ref) {
komputeManager.destroy();
}
}
static const char * ggml_backend_kompute_buffer_get_name(ggml_backend_buffer_t buffer) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buffer->buft->context);
return ctx->name.c_str();
}
static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
auto * memory = (ggml_vk_memory *)buffer->context;
if (ggml_vk_has_device()) {
ggml_vk_free_memory(*memory);
}
delete memory;
}
static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
return ((ggml_vk_memory *)buffer->context)->data;
}
static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_UNUSED(buffer);
const auto res = ggml_vk_get_tensor(tensor);
GGML_ASSERT(res);
memcpy((char *)tensor->data + offset, data, size);
komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
}
static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_UNUSED(buffer);
const auto res = ggml_vk_get_tensor(tensor);
GGML_ASSERT(res);
komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
memcpy(data, (const char *)tensor->data + offset, size);
}
static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
auto * memory = (ggml_vk_memory *)buffer->context;
memset(memory->data, value, buffer->size);
if (memory->stagingBuffer)
komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
}
static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
/* .get_name = */ ggml_backend_kompute_buffer_get_name,
/* .free_buffer = */ ggml_backend_kompute_buffer_free_buffer,
/* .get_base = */ ggml_backend_kompute_buffer_get_base,
/* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_kompute_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_kompute_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_kompute_buffer_clear,
/* .reset = */ NULL,
};
// default buffer type
static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->name.c_str();
}
static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_kompute_device_ref(buft);
auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
}
static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->buffer_alignment;
}
static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->max_alloc;
}
static bool ggml_backend_kompute_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
GGML_UNUSED(buft);
return ggml_backend_is_kompute(backend);
}
static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
/* .get_name = */ ggml_backend_kompute_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_kompute_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_kompute_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_vk_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_kompute_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
static std::vector<ggml_backend_buffer_type> bufts = []() {
std::vector<ggml_backend_buffer_type> vec;
auto devices = ggml_vk_available_devices_internal(0);
vec.reserve(devices.size());
for (const auto & dev : devices) {
vec.push_back({
/* .iface = */ ggml_backend_kompute_buffer_type_interface,
/* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc)
});
}
return vec;
}();
auto it = std::find_if(bufts.begin(), bufts.end(), [device](const ggml_backend_buffer_type & t) {
return device == static_cast<ggml_backend_kompute_buffer_type_context *>(t.context)->device;
});
return it < bufts.end() ? &*it : nullptr;
}
// backend
static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
return ctx->name.c_str();
}
static void ggml_backend_kompute_free(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
assert(ctx == s_kompute_context);
s_kompute_context = nullptr;
if (ctx != nullptr) {
delete ctx;
}
delete backend;
}
static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
return ggml_backend_kompute_buffer_type(ctx->device);
}
static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
ggml_vk_graph_compute(ctx, cgraph);
return GGML_STATUS_SUCCESS;
}
static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
GGML_UNUSED(backend);
return ggml_vk_supports_op(op);
}
static struct ggml_backend_i kompute_backend_i = {
/* .get_name = */ ggml_backend_kompute_name,
/* .free = */ ggml_backend_kompute_free,
/* .get_default_buffer_type = */ ggml_backend_kompute_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_kompute_graph_compute,
/* .supports_op = */ ggml_backend_kompute_supports_op,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_kompute_guid() {
static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
return &guid;
}
ggml_backend_t ggml_backend_kompute_init(int device) {
GGML_ASSERT(s_kompute_context == nullptr);
s_kompute_context = new ggml_kompute_context(device);
ggml_backend_t kompute_backend = new ggml_backend {
/* .guid = */ ggml_backend_kompute_guid(),
/* .interface = */ kompute_backend_i,
/* .context = */ s_kompute_context,
};
return kompute_backend;
}
bool ggml_backend_is_kompute(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
}
static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {
GGML_UNUSED(params);
return ggml_backend_kompute_init(intptr_t(user_data));
}
extern "C" int ggml_backend_kompute_reg_devices();
int ggml_backend_kompute_reg_devices() {
auto devices = ggml_vk_available_devices_internal(0);
for (const auto & device : devices) {
ggml_backend_register(
ggml_kompute_format_name(device.index).c_str(),
ggml_backend_reg_kompute_init,
ggml_backend_kompute_buffer_type(device.index),
reinterpret_cast<void *>(intptr_t(device.index))
);
}
return devices.size();
}