/**
* llama.cpp - commit ba1cb19cdd0d92e012e0f6e009e0620f854b6afd - do not edit this file
*
* MIT License
*
* Copyright (c) 2023-2024 The ggml authors
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Wpedantic"
#pragma GCC diagnostic ignored "-Wunused-local-typedefs"
#endif
#include "amx.h"
#include "mmq.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu-quants.h"
#include "ggml-quants.h"
#include <algorithm>
#include <type_traits>
#if defined(__gnu_linux__)
#include <sys/syscall.h>
#include <unistd.h>
#endif
#if (defined(_WIN32) || defined(_WIN64))
#define RESTRICT __restrict
#else
#define RESTRICT __restrict__
#endif
#if (defined(_WIN32) || defined(_WIN64))
#define ALWAYS_INLINE __forceinline
#elif __has_attribute(always_inline) || defined(__GNUC__)
#define ALWAYS_INLINE __attribute__((__always_inline__)) inline
#else
#define ALWAYS_INLINE inline
#endif
#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
namespace {
// Forced unrolling
template <int n>
struct Unroll {
template <typename Func, typename... Args>
ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
Unroll<n - 1>{}(f, args...);
f(std::integral_constant<int, n - 1>{}, args...);
}
};
template <>
struct Unroll<1> {
template <typename Func, typename... Args>
ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
f(std::integral_constant<int, 0>{}, args...);
}
};
// type traits
template <typename T> struct PackedTypes {};
template <> struct PackedTypes<block_q4_0> { using type = int8_t; };
template <> struct PackedTypes<block_q4_1> { using type = uint8_t; };
template <> struct PackedTypes<block_q8_0> { using type = int8_t; };
template <typename T> using packed_B_type = typename PackedTypes<T>::type;
template <typename T>
struct do_compensate : std::integral_constant<bool,
std::is_same<T, block_q8_0>::value> {};
template <typename T>
struct do_unpack : std::integral_constant<bool,
std::is_same<T, block_q4_0>::value ||
std::is_same<T, block_q4_1>::value> {};
template <typename T>
struct is_type_qkk : std::integral_constant<bool,
std::is_same<T, block_q4_K>::value ||
std::is_same<T, block_q5_K>::value ||
std::is_same<T, block_q6_K>::value ||
std::is_same<T, block_iq4_xs>::value> {};
#define GGML_DISPATCH_FLOATING_TYPES(TYPE, ...) \
[&] { \
switch (TYPE) { \
case GGML_TYPE_F16: { \
using type = ggml_fp16_t; \
constexpr int blck_size = 16; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_BF16: { \
using type = ggml_bf16_t; \
constexpr int blck_size = 32; \
return __VA_ARGS__(); \
} \
default: \
fprintf(stderr, "Unsupported floating data type\n"); \
} \
}()
#define GGML_DISPATCH_QTYPES(QT, ...) \
[&] { \
switch (QT) { \
case GGML_TYPE_Q4_0: { \
using type = block_q4_0; \
using vec_dot_type = block_q8_0; \
constexpr int blck_size = QK4_0; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_Q4_1: { \
using type = block_q4_1; \
using vec_dot_type = block_q8_1; \
constexpr int blck_size = QK4_1; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_Q8_0: { \
using type = block_q8_0; \
using vec_dot_type = block_q8_0; \
constexpr int blck_size = QK8_0; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_Q4_K: { \
using type = block_q4_K; \
using vec_dot_type = block_q8_K; \
constexpr int blck_size = QK_K; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_Q5_K: { \
using type = block_q5_K; \
using vec_dot_type = block_q8_K; \
constexpr int blck_size = QK_K; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_Q6_K: { \
using type = block_q6_K; \
using vec_dot_type = block_q8_K; \
constexpr int blck_size = QK_K; \
return __VA_ARGS__(); \
} \
case GGML_TYPE_IQ4_XS: { \
using type = block_iq4_xs; \
using vec_dot_type = block_q8_K; \
constexpr int blck_size = QK_K; \
return __VA_ARGS__(); \
} \
default: \
fprintf(stderr, "Unsupported quantized data type: %d\n", int(TYPE)); \
} \
}()
#define GGML_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...) \
[&] { \
if (BOOL_V) { \
constexpr bool BOOL_NAME = true; \
return __VA_ARGS__(); \
} else { \
constexpr bool BOOL_NAME = false; \
return __VA_ARGS__(); \
} \
}()
// define amx tile config data structure
struct tile_config_t{
uint8_t palette_id = 0;
uint8_t start_row = 0;
uint8_t reserved_0[14] = {0};
uint16_t colsb[16] = {0};
uint8_t rows[16] = {0};
};
// Notes: amx tile config
//
// Typically, TMUL calculates A and B of size 16 x 64 containing INT8 values,
// and accumulate the result to a 16 x 16 matrix C containing INT32 values,
//
// As many GGUF quantized types as `block_size` of 32, so a 16-16-32 config is used
// instead of the normally used 16-16-64 config.
//
// Block A: {16, 32}, dtype = int8_t
// Block B: {16, 32}, dtype = uint8_t/int8_t
// Block C: {16, 16}, dtype = int32_t
//
// Block B needs to be prepacked to vnni format before feeding into TMUL:
// packed_B: from {n, k} to {k/vnni_blk, n, vnni_blck}, viewed in 2d, we get {8, 64}
//
// Therefore, we get tileconfig:
// A B C
// rows 16 8 16
// colsb 32 64 16
//
// For tile distribution, follow a 2-2-4 pattern, e.g. A used TMM2-TMM3, B used TMM0-TMM1,
// C used TMM4-TMM7:
// B TMM0 B TMM1
// A TMM2 C TMM4 C TMM6
// A TMM3 C TMM5 C TMM7
//
// Each `amx` kernel handles 4 blocks at a time: 2MB * 2NB, when m < 2 * BLOCK_M, unpack A
// will be needed.
//
// Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16;
// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`.
//
// ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/
// advanced-matrix-extensions-intrinsics-functions.html
//
#define TC_CONFIG_TILE(i, r, cb) tc.rows[i] = r; tc.colsb[i] = cb
void ggml_tile_config_init(void) {
static thread_local bool is_first_time = true;
if (!is_first_time) {
return;
}
static thread_local tile_config_t tc;
tile_config_t current_tc;
_tile_storeconfig(¤t_tc);
// load only when config changes
if (tc.palette_id == 0 || (memcmp(¤t_tc.colsb, &tc.colsb, sizeof(uint16_t) * 8) != 0 &&
memcmp(¤t_tc.rows, &tc.rows, sizeof(uint8_t) * 8) != 0)) {
tc.palette_id = 1;
tc.start_row = 0;
TC_CONFIG_TILE(TMM0, 8, 64);
TC_CONFIG_TILE(TMM1, 8, 64);
TC_CONFIG_TILE(TMM2, 16, 32);
TC_CONFIG_TILE(TMM3, 16, 32);
TC_CONFIG_TILE(TMM4, 16, 64);
TC_CONFIG_TILE(TMM5, 16, 64);
TC_CONFIG_TILE(TMM6, 16, 64);
TC_CONFIG_TILE(TMM7, 16, 64);
_tile_loadconfig(&tc);
}
is_first_time = false;
}
// we need an extra 16 * 4B (TILE_N * int32_t) for each NB/KB block for compensation.
