vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : add AArch64 optimized GEMV and GEMM Q4 kernels (#5780)

Browse source 
* Arm AArch64: optimized GEMV and GEMM kernels for q4_0_q8_0, and q8_0_q8_0 quantization

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add optimized GEMV and GEMM asm kernels for q4_0_q8_0 quantization and refactor code to address llama.cpp pr#5780 suggestions

* Arm AArch64: add copyright claim only to ggml-aarch64.cpp and ggml-aarch64.h files

* Arm AArch64: minor code refactoring for rebase

* Arm AArch64: minor code refactoring for resolving a build issue with cmake

* Arm AArch64: minor code refactoring to split the Q4_0_AARC64 type into three separate types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code change for resolving a build issue with server-windows

* retrigger checks

* Arm AArch64: minor code changes for rebase

* Arm AArch64: minor changes to skip the pr#7433 vec_dot code for arm cpus with SVE VL not equal to 256 bits

* Arm AArch64: remove stale LLAMA_QKK_64 from CMakeLists.txt and delete build.zig

* Arm AArch64: add reference scalar gemm and gemv, and avoid dynamic memory allocations during quantization for Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: add multithreaded quantization support for the new types: Q4_0_4_4, Q4_0_4_8, and Q4_0_8_8

* Arm AArch64: minor code refactoring

* Arm AArch64: simplify logic for calling gemm and gemv functions in ggml_compute_forward_mul_mat

* Arm AArch64: minimize changes in ggml_compute_forward_mul_mat

* Arm AArch64: minor code refactoring, and add reference scalar code to quantize routines for new quant types

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* Arm AArch64: minor code refactoring

* rebase on the latest master commit 3fd62a6 and adapt to the new directory structure

* Arm AArch64: remove a redundant comment

* Arm AArch64: add pragma in ggml-aarch64.c to turn -Woverlength-strings warning off

* Arm AArch64: use __aarch64__ check to guard 64-bit neon kernels

* Arm AArch64: update docs/build.md README to include compile time flags for buiilding the Q4_0_4_4 quant type
This commit is contained in:
Dibakar Gope 2024-07-10 07:14:51 -05:00 • committed by GitHub
parent 83321c6958
commit 0f1a39f343
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
14 changed files with 2534 additions and 53 deletions
Download patch file Download diff file Expand all files Collapse all files

150
ggml/src/ggml.c
View file

@ -4,6 +4,7 @@
#include "ggml-impl.h"
#include "ggml-quants.h"
#include "ggml.h"
#include "ggml-aarch64.h"
#if defined(_MSC_VER) || defined(__MINGW32__)
@ -37,7 +38,7 @@
#include <unistd.h>
#endif
#ifdef __ARM_FEATURE_MATMUL_INT8
#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
#undef GGML_USE_LLAMAFILE
#endif
@ -692,6 +693,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
#else
.nrows = 1,
#endif
.from_float_to_mat = quantize_mat_q8_0,
},
[GGML_TYPE_Q8_1] = {
.type_name = "q8_1",
@ -889,6 +891,54 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16,
.vec_dot_type = GGML_TYPE_BF16,
.nrows = 1,
},
[GGML_TYPE_Q4_0_4_4] = {
.type_name = "q4_0_4x4",
.blck_size = QK4_0,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float = NULL,
.from_float_reference = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 4,
.interleave_blcksize = 4,
.gemv = ggml_gemv_q4_0_4x4_q8_0,
.gemm = ggml_gemm_q4_0_4x4_q8_0,
},
[GGML_TYPE_Q4_0_4_8] = {
.type_name = "q4_0_4x8",
.blck_size = QK4_0,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float = NULL,
.from_float_reference = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 4,
.interleave_blcksize = 8,
.gemv = ggml_gemv_q4_0_4x8_q8_0,
.gemm = ggml_gemm_q4_0_4x8_q8_0,
},
[GGML_TYPE_Q4_0_8_8] = {
.type_name = "q4_0_8x8",
.blck_size = QK4_0,
.type_size = sizeof(block_q4_0),
.is_quantized = true,
.to_float = NULL,
.from_float = NULL,
.from_float_reference = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 8,
.interleave_blcksize = 8,
.gemv = ggml_gemv_q4_0_8x8_q8_0,
.gemm = ggml_gemm_q4_0_8x8_q8_0,
}
};
@ -3188,6 +3238,9 @@ 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;
}
@ -9432,6 +9485,9 @@ 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;
@ -9807,6 +9863,9 @@ 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;
@ -9932,6 +9991,9 @@ 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_ASSERT(false);
@ -12134,6 +12196,12 @@ static void ggml_compute_forward_mul_mat(
enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
int64_t const vec_dot_num_rows = type_traits[type].nrows;
int64_t const matmul_num_cols = type_traits[type].ncols;
int64_t const interleave_blcksize = type_traits[type].interleave_blcksize;
ggml_from_float_to_mat_t const from_float_to_mat
= type_traits[vec_dot_type].from_float_to_mat;
ggml_gemv_t const gemv = type_traits[type].gemv;
ggml_gemm_t const gemm = type_traits[type].gemm;
GGML_ASSERT(ne0 == ne01);
GGML_ASSERT(ne1 == ne11);
@ -12192,7 +12260,16 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; ++i13) {
for (int64_t i12 = 0; i12 < ne12; ++i12) {
for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
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, interleave_blcksize);
}
i11_processed = ne11 - ne11 % 4;
}
for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
ne10);
@ -12273,6 +12350,28 @@ 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;
@ -12318,6 +12417,8 @@ static void ggml_compute_forward_mul_mat_id(
ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot;
enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
int64_t const matmul_num_cols = type_traits[type].ncols;
ggml_gemv_t const gemv = type_traits[type].gemv;
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -12403,6 +12504,34 @@ 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
@ -12704,6 +12833,9 @@ 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;
@ -12889,6 +13021,9 @@ 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_ASSERT(false);
@ -13148,6 +13283,9 @@ 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;
@ -13734,6 +13872,9 @@ 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_I8:
case GGML_TYPE_I16:
case GGML_TYPE_I32:
@ -20457,6 +20598,9 @@ 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);
@ -21759,8 +21903,6 @@ int ggml_cpu_has_neon(void) {
int ggml_cpu_has_sve(void) {
#if defined(__ARM_FEATURE_SVE)
// TODO: Currently, SVE 256 bit is only supported.
GGML_ASSERT(svcntb() == QK8_0);
return 1;
#else
return 0;