root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

ggml : refactor online repacking (#10446)

Browse Source 
* rename ggml-cpu-aarch64.c to .cpp

* reformat extra cpu backend.

- clean Q4_0_N_M and IQ4_0_N_M
  - remove from "file" tensor type
  - allow only with dynamic repack

- extract cpu extra bufts and convert to C++
  - hbm
  - "aarch64"

- more generic use of extra buffer
  - generalise extra_supports_op
  - new API for "cpu-accel":
     - amx
     - aarch64

* clang-format

* Clean Q4_0_N_M ref

Enable restrict on C++

* add op GGML_OP_MUL_MAT_ID for Q4_0_N_M with runtime repack

* added/corrected control on tensor size for Q4 repacking.

* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* add debug logs on repacks.

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Djip007 2024-12-07 13:37:50 +01:00 committed by GitHub
parent c2a16c0bdb
commit 19d8762ab6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
33 changed files with 1136 additions and 1049 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
Makefile
View File

@ -445,6 +445,10 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
MK_CFLAGS += -march=native -mtune=native
HOST_CXXFLAGS += -march=native -mtune=native
# Usage AMX build test
#MK_CFLAGS += -march=graniterapids -mtune=graniterapids
#HOST_CXXFLAGS += -march=graniterapids -mtune=graniterapids
# Usage AVX-only
#MK_CFLAGS += -mfma -mf16c -mavx
#MK_CXXFLAGS += -mfma -mf16c -mavx
@ -948,7 +952,6 @@ DIR_COMMON = common
OBJ_GGML = \
$(DIR_GGML)/src/ggml.o \
$(DIR_GGML)/src/ggml-aarch64.o \
$(DIR_GGML)/src/ggml-alloc.o \
$(DIR_GGML)/src/ggml-backend.o \
$(DIR_GGML)/src/ggml-backend-reg.o \
@ -956,9 +959,11 @@ OBJ_GGML = \
$(DIR_GGML)/src/ggml-quants.o \
$(DIR_GGML)/src/ggml-threading.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \
$(OBJ_GGML_EXT)
OBJ_LLAMA = \
@ -1098,17 +1103,10 @@ DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
# Default target
all: $(BUILD_TARGETS)
# force c++ build for source file that have same name as c file
# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
# g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
ggml/src/ggml-cpu/ggml-cpu.cpp \
ggml/include/ggml-backend.h \
ggml/include/ggml.h \
ggml/include/ggml-alloc.h \
ggml/src/ggml-backend-impl.h \
ggml/include/ggml-cpu.h \
ggml/src/ggml-impl.h
$(CXX) $(CXXFLAGS) -c $< -o $@
$(DIR_GGML)/%_cpp.o: $(DIR_GGML)/%.cpp
$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
# Rules for building object files
$(DIR_GGML)/%.o: $(DIR_GGML)/%.c

5
Package.swift
View File

@ -10,14 +10,15 @@ var sources = [
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-aarch64.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.cpp",
"ggml/src/ggml-backend-reg.cpp",
"ggml/src/ggml-cpu/ggml-cpu.c",
"ggml/src/ggml-cpu/ggml-cpu.cpp",
"ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
"ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp",
"ggml/src/ggml-cpu/ggml-cpu-hbm.cpp",
"ggml/src/ggml-cpu/ggml-cpu-quants.c",
"ggml/src/ggml-cpu/ggml-cpu-traits.cpp",
"ggml/src/ggml-threading.cpp",
"ggml/src/ggml-quants.c",
]