// See the notes `s8s8 igemm compensation in avx512-vnni` for detail.
template <typename TB>
int get_tile_size() {
int tile_size = TILE_N * sizeof(TB);
if (do_compensate<TB>::value) {
tile_size += TILE_N * sizeof(int32_t);
}
if (std::is_same<TB, block_q4_K>::value ||
std::is_same<TB, block_q5_K>::value) {
tile_size += TILE_N * 4;
}
if (std::is_same<TB, block_iq4_xs>::value) {
tile_size += TILE_N * 2;
}
return tile_size;
}
template <typename TB, int BLOCK_K>
int get_row_size(int K) {
int KB = K / BLOCK_K;
int row_size = KB * sizeof(TB);
if (do_compensate<TB>::value) {
row_size += KB * sizeof(int32_t);
}
if (std::is_same<TB, block_q4_K>::value ||
std::is_same<TB, block_q5_K>::value) {
row_size += KB * 4;
}
if (std::is_same<TB, block_iq4_xs>::value) {
row_size += KB * 2;
}
return row_size;
}
// vectorized dtype conversion
inline float FP16_TO_FP32(ggml_half val) {
__m256i v = _mm256_setr_epi16(
val, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
__m512 o = _mm512_cvtph_ps(v);
return _mm512_cvtss_f32(o);
}
inline __m512 FP16_TO_FP32_VEC(ggml_half val) {
__m256i v = _mm256_set1_epi16(val);
return _mm512_cvtph_ps(v);
}
// horizontal reduce
inline float _mm512_reduce_max_ps(const __m512 x) {
__m512 v = x;
__m512 v1 = _mm512_shuffle_f32x4(v, v, 0x4E);
v = _mm512_max_ps(v, v1);
v1 = _mm512_shuffle_f32x4(v, v, 0xB1);
v = _mm512_max_ps(v, v1);
v1 = _mm512_shuffle_ps(v, v, 0x4E);
v = _mm512_max_ps(v, v1);
v1 = _mm512_shuffle_ps(v, v, 0xB1);
v = _mm512_max_ps(v, v1);
return _mm512_cvtss_f32(v);
}
// transpose utils
#define SHUFFLE_EPI32(a, b, mask) \
_mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), mask))
inline void transpose_8x8_32bit(__m256i * v, __m256i * v1) {
// unpacking and 32-bit elements
v1[0] = _mm256_unpacklo_epi32(v[0], v[1]);
v1[1] = _mm256_unpackhi_epi32(v[0], v[1]);
v1[2] = _mm256_unpacklo_epi32(v[2], v[3]);
v1[3] = _mm256_unpackhi_epi32(v[2], v[3]);
v1[4] = _mm256_unpacklo_epi32(v[4], v[5]);
v1[5] = _mm256_unpackhi_epi32(v[4], v[5]);
v1[6] = _mm256_unpacklo_epi32(v[6], v[7]);
v1[7] = _mm256_unpackhi_epi32(v[6], v[7]);
// shuffling the 32-bit elements
v[0] = SHUFFLE_EPI32(v1[0], v1[2], 0x44);
v[1] = SHUFFLE_EPI32(v1[0], v1[2], 0xee);
v[2] = SHUFFLE_EPI32(v1[4], v1[6], 0x44);
v[3] = SHUFFLE_EPI32(v1[4], v1[6], 0xee);
v[4] = SHUFFLE_EPI32(v1[1], v1[3], 0x44);
v[5] = SHUFFLE_EPI32(v1[1], v1[3], 0xee);
v[6] = SHUFFLE_EPI32(v1[5], v1[7], 0x44);
v[7] = SHUFFLE_EPI32(v1[5], v1[7], 0xee);
// shuffling 128-bit elements
v1[0] = _mm256_permute2f128_si256(v[2], v[0], 0x02);
v1[1] = _mm256_permute2f128_si256(v[3], v[1], 0x02);
v1[2] = _mm256_permute2f128_si256(v[6], v[4], 0x02);
v1[3] = _mm256_permute2f128_si256(v[7], v[5], 0x02);
v1[4] = _mm256_permute2f128_si256(v[2], v[0], 0x13);
v1[5] = _mm256_permute2f128_si256(v[3], v[1], 0x13);
v1[6] = _mm256_permute2f128_si256(v[6], v[4], 0x13);
v1[7] = _mm256_permute2f128_si256(v[7], v[5], 0x13);
}
inline void transpose_16x4_32bit(__m512i * r, __m512i * d) {
static const __m512i index1 = _mm512_set_epi32(
0x0f, 0x0b, 0x07, 0x03,
0x0e, 0x0a, 0x06, 0x02,
0x0d, 0x09, 0x05, 0x01,
0x0c, 0x08, 0x04, 0x00);
d[0] = _mm512_permutexvar_epi32(index1, r[0]);
d[1] = _mm512_permutexvar_epi32(index1, r[1]);
d[2] = _mm512_permutexvar_epi32(index1, r[2]);
d[3] = _mm512_permutexvar_epi32(index1, r[3]);
r[0] = _mm512_shuffle_i32x4(d[0], d[1], 0x44);
r[1] = _mm512_shuffle_i32x4(d[0], d[1], 0xee);
r[2] = _mm512_shuffle_i32x4(d[2], d[3], 0x44);
r[3] = _mm512_shuffle_i32x4(d[2], d[3], 0xee);
d[0] = _mm512_shuffle_i32x4(r[0], r[2], 0x88);
d[1] = _mm512_shuffle_i32x4(r[0], r[2], 0xdd);
d[2] = _mm512_shuffle_i32x4(r[1], r[3], 0x88);
d[3] = _mm512_shuffle_i32x4(r[1], r[3], 0xdd);
}
inline void transpose_16x16_32bit(__m512i * v) {
__m512i v1[16];
v1[0] = _mm512_unpacklo_epi32(v[0], v[1]);
v1[1] = _mm512_unpackhi_epi32(v[0], v[1]);
v1[2] = _mm512_unpacklo_epi32(v[2], v[3]);
v1[3] = _mm512_unpackhi_epi32(v[2], v[3]);
v1[4] = _mm512_unpacklo_epi32(v[4], v[5]);
v1[5] = _mm512_unpackhi_epi32(v[4], v[5]);
v1[6] = _mm512_unpacklo_epi32(v[6], v[7]);
v1[7] = _mm512_unpackhi_epi32(v[6], v[7]);
v1[8] = _mm512_unpacklo_epi32(v[8], v[9]);
v1[9] = _mm512_unpackhi_epi32(v[8], v[9]);
v1[10] = _mm512_unpacklo_epi32(v[10], v[11]);
v1[11] = _mm512_unpackhi_epi32(v[10], v[11]);
v1[12] = _mm512_unpacklo_epi32(v[12], v[13]);
v1[13] = _mm512_unpackhi_epi32(v[12], v[13]);
v1[14] = _mm512_unpacklo_epi32(v[14], v[15]);
v1[15] = _mm512_unpackhi_epi32(v[14], v[15]);
v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]);
v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]);
v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]);
v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]);
v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]);
v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]);
v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]);
v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]);
v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]);
v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]);
v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]);
v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]);
v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]);
v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]);
v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]);
v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]);
v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88);
v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88);
v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88);
v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88);
v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd);
v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd);
v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd);
v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd);
v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88);
v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88);
v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88);
v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88);
v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd);
v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd);
v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd);
v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd);
v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88);
v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88);
v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88);
v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88);
v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88);
v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88);
v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88);
v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88);
v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd);
v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd);
v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd);
v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd);
v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd);
v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd);
v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd);
v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd);
}
void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_t k) {
assert(k % QK_K == 0);
const int KB = k / QK_K;
constexpr int kVecs = QK_K / 16;
block_q8_K * y = reinterpret_cast<block_q8_K *>(vy);
// hold 16 float vecs from x
__m512 v[kVecs];
// hold the quants vecs
__m512i vq[kVecs / 4];
// hold the packed quants vecs
__m512i vq_packed[kVecs / 4];
const __m512 signBit = _mm512_set1_ps(-0.f);
for (int i = 0; i < KB; ++i) {
// Compute max(abs(e)) for the block
__m512 vamax = _mm512_set1_ps(0.f);
for (int j = 0; j < kVecs; ++j) {
v[j] = _mm512_loadu_ps(x); x += 16;
vamax = _mm512_max_ps(vamax, _mm512_andnot_ps(signBit, v[j]));
}
const float amax = _mm512_reduce_max_ps(vamax);
// Quantize these floats
const float iscale = 127.f / amax;
y[i].d = GGML_FP32_TO_FP16(1 / iscale);
const float id = ( amax != 0.0f ) ? iscale : 0.f;
const __m512 vscale = _mm512_set1_ps(id);
// Apply multiplier and round to nearest integer
for (int j = 0; j < kVecs; ++j) {
v[j] = _mm512_mul_ps(v[j], vscale);
v[j] = _mm512_roundscale_ps(v[j], (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
}
// Pack to epi8 vecs
for (int j = 0; j < kVecs / 4; ++j) {
__m128i q8_0 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 0]));
__m128i q8_1 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 1]));
__m128i q8_2 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 2]));
__m128i q8_3 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 3]));
__m256i q8_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_0), (q8_1), 1);
__m256i q8_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_2), (q8_3), 1);
vq[j] = _mm512_inserti32x8(_mm512_castsi256_si512(q8_01), q8_23, 1);
_mm512_storeu_si512((__m512i *)(y[i].qs + j * 64), vq[j]);
}
// Compute the bsums with vnni
transpose_16x4_32bit(vq, vq_packed);
const __m512i one = _mm512_set1_epi8(1);
__m512i sum = _mm512_setzero_si512();
for (int k = 0; k < 4; ++k) {
sum = _mm512_dpbusd_epi32(sum, one, vq_packed[k]);
}
_mm256_storeu_si256((__m256i *)(y[i].bsums), _mm512_cvtepi32_epi16(sum));
}
}
// quantize A from float to `vec_dot_type`
template <typename T>
inline void from_float(const float * x, char * vy, int64_t k);
template <>
inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
quantize_row_q8_0(x, (block_q8_0 *)vy, k);
}
template <>
inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
quantize_row_q8_1(x, (block_q8_1 *)vy, k);
}
template <>
inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
#if 1
// TODO: this is reference impl!
quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
#else
quantize_row_q8_K_vnni(x, vy, k);
#endif
}
// load A from memory to array when nrows can not fill in whole tile
void unpack_A(int8_t * RESTRICT tile, const block_q8_0 * RESTRICT A, int lda, int nr) {
assert(nr != TILE_M);
for (int m = 0; m < nr; ++m) {
const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs));
_mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
}
}
void unpack_A(int8_t * RESTRICT tile, const block_q8_1 * RESTRICT A, int lda, int nr) {
assert(nr != TILE_M);
for (int m = 0; m < nr; ++m) {
const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs));
_mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
}
}
template <typename TB>
void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) {
assert(nr <= TILE_M);
for (int m = 0; m < nr; ++m) {
const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs + k * 32));
_mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
}
}
template <>
void unpack_A<block_q6_K>(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) {
assert(nr <= TILE_M);
// zero padding k from 16 to 32, so that we don't have to re-config amx
const __m128i zero = _mm_setzero_si128();
for (int m = 0; m < nr; ++m) {
const __m128i v = _mm_loadu_si128((const __m128i *)(A[m * lda].qs + k * 16));
const __m256i r = _mm256_insertf128_si256(_mm256_castsi128_si256(v), zero, 1);
_mm256_storeu_si256((__m256i *)(tile + m * TILE_K), r);
}
}
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
const __m256i lowMask = _mm256_set1_epi8(0xF);
return _mm256_and_si256(lowMask, bytes);
}
// used for block_q4_K
inline __m512i bytes_from_nibbles_64(const uint8_t * rsi) {
const __m256i tmp = _mm256_loadu_si256((const __m256i *)rsi);
const __m256i lowMask = _mm256_set1_epi8(0xF);
const __m256i q4l = _mm256_and_si256(tmp, lowMask);
const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(tmp, 4), lowMask);
return _mm512_inserti32x8(_mm512_castsi256_si512(q4l), q4h, 1);
}
// used for block_q5_K
inline __m512i bytes_from_nibbles_64(const uint8_t * qs, const uint8_t * qh, int k) {
const __m256i lowMask = _mm256_set1_epi8(0xF);
__m256i hmask = _mm256_set1_epi8(1);
hmask = _mm256_slli_epi16(hmask, k);
const __m256i q5bits = _mm256_loadu_si256((const __m256i *)qs);
const __m256i hbits = _mm256_loadu_si256((const __m256i *)qh);
const __m256i q5l_0 = _mm256_and_si256(q5bits, lowMask);
const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 0), 4);
const __m256i q5_0 = _mm256_add_epi8(q5l_0, q5h_0);
hmask = _mm256_slli_epi16(hmask, 1);
const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), lowMask);
const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 1), 4);
const __m256i q5_1 = _mm256_add_epi8(q5l_1, q5h_1);
return _mm512_inserti32x8(_mm512_castsi256_si512(q5_0), q5_1, 1);
}
// used for block_q6_K
inline void bytes_from_nibbles_128(__m512i& r0, __m512i& r1, const uint8_t * qs, const uint8_t * qh) {
const __m256i m4 = _mm256_set1_epi8(0xF);
const __m256i m2 = _mm256_set1_epi8(0x3);
const __m256i q6bits1 = _mm256_loadu_si256((const __m256i *)qs);
const __m256i q6bits2 = _mm256_loadu_si256((const __m256i *)(qs + 32));
const __m256i q6bitsH = _mm256_loadu_si256((const __m256i *)qh);
const __m256i q6h_0 = _mm256_slli_epi16(_mm256_and_si256( q6bitsH, m2), 4);
const __m256i q6h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 2), m2), 4);
const __m256i q6h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 4), m2), 4);
const __m256i q6h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 6), m2), 4);
const __m256i q6_0 = _mm256_or_si256(_mm256_and_si256(q6bits1, m4), q6h_0);
const __m256i q6_1 = _mm256_or_si256(_mm256_and_si256(q6bits2, m4), q6h_1);
const __m256i q6_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits1, 4), m4), q6h_2);
const __m256i q6_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits2, 4), m4), q6h_3);
r0 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_0), q6_1, 1);
r1 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_2), q6_3, 1);
}
inline __m512i packNibbles(__m512i r0, __m512i r1) {
return _mm512_or_si512(r0, _mm512_slli_epi16(r1, 4));
}
template <typename TB>
inline void pack_qs(void * RESTRICT packed_B, const TB * RESTRICT B, int KB) {
int8_t tmp[8 * 64];
__m256i v[8], v2[8];
for (int n = 0; n < 8; ++n) {
v[n] = bytes_from_nibbles_32(B[n * KB].qs);
}
transpose_8x8_32bit(v, v2);
for (int n = 0; n < 8; ++n) {
_mm256_storeu_si256((__m256i *)(tmp + n * 64), v2[n]);
}
for (int n = 0; n < 8; ++n) {
v[n] = bytes_from_nibbles_32(B[(n + 8) * KB].qs);
}
transpose_8x8_32bit(v, v2);
for (int n = 0; n < 8; ++n) {
_mm256_storeu_si256((__m256i *)(tmp + n * 64 + 32), v2[n]);
}
// pack again with 128 to fully utilize vector length
for (int n = 0; n < 8; n += 2) {
__m512i r0 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64));
__m512i r1 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64 + 64));
__m512i r1r0 = packNibbles(r0, r1);
_mm512_storeu_si512((__m512i *)((char *)packed_B + n * 32), r1r0);
}
}
template <>
inline void pack_qs<block_q8_0>(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) {
__m256i v[8], v2[8];
for (int n = 0; n < 8; ++n) {
v[n] = _mm256_loadu_si256((const __m256i *)(B[n * KB].qs));
}
transpose_8x8_32bit(v, v2);
for (int n = 0; n < 8; ++n) {
_mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64), v2[n]);
}
for (int n = 0; n < 8; ++n) {
v[n] = _mm256_loadu_si256((const __m256i *)(B[(n + 8) * KB].qs));
}
transpose_8x8_32bit(v, v2);
for (int n = 0; n < 8; ++n) {
_mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64 + 32), v2[n]);
}
}
template <>
inline void pack_qs<block_q4_K>(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) {
__m512i v[16];
// QK_K 256 with 8 groups, handle 2 groups at a time
char * pb = (char *)packed_B;
for (int k = 0; k < QK_K / 64; ++k) {
// pack 2 groups { n, g, k} to {g, k/4, 4n}
// e.g. {16, 2, 32} to {2, 8, 64}
for (int n = 0; n < TILE_N; ++n) {
v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32);
}
transpose_16x16_32bit(v);
// pack again with 128 to fully utilize vector length
for (int n = 0; n < TILE_N; n += 2) {
_mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1]));
pb += 64;
}
}
}
template <>
inline void pack_qs<block_q5_K>(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) {
__m512i v[16];
const __m512i lowMask = _mm512_set1_epi8(0xF);
// QK_K 256 with 8 groups, handle 2 groups at a time
char * pb = (char *)packed_B;
char * ph = (char *)packed_B + (QK_K / 2) * TILE_N;
for (int k = 0; k < QK_K / 64; ++k) {
// pack 2 groups { n, g, k} to {g, k/4, 4n}
// e.g. {16, 2, 32} to {2, 8, 64}
for (int n = 0; n < TILE_N; ++n) {
v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32, B[n * KB].qh, /* group */2 * k);
}
transpose_16x16_32bit(v);
// 1. pack lower 4bits with 2 groups
for (int n = 0; n < TILE_N; n += 2) {
// get lower 4 bits
const __m512i r0 = _mm512_and_si512(v[n], lowMask);
const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask);
_mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64;
}
// 2. pack higher 1bit with 2 groups
const __m512i hmask = _mm512_set1_epi8(0x10);
for (int g = 0; g < 2; ++g) {
__m512i hbits = _mm512_setzero_si512();
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 0], hmask), 4));
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 1], hmask), 3));
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 2], hmask), 2));
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 3], hmask), 1));
hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 8 + 4], hmask) );
hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 5], hmask), 1));
hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 6], hmask), 2));
hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 7], hmask), 3));
_mm512_storeu_si512((__m512i *)ph, hbits); ph += 64;
}
}
}
template <>
inline void pack_qs<block_q6_K>(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) {
__m512i v[32];
const __m512i lowMask = _mm512_set1_epi8(0xF);
// QK_K 256 with 8 groups, handle 4 groups at a time
char * pb = (char *)packed_B;
char * ph = (char *)packed_B + (QK_K / 2) * TILE_N;
for (int k = 0; k < QK_K / 128; ++k) {
for (int n = 0; n < TILE_N; ++n) {
bytes_from_nibbles_128(v[n], v[n + 16], B[n * KB].ql + k * 64, B[n * KB].qh + k * 32);
}
// top half: group 0,1 or 4,5; bottom half: group 2,3 or 6,7
transpose_16x16_32bit(v);
transpose_16x16_32bit(v + 16);
// 1. pack lower 4bits with 4 groups
for (int n = 0; n < 32; n += 2) {
const __m512i r0 = _mm512_and_si512(v[n], lowMask);
const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask);
_mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64;
}
// 2. pack higher 2bit with 4 groups
const __m512i hmask = _mm512_set1_epi8(0x30);
for (int g = 0; g < 8; ++g) {
__m512i hbits = _mm512_setzero_si512();
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 0], hmask), 4));
hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 1], hmask), 2));
hbits = _mm512_add_epi8(hbits, _mm512_and_si512(v[g * 4 + 2], hmask) );
hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 4 + 3], hmask), 2));
_mm512_storeu_si512((__m512i *)ph, hbits); ph += 64;
}
}
}
template <>
inline void pack_qs<block_iq4_xs>(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) {
__m512i v[16];
char * pb = (char *)packed_B;
for (int k = 0; k < QK_K / 64; ++k) {
for (int n = 0; n < TILE_N; ++n) {
__m256i r0 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 0);
__m256i r1 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 16);
v[n] = _mm512_inserti32x8(_mm512_castsi256_si512(r0), r1, 1);
}
transpose_16x16_32bit(v);
// pack again with 128 to fully utilize vector length
for (int n = 0; n < TILE_N; n += 2) {
_mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1]));
pb += 64;
}
}
}
// pack B to vnni formats in 4bits or 8 bits
void pack_B(void * RESTRICT packed_B, const block_q4_0 * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K / 2);
for (int n = 0; n < TILE_N; ++n) {
d0[n] = B[n * KB].d;
}
}
void pack_B(void * RESTRICT packed_B, const block_q4_1 * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K / 2);
ggml_half * m0 = d0 + TILE_N;
for (int n = 0; n < TILE_N; ++n) {
d0[n] = B[n * KB].d;
m0[n] = B[n * KB].m;
}
}
inline void s8s8_compensation(void * RESTRICT packed_B) {
// packed_B layout:
// quants {TILE_N, TILEK} int8_t
// d0 {TILE_N} ggml_half
// comp {TILE_N} int32_t
const int offset = TILE_N * TILE_K + TILE_N * sizeof(ggml_half);
__m512i vcomp = _mm512_setzero_si512();
const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
for (int k = 0; k < 8; ++k) {
__m512i vb = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + k * 64));
vcomp = _mm512_dpbusd_epi32(vcomp, off, vb);
}
_mm512_storeu_si512((__m512i *)((char *)(packed_B) + offset), vcomp);
}
void pack_B(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K);
for (int n = 0; n < TILE_N; ++n) {
d0[n] = B[n * KB].d;
}
s8s8_compensation(packed_B);
}
// convert 8 * {min, scale} from int6 to int8
inline void unpack_mins_and_scales(const uint8_t * scales, uint32_t * utmp) {
const uint32_t kmask1 = 0x3f3f3f3f;
const uint32_t kmask2 = 0x0f0f0f0f;
const uint32_t kmask3 = 0x03030303;
memcpy(utmp, scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
}
// packed_B layout:
// quants {8, TILE_N, 16} uint8
// scales {8, TILE_N} uint8
// mins {8, TILE_N} uint8
// d {TILE_N} ggml_half
// dmin {TILE_N} ggml_half
void pack_B(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N);
uint8_t * mins = scales + 8 * TILE_N;
ggml_half * d = reinterpret_cast<ggml_half *>(mins + 8 * TILE_N);
ggml_half * dmin = d + TILE_N;
union {
uint32_t u32[4];
uint8_t u8[16];
} s;
for (int n = 0; n < TILE_N; ++n) {
unpack_mins_and_scales(B[n * KB].scales, s.u32);
for (int k = 0; k < 8; ++k) {
scales[k * TILE_N + n] = s.u8[k];
mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8];
}
d[n] = B[n * KB].d;
dmin[n] = B[n * KB].dmin;
}
}
// packed_B layout:
// quants {8, TILE_N, 16} uint8
// qh {8, TILE_N, 4} uint8
// scales {8, TILE_N} uint8
// mins {8, TILE_N} uint8
// d {TILE_N} ggml_half
// dmin {TILE_N} ggml_half
void pack_B(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N);
uint8_t * mins = scales + 8 * TILE_N;
ggml_half * d = reinterpret_cast<ggml_half *>(mins + 8 * TILE_N);
ggml_half * dmin = d + TILE_N;
union {
uint32_t u32[4];
uint8_t u8[16];
} s;
for (int n = 0; n < TILE_N; ++n) {
unpack_mins_and_scales(B[n * KB].scales, s.u32);
for (int k = 0; k < 8; ++k) {
scales[k * TILE_N + n] = s.u8[k];
mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8];
}
d[n] = B[n * KB].d;
dmin[n] = B[n * KB].dmin;
}
}
// packed_B layout:
// quants {16, TILE_N, 8} uint8
// qh {16, TILE_N, 4} uint8
// scales {16, TILE_N} uint8
// d {TILE_N} ggml_half
void pack_B(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N);
ggml_half * d = reinterpret_cast<ggml_half *>(scales + 16 * TILE_N);
for (int n = 0; n < TILE_N; ++n) {
const int8_t * ps = B[n * KB].scales;
for (int k = 0; k < 16; ++k) {
scales[k * TILE_N + n] = ps[k];
}
d[n] = B[n * KB].d;
}
}
// packed_B layout:
// quants {8, TILE_N, 16} uint8
// scales {8, TILE_N} int8
// d {TILE_N} ggml_half
void pack_B(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) {
pack_qs(packed_B, B, KB);
int8_t * scales = reinterpret_cast<int8_t *>((char *)packed_B + (QK_K / 2) * TILE_N);
ggml_half * d = reinterpret_cast<ggml_half *>(scales + 8 * TILE_N);
// pack the scales
for (int n = 0; n < TILE_N; ++n) {
uint16_t sh = B[n * KB].scales_h;
for (int k = 0; k < 8; k += 2) {
const int16_t ls1 = ((B[n * KB].scales_l[k / 2] & 0xf) | ((sh << 4) & 0x30)) - 32;
const int16_t ls2 = ((B[n * KB].scales_l[k / 2] >> 4) | ((sh << 2) & 0x30)) - 32;
scales[(k + 0) * TILE_N + n] = ls1;
scales[(k + 1) * TILE_N + n] = ls2;
sh >>= 4;
}
d[n] = B[n * KB].d;
}
}
template<typename TB, typename packed_B_t = packed_B_type<TB>>
void unpack_B(packed_B_t * RESTRICT tile, const void * RESTRICT packed_B) {
GGML_UNUSED(tile);
GGML_UNUSED(packed_B);
}
template <>
void unpack_B<block_q4_0>(int8_t * RESTRICT tile, const void * RESTRICT packed_B) {
const __m512i off = _mm512_set1_epi8(8);
const __m512i lowMask = _mm512_set1_epi8(0xF);
for (int n = 0; n < 8; n += 2) {
__m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32));
const __m512i r0 = _mm512_sub_epi8(_mm512_and_si512(bytes, lowMask), off);
const __m512i r1 = _mm512_sub_epi8(_mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask), off);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
}
}
template <>
void unpack_B<block_q4_1>(uint8_t * RESTRICT tile, const void * RESTRICT packed_B) {
const __m512i lowMask = _mm512_set1_epi8(0xF);
for (int n = 0; n < 8; n += 2) {
__m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32));
const __m512i r0 = _mm512_and_si512(bytes, lowMask);
const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
}
}
// packed_B_t for QKK is int8_t
template <typename TB>
void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
const int packed_B_group_size = QK_K / 2 * TILE_N / 8;
const char * packed_B_group = (const char *)packed_B + k * packed_B_group_size;
const __m512i lowMask = _mm512_set1_epi8(0xF);
for (int n = 0; n < 8; n += 2) {