2
docs/build.md
View File

@ -55,7 +55,7 @@ cmake --build build --config Release
cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
cmake --build build-arm64-windows-llvm-release
```
Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels.
## BLAS Build

2
examples/quantize/README.md
View File

@ -54,8 +54,6 @@ As the models are currently fully loaded into memory, you will need adequate dis
Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format.
*(outdated)*
| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |

3
examples/quantize/quantize.cpp
View File

@ -48,9 +48,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", },
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", },
{ "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", },
{ "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },

17
ggml/include/ggml-cpu.h
View File

@ -103,24 +103,14 @@ extern "C" {
// Internal types and functions exposed for tests and benchmarks
typedef void (*ggml_from_float_to_mat_t)
(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
const void * GGML_RESTRICT y, size_t by, int nrc);
typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc);
struct ggml_type_traits_cpu {
ggml_from_float_t from_float;
ggml_from_float_to_mat_t from_float_to_mat;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
int64_t nrows; // number of rows to process simultaneously
int64_t ncols; // number of columns to process simultaneously
ggml_gemv_t gemv;
ggml_gemm_t gemm;
};
GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
@ -140,13 +130,6 @@ extern "C" {
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
#ifdef GGML_USE_CPU_HBM
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
#ifdef __cplusplus
}
#endif

29
ggml/include/ggml.h
View File

@ -384,15 +384,15 @@ extern "C" {
GGML_TYPE_F64 = 28,
GGML_TYPE_IQ1_M = 29,
GGML_TYPE_BF16 = 30,
GGML_TYPE_Q4_0_4_4 = 31,
GGML_TYPE_Q4_0_4_8 = 32,
GGML_TYPE_Q4_0_8_8 = 33,
// GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
// GGML_TYPE_Q4_0_4_8 = 32,
// GGML_TYPE_Q4_0_8_8 = 33,
GGML_TYPE_TQ1_0 = 34,
GGML_TYPE_TQ2_0 = 35,
GGML_TYPE_IQ4_NL_4_4 = 36,
// GGML_TYPE_IQ4_NL_4_4 = 36,
// GGML_TYPE_IQ4_NL_4_8 = 37,
// GGML_TYPE_IQ4_NL_8_8 = 38,
GGML_TYPE_COUNT,
GGML_TYPE_COUNT = 39,
};
// precision
@ -433,9 +433,6 @@ extern "C" {
GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_M = 23, // except 1d tensors
GGML_FTYPE_MOSTLY_BF16 = 24, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
};
// available tensor operations:
@ -2205,11 +2202,19 @@ extern "C" {
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
#ifdef __cplusplus
// restrict not standard in C++
#define GGML_RESTRICT
#ifdef __cplusplus
// restrict not standard in C++
# if defined(__GNUC__)
# define GGML_RESTRICT __restrict__
# elif defined(__clang__)
# define GGML_RESTRICT __restrict
# elif defined(_MSC_VER)
# define GGML_RESTRICT __restrict
# else
# define GGML_RESTRICT
# endif
#else
#define GGML_RESTRICT restrict
# define GGML_RESTRICT restrict
#endif
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);

4
ggml/src/CMakeLists.txt
View File

@ -220,9 +220,7 @@ add_library(ggml-base
ggml-threading.cpp
ggml-threading.h
ggml-quants.c
ggml-quants.h
ggml-aarch64.c
ggml-aarch64.h)
ggml-quants.h)
target_include_directories(ggml-base PRIVATE .)

129
ggml/src/ggml-aarch64.c
View File

@ -1,129 +0,0 @@
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml-aarch64.h"
#include "ggml-impl.h"
#include "ggml-quants.h"
#include <assert.h>
#define UNUSED GGML_UNUSED
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x4 out;
for (int i = 0; i < 4; i++) {
out.d[i] = in[i].d;
}
const int end = QK4_0 * 2 / blck_size_interleave;
if (blck_size_interleave == 8) {
const uint64_t xor_mask = 0x8888888888888888ULL;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
// Using memcpy to avoid unaligned memory accesses
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
} else if (blck_size_interleave == 4) {
const uint32_t xor_mask = 0x88888888;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint32_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
}
} else {
GGML_ASSERT(false);
}
return out;
}
// interleave 8 block_q4_0s in blocks of blck_size_interleave
// returns an interleaved block_q4_0x8
// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x8 out;
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].d;
}
const int end = QK4_0 * 4 / blck_size_interleave;
const uint64_t xor_mask = 0x8888888888888888ULL;
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
return out;
}
static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
assert(n_per_row % QK4_0 == 0);
const int nb = n_per_row / QK4_0;
void * out_ptr = NULL;
if (nrows_interleaved == 8) {
out_ptr = (block_q4_0x8 *) dst;
}
else if (nrows_interleaved == 4) {
out_ptr = (block_q4_0x4 *) dst;
}
assert(nrows_interleaved <= 8);
block_q4_0 dst_tmp[8];
for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
for (int64_t x = 0; x < nb; x++) {
for (int i = 0; i < nrows_interleaved; i++ ) {
quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
}
if (nrows_interleaved == 8) {
*(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
out_ptr = (block_q4_0x8 *) out_ptr + 1;
}
else if (nrows_interleaved == 4) {
*(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
out_ptr = (block_q4_0x4 *) out_ptr + 1;
}
}
}
return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
}
size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
}
size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
}
size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
}

19
ggml/src/ggml-aarch64.h
View File

@ -1,19 +0,0 @@
#pragma once
#include "ggml.h"
// GGML internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
#ifdef __cplusplus
}
#endif

2
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -2089,7 +2089,7 @@ static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, con
static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
/* .get_name = */ ggml_backend_cann_reg_get_name,
/* .get_device_count = */ ggml_backend_cann_reg_get_device_count,
/* .get_device_get = */ ggml_backend_cann_reg_get_device,
/* .get_device = */ ggml_backend_cann_reg_get_device,
/* .get_proc_address = */ ggml_backend_cann_reg_get_proc_address,
};

88
ggml/src/ggml-common.h
View File

@ -6,7 +6,20 @@
typedef uint16_t ggml_half;
typedef uint32_t ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_AGGR_U
#define GGML_COMMON_AGGR_S
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_CPP)
#include <cstdint>
typedef uint16_t ggml_half;
typedef uint32_t ggml_half2;
// std-c++ allow anonymous unions but some compiler warn on it
#define GGML_COMMON_AGGR_U data
// std-c++ do not allow it.
#define GGML_COMMON_AGGR_S data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_METAL)
@ -15,7 +28,8 @@ typedef uint32_t ggml_half2;
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_AGGR_U
#define GGML_COMMON_AGGR_S
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_CUDA)
@ -29,7 +43,8 @@ typedef half2 ggml_half2;
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_AGGR_U
#define GGML_COMMON_AGGR_S data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_HIP)
@ -39,7 +54,8 @@ typedef half2 ggml_half2;
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_AGGR_U
#define GGML_COMMON_AGGR_S data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_SYCL)
@ -49,7 +65,8 @@ typedef half2 ggml_half2;
typedef sycl::half ggml_half;
typedef sycl::half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_AGGR_U
#define GGML_COMMON_AGGR_S data
#define GGML_COMMON_DECL
#endif
@ -154,9 +171,9 @@ typedef struct {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 dm;
};
} GGML_COMMON_AGGR_U;
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
@ -175,9 +192,9 @@ typedef struct {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 dm;
};
} GGML_COMMON_AGGR_U;
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
@ -196,37 +213,13 @@ typedef struct {
struct {
ggml_half d; // delta
ggml_half s; // d * sum(qs[i])
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 ds;
};
} GGML_COMMON_AGGR_U;
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
typedef struct {
ggml_half d[4]; // deltas for 4 q4_0 blocks
uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
} block_q4_0x4;
static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
typedef struct {
ggml_half d[8]; // deltas for 8 q4_0 blocks
uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
} block_q4_0x8;
static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
typedef struct {
ggml_half d[4]; // deltas for 4 q8_0 blocks
int8_t qs[QK8_0 * 4]; // quants for 4 q8_0 blocks
} block_q8_0x4;
static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
typedef struct {
ggml_half d[8]; // deltas for 8 q8_0 blocks
int8_t qs[QK8_0 * 8]; // quants for 8 q8_0 blocks
} block_q8_0x8;
static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
//
// Ternary quantization
//
@ -261,9 +254,9 @@ typedef struct {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 dm;
};
} GGML_COMMON_AGGR_U;
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
@ -288,9 +281,9 @@ typedef struct {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 dm;
};
} GGML_COMMON_AGGR_U;
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
@ -305,9 +298,9 @@ typedef struct {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
} GGML_COMMON_AGGR_S;
ggml_half2 dm;
};
} GGML_COMMON_AGGR_U;
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
@ -418,12 +411,6 @@ typedef struct {
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
typedef struct {
ggml_half d[4]; // deltas for 4 iq4_nl blocks
uint8_t qs[QK4_NL * 2];// nibbles / quants for 4 iq4_nl blocks
} block_iq4_nlx4;
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
#endif // GGML_COMMON_DECL
#endif // GGML_COMMON_DECL
@ -437,6 +424,13 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro
#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_CPP)
#include <cstdint>
#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_METAL)
#include <metal_stdlib>

6
ggml/src/ggml-cpu/CMakeLists.txt
View File

@ -10,10 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
list (APPEND GGML_CPU_SOURCES
ggml-cpu/ggml-cpu.c
ggml-cpu/ggml-cpu.cpp
ggml-cpu/ggml-cpu-aarch64.c
ggml-cpu/ggml-cpu-aarch64.cpp
ggml-cpu/ggml-cpu-aarch64.h
ggml-cpu/ggml-cpu-hbm.cpp
ggml-cpu/ggml-cpu-hbm.h
ggml-cpu/ggml-cpu-quants.c
ggml-cpu/ggml-cpu-quants.h
ggml-cpu/ggml-cpu-traits.cpp
ggml-cpu/ggml-cpu-traits.h
ggml-cpu/amx/amx.cpp
ggml-cpu/amx/amx.h
ggml-cpu/amx/mmq.cpp

162
ggml/src/ggml-cpu/amx/amx.cpp
View File

@ -5,6 +5,7 @@
#include "ggml-backend.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-traits.h"
#if defined(__gnu_linux__)
#include <sys/syscall.h>
@ -17,31 +18,65 @@
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
// AMX type_trais
namespace ggml::cpu::amx {
class tensor_traits : public ggml::cpu::tensor_traits {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
size = ggml_backend_amx_desired_wsize(op);
return true;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT) {
ggml_backend_amx_mul_mat(params, op);
return true;
}
return false;
}
};
static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
static tensor_traits traits;
return &traits;
}
} // namespace ggml::cpu::amx
// AMX buffer interface
static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
free(buffer->context);
}
static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
return (void *)(buffer->context);
return (void *) (buffer->context);
}
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
memset((char *)tensor->data + offset, value, size);
static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
uint8_t value, size_t offset, size_t size) {
memset((char *) tensor->data + offset, value, size);
GGML_UNUSED(buffer);
}
static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
const void * data, size_t offset, size_t size) {
if (qtype_has_amx_kernels(tensor->type)) {
GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type));
ggml_backend_amx_convert_weight(tensor, data, offset, size);
} else {
memcpy((char *)tensor->data + offset, data, size);
memcpy((char *) tensor->data + offset, data, size);
}
GGML_UNUSED(buffer);
}
/*
// need to figure what we need to do with buffer->extra.
static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
memcpy(data, (const char *)tensor->data + offset, size);
@ -62,6 +97,7 @@ static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
GGML_UNUSED(buffer);
}
*/
static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
memset(buffer->context, value, buffer->size);
@ -70,13 +106,13 @@ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
/* .free_buffer = */ ggml_backend_amx_buffer_free_buffer,
/* .get_base = */ ggml_backend_amx_buffer_get_base,
/* .init_tensor = */ NULL, // no initialization required
/* .init_tensor = */ ggml_backend_amx_buffer_init_tensor,
/* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor,
/* .set_tensor = */ ggml_backend_amx_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_amx_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor,
/* .get_tensor = */ nullptr,
/* .cpy_tensor = */ nullptr,
/* .clear = */ ggml_backend_amx_buffer_clear,
/* .reset = */ NULL,
/* .reset = */ nullptr,
};
static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
@ -101,18 +137,48 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ
GGML_UNUSED(buft);
}
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
namespace ggml::cpu::amx {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
// handle only 2d gemm for now
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
};
if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) && // src0 must be contiguous
is_contiguous_2d(op->src[1]) && // src1 must be contiguous
op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x
(qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
// src1 must be host buffer
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
// src1 must be float32
if (op->src[1]->type == GGML_TYPE_F32) {
return true;
}
}
return false;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
return nullptr;
}
};
} // namespace ggml::cpu::amx
static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
return ggml_backend_amx_get_alloc_size(tensor);
GGML_UNUSED(buft);
}
static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return false;
GGML_UNUSED(buft);
}
#define ARCH_GET_XCOMP_PERM 0x1022
#define ARCH_REQ_XCOMP_PERM 0x1023
#define XFEATURE_XTILECFG 17
@ -129,68 +195,26 @@ static bool ggml_amx_init() {
return true;
#endif
}
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
/* .iface = */ {
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
/* .is_host = */ ggml_backend_amx_buffer_type_is_host,
},
/* .get_name = */ ggml_backend_amx_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size,
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
/* .context = */ new ggml::cpu::amx::extra_buffer_type(),
};
if (!ggml_amx_init()) {
return NULL;
return nullptr;
}
return &ggml_backend_buffer_type_amx;
}
bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
}
bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op) {
// handle only 2d gemm for now
auto is_contiguous_2d = [](const struct ggml_tensor * t) {
return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
};
switch (op->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
return true;
case GGML_OP_MUL_MAT: {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
const enum ggml_type type = src0->type;
const int64_t ne0 = op->ne[0];
// amx kernels enables for Q4_0, Q4_1, Q8_0, F16
// Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
bool can_use_amx =
is_contiguous_2d(src0) && // src0 must be contiguous
is_contiguous_2d(src1) && // src1 must be contiguous
src1->type == GGML_TYPE_F32 && // src1 must be float32
has_amx_kernels && // with amx kernel impls
ne0 % (TILE_N * 2) == 0; // out_features is 32x
return can_use_amx;
}
default:
return false;
}
}
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)

14
ggml/src/ggml-cpu/amx/amx.h
View File

@ -1,20 +1,8 @@
#include "ggml-backend.h"
#include "ggml-cpu-impl.h"
#ifdef __cplusplus
extern "C" {
#endif
// GGML internal header
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft);
bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op);
void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
#endif
#ifdef __cplusplus
}
#endif

19
ggml/src/ggml-cpu/amx/common.h
View File

@ -7,7 +7,7 @@
#include <memory>
#include <type_traits>
#if defined(_OPENMP)
#if defined(GGML_USE_OPENMP)
#include <omp.h>
#endif
@ -56,11 +56,11 @@ inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
}
template <typename func_t>
inline void parallel_for(int nth, int n, const func_t& f) {
#if defined(_OPENMP)
#pragma omp parallel num_threads(nth)
inline void parallel_for(int n, const func_t& f) {
#if defined(GGML_USE_OPENMP)
#pragma omp parallel
{
//int nth = omp_get_num_threads();
int nth = omp_get_num_threads();
int ith = omp_get_thread_num();
int tbegin, tend;
balance211(n, nth, ith, tbegin, tend);
@ -68,8 +68,6 @@ inline void parallel_for(int nth, int n, const func_t& f) {
}
#else
f(0, n);
GGML_UNUSED(nth);
#endif
}
@ -91,10 +89,3 @@ inline bool qtype_has_amx_kernels(const enum ggml_type type) {
(type == GGML_TYPE_Q6_K) ||
(type == GGML_TYPE_IQ4_XS);
}
// ggml backend context
struct ggml_backend_amx_context {
int n_threads = GGML_DEFAULT_N_THREADS;
std::unique_ptr<char[]> work_data;
size_t work_size = 0;
};

17
ggml/src/ggml-cpu/amx/mmq.cpp
View File

@ -18,10 +18,6 @@
#include <unistd.h>
#endif
#if defined(_OPENMP)
#include <omp.h>
#endif
#if (defined(_WIN32) || defined(_WIN64))
#define RESTRICT __restrict
#else
@ -1382,13 +1378,13 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
#define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size
template<typename TB, int BLOCK_K>
void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K, int n_threads) {
void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K) {
const int NB = N / TILE_N;
const int KB = K / BLOCK_K;
const int TILE_SIZE = get_tile_size<TB>();
// parallel on NB should be enough
parallel_for(n_threads, NB, [&](int begin, int end) {
parallel_for(NB, [&](int begin, int end) {
for (int n = begin; n < end; ++n) {
for (int k = 0; k < KB; ++k) {
int n0 = n * TILE_N;
@ -2334,15 +2330,8 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d
const int K = tensor->ne[0]; // ne0: in_features
const int N = tensor->ne[1]; // ne1: out_features
#if defined(_OPENMP)
// the buffer ctx is not initialized when .set_tensor is called
int n_threads = omp_get_num_threads();
#else
int n_threads = 1;
#endif
GGML_DISPATCH_QTYPES(TYPE, [&] {
convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads);
convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K);
});
}

8
ggml/src/ggml-cpu/amx/mmq.h
View File

@ -1,16 +1,10 @@
#pragma once
#include "common.h"
#ifdef __cplusplus
extern "C" {
#endif
size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
#ifdef __cplusplus
}