__m512i bytes = _mm512_loadu_si512(packed_B_group + n * 32);
const __m512i r0 = _mm512_and_si512(bytes, lowMask);
const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
}
}
template <>
void unpack_B<block_q5_K>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
// lower 4bits, stride 256 bytes
const int packed_l4_group_size = QK_K / 2 * TILE_N / 8;
const char * pb = (const char *)packed_B + k * packed_l4_group_size;
// higher 1bit, stride 64 bytes
const int packed_h1_group_size = QK_K / 8 * TILE_N / 8;
const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h1_group_size;
const __m512i hbits = _mm512_loadu_si512(ph);
const __m512i lowMask = _mm512_set1_epi8(0xF);
__m512i hmask0 = _mm512_set1_epi8(0x1);
__m512i hmask1 = _mm512_set1_epi8(0x2);
for (int n = 0; n < 8; n += 2) {
__m512i bytes = _mm512_loadu_si512(pb + n * 32);
__m512i r0 = _mm512_and_si512(bytes, lowMask);
__m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
__m512i h0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), n), 4);
__m512i h1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), n + 1), 4);
hmask0 = _mm512_slli_epi16(hmask0, 2);
hmask1 = _mm512_slli_epi16(hmask1, 2);
r0 = _mm512_add_epi8(r0, h0);
r1 = _mm512_add_epi8(r1, h1);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
}
}
template <>
void unpack_B<block_q6_K>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
// lower 4bits, stride 128 bytes
const int packed_l4_group_size = QK_K / 2 * TILE_N / 16;
const char * pb = (const char *)packed_B + k * packed_l4_group_size;
// higher 2bits, stride 64 bytes
const int packed_h2_group_size = QK_K / 4 * TILE_N / 16;
const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h2_group_size;
const __m512i hbits = _mm512_loadu_si512(ph);
const __m512i off = _mm512_set1_epi8(32);
const __m512i lowMask = _mm512_set1_epi8(0xF);
__m512i hmask0 = _mm512_set1_epi8(0x3); // 0011
__m512i hmask1 = _mm512_set1_epi8(0xC); // 1100
// notes: skip zero padding from row4 to row7 as we have done so in `unpack_A`
__m512i bytes = _mm512_loadu_si512(pb);
__m512i r0 = _mm512_and_si512(bytes, lowMask);
__m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
__m512i h0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask0), 4);
__m512i h1 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask1), 2);
_mm512_storeu_si512((__m512i *)(tile + 0), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off));
_mm512_storeu_si512((__m512i *)(tile + 64), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off));
hmask0 = _mm512_slli_epi16(hmask0, 4);
hmask1 = _mm512_slli_epi16(hmask1, 4);
bytes = _mm512_loadu_si512(pb + 64);
r0 = _mm512_and_si512(bytes, lowMask);
r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
h0 = _mm512_and_si512(hbits, hmask0);
h1 = _mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), 2);
_mm512_storeu_si512((__m512i *)(tile + 128), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off));
_mm512_storeu_si512((__m512i *)(tile + 192), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off));
}
template <>
void unpack_B<block_iq4_xs>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
static const __m512i values128 = _mm512_set_epi8(
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127
);
const int packed_B_group_size = QK_K / 2 * TILE_N / 8;
const char * pb = (const char *)packed_B + k * packed_B_group_size;
const __m512i lowMask = _mm512_set1_epi8(0xF);
for (int n = 0; n < 8; n += 2) {
__m512i bytes = _mm512_loadu_si512(pb + n * 32);
const __m512i r0 = _mm512_shuffle_epi8(values128, _mm512_and_si512(bytes, lowMask));
const __m512i r1 = _mm512_shuffle_epi8(values128, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask));
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 0), r0);
_mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
}
}
template <typename TA, typename TB, bool is_acc>
struct acc_C {};
template <bool is_acc>
struct acc_C<block_q8_0, block_q4_0, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) {
const int offset = TILE_N * TILE_K / 2;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_1, block_q4_1, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_1 * A, int lda, const void * packed_B, int nr) {
const int offset = TILE_N * TILE_K / 2;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
vsum = _mm512_fmadd_ps(vm0, vs1, vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_0, block_q8_0, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) {
const int offset = TILE_N * TILE_K;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
for (int m = 0; m < nr; ++m) {
const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_K, block_q4_K, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N);
const uint8_t * mins = scales + 8 * TILE_N;
const ggml_half * d0 = reinterpret_cast<const ggml_half *>(mins + 8 * TILE_N);
const ggml_half * dmin = d0 + TILE_N;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin));
for (int m = 0; m < nr; ++m) {
const float d1 = A[m * lda].d;
const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin);
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
__m512i acc_m = _mm512_setzero_si512();
for (int k = 0; k < 4; ++k) {
__m512i vmask = _mm512_set1_epi32(k);
__m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s));
__m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32)));
acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
}
vsum = _mm512_fmadd_ps(vtile, vd, vsum);
vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_K, block_q5_K, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N);
const uint8_t * mins = scales + 8 * TILE_N;
const ggml_half * d0 = reinterpret_cast<const ggml_half *>(mins + 8 * TILE_N);
const ggml_half * dmin = d0 + TILE_N;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin));
for (int m = 0; m < nr; ++m) {
const float d1 = A[m * lda].d;
const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin);
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
__m512i acc_m = _mm512_setzero_si512();
for (int k = 0; k < 4; ++k) {
__m512i vmask = _mm512_set1_epi32(k);
__m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s));
__m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32)));
acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
}
vsum = _mm512_fmadd_ps(vtile, vd, vsum);
vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_K, block_q6_K, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N);
const ggml_half * d0 = reinterpret_cast<const ggml_half *>(scales + 16 * TILE_N);
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
for (int m = 0; m < nr; ++m) {
const float d1 = A[m * lda].d;
const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
vsum = _mm512_fmadd_ps(vtile, vd, vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <bool is_acc>
struct acc_C<block_q8_K, block_iq4_xs, is_acc> {
static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
const int8_t * scales = reinterpret_cast<const int8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N);
const ggml_half * d0 = reinterpret_cast<const ggml_half *>(scales + 8 * TILE_N);
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
for (int m = 0; m < nr; ++m) {
const float d1 = A[m * lda].d;
const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
__m512 vsum;
if (is_acc) {
vsum = _mm512_loadu_ps(C + m * ldc);
} else {
vsum = _mm512_set1_ps(0.f);
}
vsum = _mm512_fmadd_ps(vtile, vd, vsum);
_mm512_storeu_ps(C + m * ldc, vsum);
}
}
};
template <typename TB> constexpr int get_quants_size();
template <> constexpr int get_quants_size<block_q4_K>() { return (QK_K / 2) * TILE_N; }
template <> constexpr int get_quants_size<block_q5_K>() { return (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; }