#endif

723
ggml/src/ggml-cpu/ggml-cpu-aarch64.c → ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
View File

@ -1,20 +1,57 @@
#define GGML_COMMON_IMPL_C
#define GGML_COMMON_IMPL_CPP
#define GGML_COMMON_DECL_CPP
#include "ggml-common.h"
#include "ggml-backend-impl.h"
#include "ggml-quants.h"
#include "ggml-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu/ggml-cpu-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu-traits.h"
#include <math.h>
#include <string.h>
#include <assert.h>
#include <float.h>
#include <stdlib.h> // for qsort
#include <stdio.h> // for GGML_ASSERT
#include <cmath>
#include <cstring>
#include <cassert>
#include <cfloat>
#include <cstdlib> // for qsort
#include <cstdio> // for GGML_ASSERT
#include "ggml-cpu-aarch64.h"
// TODO: move to include file?
template <int K> constexpr int QK_0() {
if constexpr (K == 4) {
return QK4_0;
}
if constexpr (K == 8) {
return QK8_0;
}
return -1;
}
template <int K, int N> struct block {
ggml_half d[N]; // deltas for N qK_0 blocks
int8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_0 blocks
};
// control size
static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
using block_q4_0x4 = block<4, 4>;
using block_q4_0x8 = block<4, 8>;
using block_q8_0x4 = block<8, 4>;
using block_q8_0x8 = block<8, 8>;
struct block_iq4_nlx4 {
ggml_half d[4]; // deltas for 4 iq4_nl blocks
uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks
};
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Woverlength-strings"
#elif defined(_MSC_VER)
@ -185,12 +222,12 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y)
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
static void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
assert(QK8_0 == 32);
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
#if defined(__ARM_NEON)
float32x4_t srcv[4][8];
@ -279,12 +316,12 @@ static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int6
#endif
}
static void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
static void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
assert(QK8_0 == 32);
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
#if defined(__ARM_NEON)
float32x4_t srcv[4][8];
@ -494,7 +531,7 @@ static void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int6
#endif
}
void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
static void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
assert(nrow == 4);
UNUSED(nrow);
if (blck_size_interleave == 4) {
@ -506,7 +543,7 @@ void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nro
}
}
void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -591,7 +628,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -701,7 +738,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
@ -974,7 +1011,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -1070,7 +1107,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void
}
}
void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -1586,7 +1623,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -2040,7 +2077,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
@ -2560,31 +2597,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
// Shuffle pattern one - right side input
const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
// Shuffle pattern two - right side input
const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
// Scale values - Load the weight scale values of two block_q4_0x8
const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
@ -2618,31 +2655,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
// Shuffle pattern one - left side input
const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
// Shuffle pattern two - left side input
const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
// The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
// Resembles MMLAs into 2x2 matrices in ARM Version
@ -2671,10 +2708,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
// Straighten out to make 4 row vectors
__m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
__m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
__m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
__m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
__m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
__m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
__m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
__m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
@ -2753,31 +2790,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
// Shuffle pattern one - right side input
const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
// Shuffle pattern two - right side input
const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
// Scale values - Load the weight scale values of two block_q4_0x8
@ -2809,31 +2846,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
// Shuffle pattern one - left side input
const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
// Shuffle pattern two - left side input
const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
// The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
// Resembles MMLAs into 2x2 matrices in ARM Version
@ -2862,10 +2899,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
// Straighten out to make 4 row vectors
__m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
__m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
__m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
__m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
__m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
__m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
__m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
__m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
// Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
@ -3460,7 +3497,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
}
}
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
static void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
@ -3571,7 +3608,6 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void
}
}
// FIXME: this code is duplicated from ggml-aarch64.c
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x4 out;
@ -3641,20 +3677,20 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_in
return out;
}
static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) {
static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
constexpr int nrows_interleaved = 4;
block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
const block_q4_0 * src = (const block_q4_0 *)data;
block_q4_0 dst_tmp[4];
int nrow = t->ne[1]; // Number of rows
int nrows_interleaved = 4;
int nrow = ggml_nrows(t);
int nblocks = t->ne[0] / QK4_0;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
return -1;
}
@ -3672,20 +3708,20 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
GGML_UNUSED(data_size);
}
static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block, const void * restrict data, size_t data_size) {
static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 8);
constexpr int nrows_interleaved = 8;
block_q4_0x8 * dst = (block_q4_0x8*)t->data;
const block_q4_0 * src = (const block_q4_0*) data;
block_q4_0 dst_tmp[8];
int nrow = t->ne[1]; // Number of rows
int nrows_interleaved = 8;
int nrow = ggml_nrows(t);
int nblocks = t->ne[0] / QK4_0;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
return -1;
}
@ -3736,20 +3772,20 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s
return out;
}
static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) {
static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
const block_iq4_nl * src = (const block_iq4_nl *)data;
block_iq4_nl dst_tmp[4];
int nrow = t->ne[1]; // Number of rows
int nrow = ggml_nrows(t);
int nrows_interleaved = 4;
int nblocks = t->ne[0] / QK4_0;
GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
return -1;
}
@ -3767,57 +3803,456 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b
GGML_UNUSED(data_size);
}
// Prepare for optimized kernels if applicable
void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * restrict data, size_t data_size) {
if (cur->type == repack_type) {
memcpy(cur->data, data, data_size);
return;
}
namespace ggml::cpu::aarch64 {
// repack
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
int repack(struct ggml_tensor *, const void *, size_t);
if (cur->type == GGML_TYPE_Q4_0) {
switch (repack_type) {
case GGML_TYPE_Q4_0_8_8:
repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size);
break;
case GGML_TYPE_Q4_0_4_8:
repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size);
break;
case GGML_TYPE_Q4_0_4_4:
repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size);
break;
default:
GGML_ABORT("Unsupported type");
}
} else if (cur->type == GGML_TYPE_IQ4_NL) {
switch (repack_type) {
case GGML_TYPE_IQ4_NL_4_4:
repack_iq4_nl_to_iq4_nl_4_bl(cur, 4, data, data_size);
break;
default:
GGML_ABORT("Unsupported type");
}
} else {
GGML_ABORT("Unsupported type");
}
// TODO: generalise.
template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size);
}
enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) {
template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size);
}
template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size);
}
template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
}
template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
}
// gemv
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
void gemv(int, float *, size_t, const void *, const void *, int, int);
template <> void gemv<block_q4_0, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q4_0, 8, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q4_0, 8, 8>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <>
void gemv<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
// gemm
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
void gemm(int, float *, size_t, const void *, const void *, int, int);
template <> void gemm<block_q4_0, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q4_0, 8, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q4_0, 8, 8>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
}
template <>
void gemm<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
class tensor_traits_base : public ggml::cpu::tensor_traits {
public:
virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
};
template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_traits : public tensor_traits_base {
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
// not realy a GGML_TYPE_Q8_0 but same size.
switch (op->op) {
case GGML_OP_MUL_MAT:
size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1]));
return true;
case GGML_OP_MUL_MAT_ID:
size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1]));
size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
return true;
default:
// GGML_ABORT("fatal error");
break;
}
return false;
}
bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
switch (op->op) {
case GGML_OP_MUL_MAT:
forward_mul_mat(params, op);
return true;
case GGML_OP_MUL_MAT_ID:
forward_mul_mat_id(params, op);
return true;
default:
// GGML_ABORT("fatal error");
break;
}
return false;
}
void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
const ggml_tensor * src0 = op->src[0];
const ggml_tensor * src1 = op->src[1];
ggml_tensor * dst = op;
GGML_TENSOR_BINARY_OP_LOCALS
const int ith = params->ith;
const int nth = params->nth;
GGML_ASSERT(ne0 == ne01);
GGML_ASSERT(ne1 == ne11);
GGML_ASSERT(ne2 == ne12);
GGML_ASSERT(ne3 == ne13);
// dst cannot be transposed or permuted
GGML_ASSERT(nb0 == sizeof(float));
GGML_ASSERT(nb0 <= nb1);
GGML_ASSERT(nb1 <= nb2);
GGML_ASSERT(nb2 <= nb3);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
// GGML_ASSERT(ggml_n_dims(op->src[1]) == 2);
char * wdata = static_cast<char *>(params->wdata);
const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10);
assert(params->wsize >= nbw1 * ne11);
const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float;
int64_t i11_processed = 0;
for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10,
INTER_SIZE);
}
i11_processed = ne11 - ne11 % 4;
for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
}
ggml_barrier(params->threadpool);
const void * src1_wdata = params->wdata;
const size_t src1_col_stride = ggml_row_size(GGML_TYPE_Q8_0, ne10);
int64_t src0_start = (ith * ne01) / nth;
int64_t src0_end = ((ith + 1) * ne01) / nth;
src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end;
if (src0_start >= src0_end) {
return;
}
// If there are more than three rows in src1, use gemm; otherwise, use gemv.
if (ne11 > 3) {
gemm<BLOC_TYPE, INTER_SIZE, NB_COLS>(ne00, (float *) ((char *) dst->data) + src0_start, ne01,
(const char *) src0->data + src0_start * nb01,
(const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
}
for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
gemv<BLOC_TYPE, INTER_SIZE, NB_COLS>(ne00, (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
(const char *) src0->data + src0_start * nb01,
(const char *) src1_wdata + (src1_col_stride * iter), 1,
src0_end - src0_start);
}
}
void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) {
const ggml_tensor * src0 = op->src[0];
const ggml_tensor * src1 = op->src[1];
const ggml_tensor * ids = op->src[2];
ggml_tensor * dst = op;
GGML_TENSOR_BINARY_OP_LOCALS
const int ith = params->ith;
const int nth = params->nth;
const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float;
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(src0->type));
GGML_ASSERT(nb10 == ggml_type_size(src1->type));
// dst cannot be transposed or permuted
GGML_ASSERT(nb0 == sizeof(float));
GGML_ASSERT(nb0 <= nb1);
GGML_ASSERT(nb1 <= nb2);
GGML_ASSERT(nb2 <= nb3);
GGML_ASSERT(ne03 == 1);
GGML_ASSERT(ne13 == 1);
GGML_ASSERT(ne3 == 1);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
// row groups
const int n_ids = ids->ne[0]; // n_expert_used
const int n_as = ne02; // n_expert
const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10);
const size_t nbw2 = nbw1*ne11;
const size_t nbw3 = nbw2*ne12;
struct mmid_row_mapping {
int32_t i1;
int32_t i2;
};
GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
n_as * ne12 * sizeof(mmid_row_mapping)));
auto wdata = (char *) params->wdata;
auto wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t));
int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
// src1: float32 => block_q8_0
for (int64_t i12 = 0; i12 < ne12; ++i12) {
for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11),
(void *) (wdata + i12 * nbw2 + i11 * nbw1),
ne10);
}
}
#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id) * ne12 + (i1)]
if (ith == 0) {
// initialize matrix_row_counts
memset(matrix_row_counts, 0, n_as * sizeof(int64_t));
// group rows by src0 matrix
for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
for (int32_t id = 0; id < n_ids; ++id) {
const int32_t i02 =
*(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
matrix_row_counts[i02] += 1;
}
}
}
ggml_barrier(params->threadpool);
// compute each matrix multiplication in sequence
for (int cur_a = 0; cur_a < n_as; ++cur_a) {
const int64_t cne1 = matrix_row_counts[cur_a];
if (cne1 == 0) {
continue;
}
auto src0_cur = (const char *) src0->data + cur_a*nb02;
//const int64_t nr0 = ne01; // src0 rows
const int64_t nr1 = cne1; // src1 rows
int64_t src0_cur_start = (ith * ne01) / nth;
int64_t src0_cur_end = ((ith + 1) * ne01) / nth;
src0_cur_start =
(src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end;
if (src0_cur_start >= src0_cur_end) return;
for (int ir1 = 0; ir1 < nr1; ir1++) {
struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
const int id = row_mapping.i1; // selected expert index
const int64_t i11 = id % ne11;
const int64_t i12 = row_mapping.i2; // row index in src1
const int64_t i1 = id; // selected expert index
const int64_t i2 = i12; // row
auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
gemv<BLOC_TYPE, INTER_SIZE, NB_COLS>(
ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start,
ne01, src0_cur + src0_cur_start * nb01,
src1_col, 1, src0_cur_end - src0_cur_start);
}
}
#undef MMID_MATRIX_ROW
}
int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
(int) NB_COLS, (int) INTER_SIZE);
return ggml::cpu::aarch64::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size);
}
};
// instance for Q4
static const tensor_traits<block_q4_0, 4, 4> q4_0_4x4_q8_0;
static const tensor_traits<block_q4_0, 8, 4> q4_0_4x8_q8_0;
static const tensor_traits<block_q4_0, 8, 8> q4_0_8x8_q8_0;
// instance for IQ4
static const tensor_traits<block_iq4_nl, 4, 4> iq4_nl_4x4_q8_0;
} // namespace ggml::cpu::aarch64
static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) {
if (cur->type == GGML_TYPE_Q4_0) {
// TODO: enable for AVX2 - currently disabled due to bad gemv performance
if (/* ggml_cpu_has_avx2() || */ (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
return GGML_TYPE_Q4_0_8_8;
if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
if (cur->ne[1] % 8 == 0) {
return &ggml::cpu::aarch64::q4_0_8x8_q8_0;
}
}
if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
return GGML_TYPE_Q4_0_4_8;
if (cur->ne[1] % 4 == 0) {
return &ggml::cpu::aarch64::q4_0_4x8_q8_0;
}
}
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
return GGML_TYPE_Q4_0_4_4;
if (cur->ne[1] % 4 == 0) {
return &ggml::cpu::aarch64::q4_0_4x4_q8_0;
}
}
} else if (cur->type == GGML_TYPE_IQ4_NL) {
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
return GGML_TYPE_IQ4_NL_4_4;
if (cur->ne[1] % 4 == 0) {
return &ggml::cpu::aarch64::iq4_nl_4x4_q8_0;
}
}
}
return cur->type;
return nullptr;
}
static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_aarch64_get_optimal_repack_type(tensor));
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
auto tensor_traits = (ggml::cpu::aarch64::tensor_traits_base *) tensor->extra;
auto OK = tensor_traits->repack(tensor, data, size);
GGML_ASSERT(OK == 0);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_AARCH64";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == nullptr) {
return nullptr;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
return buffer;
}
static size_t ggml_backend_cpu_aarch64_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return TENSOR_ALIGNMENT;
GGML_UNUSED(buft);
}
namespace ggml::cpu::aarch64 {
class extra_buffer_type : ggml::cpu::extra_buffer_type {
bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
if ( op->op == GGML_OP_MUL_MAT &&
op->src[0]->buffer &&
(ggml_n_dims(op->src[0]) == 2) &&
op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type() &&
ggml_aarch64_get_optimal_repack_type(op->src[0])
) {
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
if (op->src[1]->type == GGML_TYPE_F32) {
return true;
}
//if (op->src[1]->type == GGML_TYPE_Q8_0) {
// return true;
//}
// may be possible if Q8_0 packed...
} else if (op->op == GGML_OP_MUL_MAT_ID
&& op->src[0]->buffer
&& (ggml_n_dims(op->src[0]) == 3)
&& op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()
&& ggml_aarch64_get_optimal_repack_type(op->src[0])
) {
if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
return false;
}
if (op->src[1]->type == GGML_TYPE_F32) {
return true;
}