template <> constexpr int get_quants_size<block_q6_K>() { return (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; }
template <> constexpr int get_quants_size<block_iq4_xs>() { return (QK_K / 2) * TILE_N; }
// used for QKK format
template <typename TB, bool is_acc,
typename std::enable_if<is_type_qkk<TB>::value, int>::type = 0>
inline void scale_C(const int32_t * RESTRICT tile, int32_t * RESTRICT sumi, const void * packed_B, int k, int nr) {
const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + get_quants_size<TB>());
const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(scales + k * TILE_N)));
for (int m = 0; m < nr; ++m) {
__m512i vsumi;
if (is_acc) {
vsumi = _mm512_loadu_si512(sumi + m * TILE_N);
} else {
vsumi = _mm512_setzero_si512();
}
__m512i vtile = _mm512_loadu_si512(tile + m * TILE_N);
vsumi = _mm512_add_epi32(vsumi, _mm512_mullo_epi32(vtile, vscale));
_mm512_storeu_si512((__m512i *)(sumi + m * TILE_N), vsumi);
}
}
template <typename TA, typename TB, typename TC, int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_avx {
static void apply(int K, const TA * RESTRICT A, const TB * RESTRICT B, TC * RESTRICT C, int ldc) {
GGML_UNUSED(K);
GGML_UNUSED(A);
GGML_UNUSED(B);
GGML_UNUSED(C);
GGML_UNUSED(ldc);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int K, const float * RESTRICT A, const ggml_fp16_t * RESTRICT B, float * RESTRICT C, int ldc) {
constexpr int ROWS = BLOCK_M;
constexpr int COLS = BLOCK_N;
assert(BLOCK_K == 16);
__m512 va;
__m512 vb[COLS];
__m512 vc[ROWS * COLS];
auto loadc = [&](auto idx) {
vc[idx] = _mm512_setzero_ps();
};
Unroll<ROWS * COLS>{}(loadc);
auto compute = [&](auto idx, auto k) {
constexpr int row = idx / COLS;
constexpr int col = idx % COLS;
if constexpr (col == 0) {
va = _mm512_loadu_ps(A + row * K + k);
}
if constexpr (row == 0) {
vb[col] = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k)));
}
vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]);
};
for (int k = 0; k < K; k += 16) {
Unroll<ROWS * COLS>{}(compute, k);
}
auto storec = [&](auto idx) {
constexpr int row = idx / COLS;
constexpr int col = idx % COLS;
C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]);
};
Unroll<ROWS * COLS>{}(storec);
}
};
#define LAUNCH_TINYGEMM_KERNEL_AVX(MB_SIZE, NB_SIZE) \
tinygemm_kernel_avx<float, type, float, MB_SIZE, NB_SIZE, blck_size>::apply( \
K, (const float *)src1->data + mb_start * K, \
(const type *)src0->data + nb_start * K, \
(float *)dst->data + mb_start * ldc + nb_start, ldc);
// re-organize in the format {NB, KB, TILE_SIZE}:
#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) {
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(NB, [&](int begin, int end) {
for (int n = begin; n < end; ++n) {
for (int k = 0; k < KB; ++k) {
int n0 = n * TILE_N;
pack_B((char *)packed_B + PACKED_INDEX(n, k, KB, TILE_SIZE), &B[n0 * KB + k], KB);
}
}
});
}
template <typename TA, typename TB, typename TC, int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni {};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q4_0);
const block_q8_0 * RESTRICT A = static_cast<const block_q8_0 *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
__m512i va[8];
__m512 vc[COLS];
__m512 vd1;
// sum of offsets, shared across COLS
//
// avx512-vnni does not have `_mm512_dpbssd_epi32`,
// need to transfrom ss to us:
// a * (b - 8) is equavilent to b * a - 8 * a
// s u u u s u s
//
__m512i vcomp;
const __m512i off = _mm512_set1_epi8(8);
const __m512i lowMask = _mm512_set1_epi8(0xF);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
auto compute = [&](auto col, auto i) {
// load a and compute compensation
if constexpr (col == 0) {
const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
vcomp = _mm512_setzero_si512();
for (int k = 0; k < 8; ++k) {
va[k] = _mm512_set1_epi32(a_ptr[k]);
vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
}
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
}
// load b
__m512i vsum = _mm512_setzero_si512();
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
for (int k = 0; k < 8; k += 2) {
__m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32));
__m512i vb0 = _mm512_and_si512(bytes, lowMask);
vsum = _mm512_dpbusd_epi32(vsum, vb0, va[k + 0]);
__m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va[k + 1]);
}
const int offset = TILE_N * TILE_K / 2;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
vsum = _mm512_sub_epi32(vsum, vcomp);
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q4_1);
const block_q8_1 * RESTRICT A = static_cast<const block_q8_1 *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
__m512i va[8];
__m512i vb[8];
__m512 vc[COLS];
__m512 vd1, vs1;
const __m512i lowMask = _mm512_set1_epi8(0xF);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
auto compute = [&](auto col, auto i) {
// load a
if constexpr (col == 0) {
const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
for (int k = 0; k < 8; ++k) {
va[k] = _mm512_set1_epi32(a_ptr[k]);
}
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
}
// load b
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
for (int k = 0; k < 8; k += 2) {
__m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32));
vb[k + 0] = _mm512_and_si512(bytes, lowMask);
vb[k + 1] = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
}
const int offset = TILE_N * TILE_K / 2;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset + TILE_N * sizeof(ggml_half))));
__m512i vsum = _mm512_setzero_si512();
for (int k = 0; k < 8; ++k) {
vsum = _mm512_dpbusd_epi32(vsum, vb[k], va[k]);
}
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
vc[col] = _mm512_fmadd_ps(vm0, vs1, vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q8_0) + TILE_N * sizeof(int32_t);
const block_q8_0 * RESTRICT A = static_cast<const block_q8_0 *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
__m512i va[8];
__m512i vb[8];
__m512 vc[COLS];
__m512 vd1;
// Notes: s8s8 igemm compensation in avx512-vnni
// change s8s8 to u8s8 with compensate
// a * b = (a + 128) * b - 128 * b
// s s u s u s
//
// (128 * b is pre-computed when packing B to vnni formats)
//
const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
auto compute = [&](auto col, auto i) {
// load a and add offset 128
if constexpr (col == 0) {
const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
for (int k = 0; k < 8; ++k) {
va[k] = _mm512_set1_epi32(a_ptr[k]);
va[k] = _mm512_add_epi8(va[k], off);
}
vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
}
// load b
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
for (int k = 0; k < 8; ++k) {
vb[k] = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 64));
}
const int offset = TILE_N * TILE_K;
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
const int offset2 = TILE_N * TILE_K + TILE_N * sizeof(ggml_half);
const __m512i vcomp = _mm512_loadu_si512((const __m512i *)(b_ptr + offset2));
__m512i vsum = _mm512_setzero_si512();
for (int k = 0; k < 8; ++k) {
vsum = _mm512_dpbusd_epi32(vsum, va[k], vb[k]);
}
vsum = _mm512_sub_epi32(vsum, vcomp);
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q4_K) + TILE_N * 4;
const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
// a.qs: 8 groups, 32 bytes each group (m256i)
__m512i va[8];
// a.bsum: 8 groups, 2 bytes each group (m128i)
__m512i va_bsum;
__m512 vc[COLS];
__m512 vd1;
// packed_B:
const int offset_scales = (QK_K / 2) * TILE_N;
const int offset_mins = (QK_K / 2) * TILE_N + 8 * TILE_N;
const int offset_d0 = (QK_K / 2) * TILE_N + 16 * TILE_N;
const int offset_dmin = (QK_K / 2) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half);