//if (op->src[1]->type == GGML_TYPE_Q8_0) {
// return true;
//}
}
return false;
}
ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()) {
return (ggml::cpu::tensor_traits *) op->src[0]->extra;
}
}
return nullptr;
}
};
} // namespace ggml::cpu::aarch64
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_aarch64_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .is_host = */ nullptr,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ new ggml::cpu::aarch64::extra_buffer_type(),
};
return &ggml_backend_cpu_buffer_type_aarch64;
}

28
ggml/src/ggml-cpu/ggml-cpu-aarch64.h
View File

@ -1,32 +1,8 @@
#pragma once
#include "ggml-cpu-traits.h"
#include "ggml.h"
// GGML internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// GEMV
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
#ifdef __cplusplus
}
#endif
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

55
ggml/src/ggml-cpu/ggml-cpu-hbm.cpp Normal file
View File

@ -0,0 +1,55 @@
#ifdef GGML_USE_CPU_HBM
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "ggml-cpu-hbm.h"
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ nullptr, // defaults to SIZE_MAX
/* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif

8
ggml/src/ggml-cpu/ggml-cpu-hbm.h Normal file
View File

@ -0,0 +1,8 @@
#pragma once
#include "ggml-backend.h"
#include "ggml.h"
// GGML CPU internal header
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);

36
ggml/src/ggml-cpu/ggml-cpu-traits.cpp Normal file
View File

@ -0,0 +1,36 @@
#include "ggml-cpu-traits.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
namespace ggml::cpu {
tensor_traits::~tensor_traits() {}
extra_buffer_type::~extra_buffer_type() {}
} // namespace ggml::cpu
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->compute_forward(params, op)) {
return true;
}
}
}
return false;
}
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra->context) {
auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context;
auto tensor_traits = buf_extra->get_tensor_traits(op);
if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
return true;
}
}
}
return false;
}

38
ggml/src/ggml-cpu/ggml-cpu-traits.h Normal file
View File

@ -0,0 +1,38 @@
#pragma once
#include "ggml-backend-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml.h"
#ifdef __cplusplus
# include <vector>
extern "C" {
#endif
// return true if op part of extra "accelerator"
bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
#ifdef __cplusplus
}
namespace ggml::cpu {
// register in tensor->extra
class tensor_traits {
public:
virtual ~tensor_traits();
virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size) = 0;
virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
};
class extra_buffer_type {
public:
virtual ~extra_buffer_type();
virtual bool supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op) = 0;
};
} // namespace ggml::cpu
// implemented in ggml-cpu.cpp.
std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
#endif

436
ggml/src/ggml-cpu/ggml-cpu.c
View File

@ -3,7 +3,7 @@
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-cpu-traits.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
@ -224,10 +224,6 @@ typedef void * thread_ret_t;
typedef pthread_t ggml_thread_t;
#ifdef GGML_USE_CPU_HBM
#include <hbwmalloc.h>
#endif
#if defined(__APPLE__)
#include <unistd.h>
#include <mach/mach.h>
@ -301,7 +297,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
},
[GGML_TYPE_Q8_0] = {
.from_float = quantize_row_q8_0,
.from_float_to_mat = quantize_mat_q8_0,
.vec_dot = ggml_vec_dot_q8_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
#if defined (__ARM_FEATURE_MATMUL_INT8)
@ -409,33 +404,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_BF16,
.nrows = 1,
},
[GGML_TYPE_Q4_0_4_4] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 4,
.gemv = ggml_gemv_q4_0_4x4_q8_0,
.gemm = ggml_gemm_q4_0_4x4_q8_0,
},
[GGML_TYPE_Q4_0_4_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 4,
.gemv = ggml_gemv_q4_0_4x8_q8_0,
.gemm = ggml_gemm_q4_0_4x8_q8_0,
},
[GGML_TYPE_Q4_0_8_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 8,
.gemv = ggml_gemv_q4_0_8x8_q8_0,
.gemm = ggml_gemm_q4_0_8x8_q8_0,
},
[GGML_TYPE_TQ1_0] = {
.from_float = quantize_row_tq1_0,
.vec_dot = ggml_vec_dot_tq1_0_q8_K,
@ -448,15 +416,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ4_NL_4_4] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 4,
.gemv = ggml_gemv_iq4_nl_4x4_q8_0,
.gemm = ggml_gemm_iq4_nl_4x4_q8_0,
},
};
const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type) {
@ -4509,9 +4468,6 @@ static void ggml_compute_forward_add(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
{
ggml_compute_forward_add_q_f32(params, dst);
} break;
@ -4889,9 +4845,6 @@ static void ggml_compute_forward_add1(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
{
ggml_compute_forward_add1_q_f32(params, dst);
} break;
@ -5019,9 +4972,6 @@ static void ggml_compute_forward_acc(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
default:
{
GGML_ABORT("fatal error");
@ -7437,27 +7387,9 @@ static void ggml_compute_forward_mul_mat(
const int ith = params->ith;
const int nth = params->nth;
enum ggml_type type = src0->type;
if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
type = (enum ggml_type)(intptr_t)src0->extra;
}
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
if (src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) {
ggml_backend_amx_mul_mat(params, dst);
return;
}
#endif
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
enum ggml_type const vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
ggml_from_float_to_mat_t const from_float_to_mat = type_traits_cpu[vec_dot_type].from_float_to_mat;
int64_t const vec_dot_num_rows = type_traits_cpu[type].nrows;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
int64_t const blck_size_interleave = ggml_get_type_traits(type)->blck_size_interleave;
ggml_gemv_t const gemv = type_traits_cpu[type].gemv;
ggml_gemm_t const gemm = type_traits_cpu[type].gemm;
int64_t const vec_dot_num_rows = type_traits_cpu[src0->type].nrows;
GGML_ASSERT(ne0 == ne01);
GGML_ASSERT(ne1 == ne11);
@ -7465,7 +7397,7 @@ static void ggml_compute_forward_mul_mat(
GGML_ASSERT(ne3 == ne13);
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
GGML_ASSERT(nb00 == ggml_type_size(src0->type));
GGML_ASSERT(nb10 == ggml_type_size(src1->type));
// dst cannot be transposed or permuted
@ -7477,6 +7409,7 @@ static void ggml_compute_forward_mul_mat(
// nb01 >= nb00 - src0 is not transposed
// compute by src0 rows
// TODO: extract to "extra_op"
#if GGML_USE_LLAMAFILE
// broadcast factors
const int64_t r2 = ne12 / ne02;
@ -7487,15 +7420,15 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(type),
nb01/ggml_type_size(src0->type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
type,
src0->type,
src1->type,
dst->type))
goto UseGgmlGemm1;
@ -7516,19 +7449,10 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; ++i13) {
for (int64_t i12 = 0; i12 < ne12; ++i12) {
int64_t i11_processed = 0;
if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) {
for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
4, ne10, blck_size_interleave);
}
i11_processed = ne11 - ne11 % 4;
}
for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
ne10);
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
ne10);
}
}
}
@ -7548,15 +7472,15 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(type),
nb01/ggml_type_size(src0->type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
type,
src0->type,
vec_dot_type,
dst->type))
goto UseGgmlGemm2;
@ -7598,28 +7522,6 @@ UseGgmlGemm2:;
const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
if ((ggml_n_dims(src0) == 2) && gemv) {
const void * src1_wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11;
int64_t src0_start = (ith * ne01) / nth;
int64_t src0_end = ((ith + 1) * ne01) / nth;
src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end;
if (src0_start >= src0_end) return;
// If there are more than three rows in src1, use gemm; otherwise, use gemv.
if (gemm && (ne11 > 3)) {
gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01,
(const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
}
for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) {
gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01,
(const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1,
src0_end - src0_start);
}
return;
}
// The first chunk comes from our thread_id, the rest will get auto-assigned.
int current_chunk = ith;
@ -7642,7 +7544,7 @@ UseGgmlGemm2:;
num_rows_per_vec_dot = 1;
}
ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
if (nth >= nchunk0 * nchunk1) {
break;
@ -7674,8 +7576,6 @@ static void ggml_compute_forward_mul_mat_id(
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
ggml_gemv_t const gemv = type_traits_cpu[type].gemv;
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -7761,34 +7661,6 @@ static void ggml_compute_forward_mul_mat_id(
const int64_t nr0 = ne01; // src0 rows
const int64_t nr1 = cne1; // src1 rows
if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
int64_t src0_cur_start = (ith * ne01) / nth;
int64_t src0_cur_end = ((ith + 1) * ne01) / nth;
src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
src0_cur_end = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
if (src0_cur_start >= src0_cur_end) return;
for (int ir1 = 0; ir1 < nr1; ir1++) {
struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
const int id = row_mapping.i1; // selected expert index
const int64_t i11 = id % ne11;
const int64_t i12 = row_mapping.i2; // row index in src1