const __m512i lowMask = _mm512_set1_epi8(0xF);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
// Notes: vnni formats in QK_K
// a) quants vnni format
// int8 {k/4, n, 4}, viewed as 2d {k/4, 4n}, k = 32
// from {16, 32} to {8, 64}
//
// b) min vnni format
// int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8
// from {16, 8} to {4, 32}
//
auto compute = [&](auto col, auto i) {
// load a
if constexpr (col == 0) {
for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
}
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
va_bsum = _mm512_castsi128_si512(q8s);
vd1 = _mm512_set1_ps(A[0 * KB + i].d);
}
// step 1: accumultate the quants
__m512i acc = _mm512_setzero_si512();
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
const char * b_qs = b_ptr;
for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
__m512i vsum = _mm512_setzero_si512();
for (int k = 0; k < 8; k += 2) {
__m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]);
__m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]);
__m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs);
__m512i vb0 = _mm512_and_si512(bytes, lowMask);
vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
__m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
b_qs += 64;
}
// vacc += scale * (q8 @ q4)
const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
}
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
// step 2: accumulate the mins
__m512i acc_m = _mm512_setzero_si512();
for (int k = 0; k < 4; ++k) {
__m512i vmask = _mm512_set1_epi32(k);
__m512i va = _mm512_permutexvar_epi32(vmask, va_bsum);
__m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32)));
acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
}
const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin)));
vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q5_K) + TILE_N * 4;
const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
// a.qs: 8 groups, 32 bytes each group (m256i)
__m512i va[8];
// a.bsum: 8 groups, 2 bytes each group (m128i)
__m512i va_bsum;
__m512 vc[COLS];
__m512 vd1;
// packed_B:
const int offset_qh = (QK_K / 2) * TILE_N;
const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N;
const int offset_mins = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 8 * TILE_N;
const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N;
const int offset_dmin = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half);
const __m512i lowMask = _mm512_set1_epi8(0xF);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
// Q5_K and Q4_K shares the same vnni formats, refer to notes above.
auto compute = [&](auto col, auto i) {
// load a
if constexpr (col == 0) {
for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
}
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
va_bsum = _mm512_castsi128_si512(q8s);
vd1 = _mm512_set1_ps(A[0 * KB + i].d);
}
// step 1: accumultate the quants
__m512i acc = _mm512_setzero_si512();
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
const char * b_qs = b_ptr;
const char * b_qh = b_ptr + offset_qh;
for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
__m512i vsum = _mm512_setzero_si512();
__m512i hmask0 = _mm512_set1_epi8(0x1);
__m512i hmask1 = _mm512_set1_epi8(0x2);
__m512i hbits = _mm512_loadu_si512((const __m512i *)(b_qh + k_group * 64));
for (int k = 0; k < 8; k += 2) {
__m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]);
__m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]);
__m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs);
__m512i vb0 = _mm512_and_si512(bytes, lowMask);
__m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
__m512i vh0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), k), 4);
__m512i vh1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), k + 1), 4);
hmask0 = _mm512_slli_epi16(hmask0, 2);
hmask1 = _mm512_slli_epi16(hmask1, 2);
vb0 = _mm512_add_epi8(vb0, vh0);
vb1 = _mm512_add_epi8(vb1, vh1);
vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
b_qs += 64;
}
// vacc += scale * (q8 @ q5)
const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
}
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
// step 2: accumulate the mins
__m512i acc_m = _mm512_setzero_si512();
for (int k = 0; k < 4; ++k) {
__m512i vmask = _mm512_set1_epi32(k);
__m512i va = _mm512_permutexvar_epi32(vmask, va_bsum);
__m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32)));
acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
}
const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin)));
vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_K, block_q6_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_q6_K);
const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
// load the 256 bytes from A to 4 avx512 vectors
__m512i va[4];
__m512 vc[COLS];
__m512 vd1;
// packed_B:
const int offset_qh = (QK_K / 2) * TILE_N;
const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N;
const int offset_d0 = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N + 16 * TILE_N;
// compensation
__m512i vcomp;
const __m512i m32s = _mm512_set1_epi32(32);
const __m512i lowMask = _mm512_set1_epi8(0xF);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
auto compute = [&](auto col, auto i) {
if constexpr (col == 0) {
// load a
va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0));
va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64));
va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128));
va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192));
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
vcomp = _mm512_mullo_epi32(_mm512_cvtepi16_epi32(q8sums), m32s);
vd1 = _mm512_set1_ps(A[0 * KB + i].d);
}
// accmulate the quants
__m512i acc = _mm512_setzero_si512();
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
const char * b_qs = b_ptr;
const char * b_qh = b_ptr + offset_qh;
int mask = 0;
for (int k_group = 0; k_group < QK_K / 16; ++k_group) {
int r = k_group >> 2;
__m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
__m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
__m512i vsum = _mm512_setzero_si512();
__m512i hmask = _mm512_set1_epi8(0x3);
__m512i bytes = _mm512_loadu_si512(b_qs);
__m512i hbits = _mm512_loadu_si512(b_qh);
__m512i vb0 = _mm512_and_si512(bytes, lowMask);
__m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
__m512i vh0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask), 4);
__m512i vh1 = _mm512_slli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 2)), 2);
vb0 = _mm512_add_epi8(vb0, vh0);
vb1 = _mm512_add_epi8(vb1, vh1);
vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
b_qs += 64;
va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
bytes = _mm512_loadu_si512(b_qs);
vb0 = _mm512_and_si512(bytes, lowMask);
vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
vh0 = _mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 4));
vh1 = _mm512_srli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 6)), 2);
vb0 = _mm512_add_epi8(vb0, vh0);
vb1 = _mm512_add_epi8(vb1, vh1);
vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
b_qs += 64;
b_qh += 64;
// B * A - 32 * A
__m512i vmask = _mm512_set1_epi32(k_group);
vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp));
// vacc += scale * (q8 @ q6)
const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
}
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](int col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, BLOCK_K> {
static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
constexpr int COLS = BLOCK_N / 16;
const int TILE_SIZE = TILE_N * sizeof(block_iq4_xs) + TILE_N * 2;