const int64_t i1 = id; // selected expert index
const int64_t i2 = i12; // row
const char * src1_col = (const char *) wdata +
(src1_cont || src1->type != vec_dot_type
? (i11 + i12 * ne11) * row_size
: (i11 * nb11 + i12 * nb12));
gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
(const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
}
continue;
}
// distribute the thread work across the inner or outer loop based on which one is larger
const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
@ -8096,9 +7968,6 @@ static void ggml_compute_forward_out_prod(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
{
ggml_compute_forward_out_prod_q_f32(params, dst);
} break;
@ -8361,9 +8230,6 @@ static void ggml_compute_forward_set(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
default:
{
GGML_ABORT("fatal error");
@ -8625,9 +8491,6 @@ static void ggml_compute_forward_get_rows(
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
{
ggml_compute_forward_get_rows_q(params, dst);
} break;
@ -9217,10 +9080,6 @@ static void ggml_compute_forward_clamp(
case GGML_TYPE_IQ3_S:
case GGML_TYPE_IQ2_S:
case GGML_TYPE_Q8_K:
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
case GGML_TYPE_Q4_0_8_8:
case GGML_TYPE_IQ4_NL_4_4:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
case GGML_TYPE_I32:
@ -12426,6 +12285,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
return;
}
// extra_buffer op?
if (ggml_cpu_extra_compute_forward(params, tensor)) return;
switch (tensor->op) {
case GGML_OP_DUP:
{
@ -13373,146 +13235,142 @@ struct ggml_cplan ggml_graph_plan(
size_t cur = 0;
switch (node->op) {
case GGML_OP_CPY:
case GGML_OP_DUP:
{
if (ggml_is_quantized(node->type) ||
// F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
(node->src[0]->type == GGML_TYPE_F16 && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
(node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) {
switch (node->op) {
case GGML_OP_CPY:
case GGML_OP_DUP:
{
if (ggml_is_quantized(node->type) ||
// F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
(node->src[0]->type == GGML_TYPE_F16 && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
(node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
}
} break;
case GGML_OP_ADD:
case GGML_OP_ADD1:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_ACC:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
}
} break;
case GGML_OP_COUNT_EQUAL:
{
cur = ggml_type_size(node->type)*n_tasks;
} break;
case GGML_OP_MUL_MAT:
{
const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
if (node->src[1]->type != vec_dot_type) {
cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
}
} break;
case GGML_OP_MUL_MAT_ID:
{
cur = 0;
const struct ggml_tensor * src0 = node->src[0];
const struct ggml_tensor * src1 = node->src[1];
const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
if (src1->type != vec_dot_type) {
cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
}
const int n_as = src0->ne[2];
cur += GGML_PAD(cur, sizeof(int64_t)); // align
cur += n_as * sizeof(int64_t); // matrix_row_counts
cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
} break;
case GGML_OP_OUT_PROD:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SOFT_MAX:
case GGML_OP_ROPE:
{
cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
}
} break;
case GGML_OP_ADD:
case GGML_OP_ADD1:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_ACC:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
}
} break;
case GGML_OP_COUNT_EQUAL:
{
cur = ggml_type_size(node->type)*n_tasks;
} break;
case GGML_OP_MUL_MAT:
{
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
if (node->src[0]->buffer && ggml_backend_amx_buft_is_amx(node->src[0]->buffer->buft)) {
cur = ggml_backend_amx_desired_wsize(node);
}
#endif
const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{
GGML_ASSERT(node->src[0]->ne[3] == 1);
GGML_ASSERT(node->src[1]->ne[2] == 1);
GGML_ASSERT(node->src[1]->ne[3] == 1);
if (node->src[1]->type != vec_dot_type) {
size_t cur2 = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
cur = MAX(cur, cur2);
}
} break;
case GGML_OP_MUL_MAT_ID:
{
cur = 0;
const struct ggml_tensor * src0 = node->src[0];
const struct ggml_tensor * src1 = node->src[1];
const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
if (src1->type != vec_dot_type) {
cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
}
const int n_as = src0->ne[2];
cur += GGML_PAD(cur, sizeof(int64_t)); // align
cur += n_as * sizeof(int64_t); // matrix_row_counts
cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
} break;
case GGML_OP_OUT_PROD:
{
if (ggml_is_quantized(node->src[0]->type)) {
cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
}
} break;
case GGML_OP_SOFT_MAX:
case GGML_OP_ROPE:
{
cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
} break;
case GGML_OP_CONV_TRANSPOSE_1D:
{
GGML_ASSERT(node->src[0]->ne[3] == 1);
GGML_ASSERT(node->src[1]->ne[2] == 1);
GGML_ASSERT(node->src[1]->ne[3] == 1);
const int64_t ne00 = node->src[0]->ne[0]; // K
const int64_t ne01 = node->src[0]->ne[1]; // Cout
const int64_t ne02 = node->src[0]->ne[2]; // Cin
const int64_t ne10 = node->src[1]->ne[0]; // L
const int64_t ne11 = node->src[1]->ne[1]; // Cin
const int64_t ne00 = node->src[0]->ne[0]; // K
const int64_t ne01 = node->src[0]->ne[1]; // Cout
const int64_t ne02 = node->src[0]->ne[2]; // Cin
if ((node->src[0]->type == GGML_TYPE_F16 ||
node->src[0]->type == GGML_TYPE_BF16) &&
node->src[1]->type == GGML_TYPE_F32) {
cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
cur += sizeof(ggml_fp16_t)*ne10*ne11;
} else if (node->src[0]->type == GGML_TYPE_F32 &&
node->src[1]->type == GGML_TYPE_F32) {
cur += sizeof(float)*ne00*ne01*ne02;
cur += sizeof(float)*ne10*ne11;
} else {
GGML_ABORT("fatal error");
}
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
const int64_t ne00 = node->src[0]->ne[0]; // W
const int64_t ne01 = node->src[0]->ne[1]; // H
const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
const int64_t ne03 = node->src[0]->ne[3]; // Channels In
const int64_t ne10 = node->src[1]->ne[0]; // L
const int64_t ne11 = node->src[1]->ne[1]; // Cin
const int64_t ne10 = node->src[1]->ne[0]; // W
const int64_t ne11 = node->src[1]->ne[1]; // H
const int64_t ne12 = node->src[1]->ne[2]; // Channels In
if ((node->src[0]->type == GGML_TYPE_F16 ||
node->src[0]->type == GGML_TYPE_BF16) &&
node->src[1]->type == GGML_TYPE_F32) {
cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
cur += sizeof(ggml_fp16_t)*ne10*ne11;
} else if (node->src[0]->type == GGML_TYPE_F32 &&
node->src[1]->type == GGML_TYPE_F32) {
cur += sizeof(float)*ne00*ne01*ne02;
cur += sizeof(float)*ne10*ne11;
} else {
cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
const int64_t ne00 = node->src[0]->ne[0]; // D
cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
const int64_t D = node->src[0]->ne[0];
const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
if (node->src[1]->type == GGML_TYPE_F32) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
} else if (node->src[1]->type == GGML_TYPE_F16) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
} else if (node->src[1]->type == GGML_TYPE_BF16) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
}
} break;
case GGML_OP_CROSS_ENTROPY_LOSS:
{
cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
} break;
case GGML_OP_COUNT:
{
GGML_ABORT("fatal error");
}
} break;
case GGML_OP_CONV_TRANSPOSE_2D:
{
const int64_t ne00 = node->src[0]->ne[0]; // W
const int64_t ne01 = node->src[0]->ne[1]; // H
const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
const int64_t ne03 = node->src[0]->ne[3]; // Channels In
const int64_t ne10 = node->src[1]->ne[0]; // W
const int64_t ne11 = node->src[1]->ne[1]; // H
const int64_t ne12 = node->src[1]->ne[2]; // Channels In
cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
const int64_t ne00 = node->src[0]->ne[0]; // D
cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
} break;
case GGML_OP_FLASH_ATTN_BACK:
{
const int64_t D = node->src[0]->ne[0];
const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
if (node->src[1]->type == GGML_TYPE_F32) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
} else if (node->src[1]->type == GGML_TYPE_F16) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
} else if (node->src[1]->type == GGML_TYPE_BF16) {
cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
}
} break;
case GGML_OP_CROSS_ENTROPY_LOSS:
{
cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
} break;
case GGML_OP_COUNT:
{
GGML_ABORT("fatal error");
}
default:
break;
default:
break;
}
}
work_size = MAX(work_size, cur);

172
ggml/src/ggml-cpu/ggml-cpu.cpp
View File

@ -2,12 +2,18 @@
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-cpu-traits.h"
#include "ggml-impl.h"
#include "amx/amx.h"
#include <cctype>
#include <string>
#include <vector>
#ifdef GGML_USE_CPU_HBM
#include "ggml-cpu-hbm.h"
#endif
#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
@ -23,115 +29,7 @@
// ggml-backend interface
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
// buffer type AARCH64
static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra;
ggml_aarch64_repack_tensor(tensor, repack_type, data, size);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_AARCH64";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
auto * buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == NULL) {
return NULL;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ NULL,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_aarch64;
}
bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) {
return buft == ggml_backend_cpu_aarch64_buffer_type();
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
@ -152,11 +50,22 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backen
return bufts;
}();
return bufts.data();
return bufts;
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
return ggml_backend_cpu_get_extra_buffers_type().data();
GGML_UNUSED(device);
}
static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra && extra == buft) return true;
}
return false;
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
@ -465,25 +374,19 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
return true;
}
if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
if (op->op != GGML_OP_MUL_MAT || src0->type == ggml_aarch64_get_optimal_repack_type(src0)) {
return false;
// extra_buffer_op?
for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
if (extra) {
auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
if (buf_extra && buf_extra->supports_op(dev, op)) {
return true;
}
}
}
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
if (src0 && src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) {
return ggml_backend_amx_device_supports_op(op);
}
for (int i = 1; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && ggml_backend_amx_buft_is_amx(op->src[i]->buffer->buft)) {
return false;
}
}
#endif
for (int i = 1; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) {
// the other case need host buffer.
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) {
return false;
}
}
@ -506,19 +409,10 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
default:
return true;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
bool supported = ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft);
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
supported = supported || ggml_backend_amx_buft_is_amx(buft);
#endif
return supported;
return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_is_extra_buffer_type(buft);
GGML_UNUSED(dev);
}
@ -666,10 +560,12 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_cpu_set_n_threads;
ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
return (void *)ggml_backend_cpu_get_extra_bufts;
ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type;
return (void *)fct;
}
if (strcmp(name, "ggml_backend_get_features") == 0) {
return (void *)ggml_backend_cpu_get_features;

2
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -3210,7 +3210,7 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
/* .get_name = */ ggml_backend_cuda_reg_get_name,
/* .get_device_count = */ ggml_backend_cuda_reg_get_device_count,
/* .get_device_get = */ ggml_backend_cuda_reg_get_device,
/* .get_device = */ ggml_backend_cuda_reg_get_device,
/* .get_proc_address = */ ggml_backend_cuda_reg_get_proc_address,
};

9
ggml/src/ggml-quants.c
View File

@ -5220,15 +5220,6 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
{
VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
} break;
case GGML_TYPE_Q4_0_4_4:
case GGML_TYPE_Q4_0_4_8:
{
VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
} break;
case GGML_TYPE_Q4_0_8_8:
{
VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
} break;
case GGML_TYPE_I8:
case GGML_TYPE_I16:

2
ggml/src/ggml-sycl/ggml-sycl.cpp
View File

@ -4630,7 +4630,7 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
static const ggml_backend_reg_i ggml_backend_sycl_reg_interface = {
/* .get_name = */ ggml_backend_sycl_reg_get_name,
/* .get_device_count = */ ggml_backend_sycl_reg_get_device_count,
/* .get_device_get = */ ggml_backend_sycl_reg_get_device,
/* .get_device = */ ggml_backend_sycl_reg_get_device,
/* .get_proc_address = */ ggml_backend_sycl_reg_get_proc_address,
};

86
ggml/src/ggml.c
View File

@ -8,7 +8,10 @@
// FIXME: required here for quantization functions
#include "ggml-quants.h"
#include "ggml-aarch64.h"
#ifdef GGML_USE_CPU_HBM
#include <hbwmalloc.h>
#endif
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <malloc.h> // using malloc.h with MSC/MINGW
@ -788,32 +791,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
.to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row,
.from_float_ref = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
},
[GGML_TYPE_Q4_0_4_4] = {
.type_name = "q4_0_4x4",
.blck_size = QK4_0,
.blck_size_interleave = 4,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[31] = { // GGML_TYPE_Q4_0_4_4
.type_name = "TYPE_Q4_0_4_4 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_Q4_0_4_8] = {
.type_name = "q4_0_4x8",
.blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[32] = { // GGML_TYPE_Q4_0_4_8
.type_name = "TYPE_Q4_0_4_8 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_Q4_0_8_8] = {
.type_name = "q4_0_8x8",
.blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[33] = { // GGML_TYPE_Q4_0_8_8
.type_name = "TYPE_Q4_0_8_8 REMOVED, use Q4_0 with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[GGML_TYPE_TQ1_0] = {
.type_name = "tq1_0",
@ -831,14 +825,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
.to_float = (ggml_to_float_t) dequantize_row_tq2_0,
.from_float_ref = (ggml_from_float_t) quantize_row_tq2_0_ref,
},
[GGML_TYPE_IQ4_NL_4_4] = {
.type_name = "iq4_nl_4x4",
.blck_size = QK4_NL,
.blck_size_interleave = 4,
.type_size = sizeof(block_iq4_nl),
.is_quantized = true,
.to_float = NULL,
.from_float_ref = NULL,
[36] = { // GGML_TYPE_IQ4_NL_4_4
.type_name = "TYPE_IQ4_NL_4_4 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[37] = { // GGML_TYPE_IQ4_NL_4_8
.type_name = "TYPE_IQ4_NL_4_8 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
[38] = { // GGML_TYPE_IQ4_NL_8_8
.type_name = "TYPE_IQ4_NL_8_8 REMOVED, use IQ4_NL with runtime repacking",
.blck_size = 0,
.type_size = 0,
.is_quantized = false,
},
};
@ -1270,9 +1273,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break;
case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break;
case GGML_FTYPE_MOSTLY_IQ2_S: wtype = GGML_TYPE_IQ2_S; break;
case GGML_FTYPE_MOSTLY_Q4_0_4_4: wtype = GGML_TYPE_Q4_0_4_4; break;
case GGML_FTYPE_MOSTLY_Q4_0_4_8: wtype = GGML_TYPE_Q4_0_4_8; break;
case GGML_FTYPE_MOSTLY_Q4_0_8_8: wtype = GGML_TYPE_Q4_0_8_8; break;
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
}
@ -6304,9 +6304,6 @@ size_t ggml_quantize_chunk(
case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_F16:
{
size_t elemsize = sizeof(ggml_fp16_t);
@ -6838,7 +6835,16 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
(int64_t) info->ne[2] *
(int64_t) info->ne[3];
if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
if (ggml_blck_size(info->type) == 0 ) {
// this tensor type support have been removed:
fprintf(stderr, "%s: tensor '%s' of type %d: %s\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type));
fclose(file);
gguf_free(ctx);
return NULL;
}
if (ne % ggml_blck_size(info->type) != 0) {
fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
fclose(file);

12
gguf-py/gguf/constants.py
View File

@ -1432,9 +1432,6 @@ class GGMLQuantizationType(IntEnum):
F64 = 28
IQ1_M = 29
BF16 = 30
Q4_0_4_4 = 31
Q4_0_4_8 = 32
Q4_0_8_8 = 33
TQ1_0 = 34
TQ2_0 = 35
@ -1478,9 +1475,9 @@ class LlamaFileType(IntEnum):
MOSTLY_IQ4_XS = 30 # except 1d tensors
MOSTLY_IQ1_M = 31 # except 1d tensors
MOSTLY_BF16 = 32 # except 1d tensors
MOSTLY_Q4_0_4_4 = 33 # except 1d tensors
MOSTLY_Q4_0_4_8 = 34 # except 1d tensors
MOSTLY_Q4_0_8_8 = 35 # except 1d tensors
# MOSTLY_Q4_0_4_4 = 33 # removed from gguf files, use Q4_0 and runtime repack
# MOSTLY_Q4_0_4_8 = 34 # removed from gguf files, use Q4_0 and runtime repack
# MOSTLY_Q4_0_8_8 = 35 # removed from gguf files, use Q4_0 and runtime repack
MOSTLY_TQ1_0 = 36 # except 1d tensors
MOSTLY_TQ2_0 = 37 # except 1d tensors
@ -1556,9 +1553,6 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
GGMLQuantizationType.F64: (1, 8),
GGMLQuantizationType.IQ1_M: (256, QK_K // 8 + QK_K // 16 + QK_K // 32),
GGMLQuantizationType.BF16: (1, 2),
GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
GGMLQuantizationType.TQ1_0: (256, 2 + 4 * 13),
GGMLQuantizationType.TQ2_0: (256, 2 + 64),
}

6
include/llama.h
View File

@ -172,9 +172,9 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors
LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors
//LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // removed from gguf files, use Q4_0 and runtime repack
//LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // removed from gguf files, use Q4_0 and runtime repack
//LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // removed from gguf files, use Q4_0 and runtime repack
LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors
LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors

24
src/llama.cpp
View File

@ -4578,9 +4578,6 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -5344,9 +5341,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
default: return "unknown, may not work";
}
@ -18367,10 +18361,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
new_type == GGML_TYPE_Q4_0_8_8) {
new_type = GGML_TYPE_Q4_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
new_type = GGML_TYPE_Q4_K;
}
@ -18693,9 +18683,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -19034,14 +19021,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data = (float *) f32_conv_buf.data();
}
int chunk_size_multiplier = 1;
if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
@ -19054,8 +19033,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const int64_t nrows = tensor->ne[1];
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
chunk_size_multiplier;
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;