const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
// load the 256 bytes from A to 4 avx512 vectors
__m512i va[4];
__m512 vc[COLS];
__m512 vd1;
// packed_B:
const int offset_scales = (QK_K / 2) * TILE_N ;
const int offset_d0 = (QK_K / 2) * TILE_N + 8 * TILE_N;
// compensation
__m512i vcomp;
const __m256i m128s = _mm256_set1_epi16(128);
const __m512i lowMask = _mm512_set1_epi8(0xF);
const __m512i values128 = _mm512_set_epi8(
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127
);
const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
const __m512i values256 = _mm512_add_epi8(values128, off);
auto loadc = [&](auto col) {
vc[col] = _mm512_setzero_ps();
};
Unroll<COLS>{}(loadc);
auto compute = [&](auto col, auto i) {
if constexpr (col == 0) {
// load a
va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 0));
va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 64));
va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128));
va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192));
// compensation: 128 * A
const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
vcomp = _mm512_castsi256_si512(_mm256_madd_epi16(q8sums, m128s));
vd1 = _mm512_set1_ps(A[0 * KB + i].d);
}
// accmulate the quants
__m512i acc = _mm512_setzero_si512();
const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
const char * b_qs = b_ptr;
int mask = 0;
for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
int r = k_group >> 1;
__m512i vmask = _mm512_set1_epi32(k_group);
__m512i vsum = _mm512_setzero_si512();
for (int k = 0; k < 8; k += 2) {
__m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
__m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
__m512i bytes = _mm512_loadu_si512(b_qs);
__m512i vb0 = _mm512_shuffle_epi8(values256, _mm512_and_si512(bytes, lowMask));
__m512i vb1 = _mm512_shuffle_epi8(values256, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask));
vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
b_qs += 64;
}
// (B + 128) * A - 128 * A
vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp));
// vacc += scale * (q8 @ q4)
const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
}
const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
};
for (int i = 0; i < KB; ++i) {
Unroll<COLS>{}(compute, i);
}
//store to C
auto storec = [&](auto col) {
_mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
};
Unroll<COLS>{}(storec);
}
};
#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE) \
tinygemm_kernel_vnni<vec_dot_type, type, float, 1, NB_SIZE, blck_size>::apply( \
KB, (const char *)wdata + 0 * row_size_A, \
(const char *)src0->data + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE), \
(float *) dst->data + 0 * N + nb_start, ldc)
template <typename TA, typename TB, typename TC, int BLOCK_K,
typename std::enable_if<!is_type_qkk<TB>::value, int>::type = 0>
void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, TC * RESTRICT C, int ldc) {
using packed_B_t = packed_B_type<TB>;
const int TILE_SIZE = get_tile_size<TB>();
const bool need_unpack = do_unpack<TB>::value;
GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N);
const TA * RESTRICT A = static_cast<const TA *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
const int m0 = std::min(M, TILE_M);
const int m1 = std::max(M - TILE_M, 0);
const int lda = KB * sizeof(TA);
//const int ldb = KB * sizeof(TB);
static thread_local packed_B_t Tile0[TILE_N * TILE_K];
static thread_local packed_B_t Tile1[TILE_N * TILE_K];
static thread_local int8_t Tile23[TILE_M * TILE_K];
static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
// double buffering C to interleave avx512 and amx
int32_t * C_cur = TileC0;
int32_t * C_pre = TileC1;
auto Tile4 = [&](int32_t * base) { return base; };
auto Tile5 = [&](int32_t * base) { return base + TILE_M * TILE_N; };
auto Tile6 = [&](int32_t * base) { return base + 2 * TILE_M * TILE_N; };
auto Tile7 = [&](int32_t * base) { return base + 3 * TILE_M * TILE_N; };
if (M == 2 * TILE_M) {
// i = 0
const char * B_blk0 = B + PACKED_INDEX(0, 0, KB, TILE_SIZE);
const char * B_blk1 = B + PACKED_INDEX(1, 0, KB, TILE_SIZE);
if (need_unpack) {
unpack_B<TB>(Tile0, B_blk0);
_tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
}
_tile_zero(TMM4);
_tile_loadd(TMM2, A[0].qs, lda);
_tile_dpbssd(TMM4, TMM2, TMM0);
_tile_stored(TMM4, Tile4(C_pre), TILE_N * sizeof(int32_t));
_tile_zero(TMM5);
_tile_loadd(TMM3, A[TILE_M * KB + 0].qs, lda);
_tile_dpbssd(TMM5, TMM3, TMM0);
_tile_stored(TMM5, Tile5(C_pre), TILE_N * sizeof(int32_t));
if (need_unpack) {
unpack_B<TB>(Tile1, B_blk0);
_tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
}
_tile_zero(TMM6);
_tile_dpbssd(TMM6, TMM2, TMM1);
_tile_stored(TMM6, Tile6(C_pre), TILE_N * sizeof(int32_t));
_tile_zero(TMM7);
_tile_dpbssd(TMM7, TMM3, TMM1);
_tile_stored(TMM7, Tile7(C_pre), TILE_N * sizeof(int32_t));
for (int i = 1; i < KB; ++i) {
// index of previous iter
const int ii = i - 1;
const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE);
const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE);
GGML_DISPATCH_BOOL(ii > 0, is_acc, [&] {
if (need_unpack) {
unpack_B<TB>(Tile0, B_blk0);
_tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
}
_tile_zero(TMM4);
_tile_loadd(TMM2, A[i].qs, lda);
acc_C<TA, TB, is_acc>::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
_tile_dpbssd(TMM4, TMM2, TMM0);
_tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t));
_tile_zero(TMM5);
_tile_loadd(TMM3, A[TILE_M * KB + i].qs, lda);
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
_tile_dpbssd(TMM5, TMM3, TMM0);
_tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t));
if (need_unpack) {
unpack_B<TB>(Tile1, B_blk1);
_tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
}
_tile_zero(TMM6);
acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
_tile_dpbssd(TMM6, TMM2, TMM1);
_tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t));
_tile_zero(TMM7);
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
_tile_dpbssd(TMM7, TMM3, TMM1);
_tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t));
std::swap(C_cur, C_pre);
});
}
// final accumulation
{
int ii = KB - 1;
acc_C<TA, TB, true>::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
acc_C<TA, TB, true>::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
acc_C<TA, TB, true>::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
acc_C<TA, TB, true>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
}
} else {
for (int i = 0; i < KB; ++i) {
_tile_zero(TMM4);
_tile_zero(TMM6);
if (m1 != 0) {
_tile_zero(TMM5);
_tile_zero(TMM7);
}
const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE);
const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE);
if (need_unpack) {
unpack_B<TB>(Tile0, B_blk0);
_tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
}
if (need_unpack) {
unpack_B<TB>(Tile1, B_blk1);
_tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
} else {
_tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
}
if (m0 == TILE_M) {
_tile_loadd(TMM2, A[i].qs, lda);
} else {
unpack_A(Tile23, &A[i], KB, m0);
_tile_loadd(TMM2, Tile23, TILE_K);
}
_tile_dpbssd(TMM4, TMM2, TMM0);
_tile_dpbssd(TMM6, TMM2, TMM1);
_tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t));
_tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t));
GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
acc_C<TA, TB, is_acc>::apply(C, ldc, Tile4(C_cur), &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0);
acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Tile6(C_cur), &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0);
});
if (m1 != 0) {
unpack_A(Tile23, &A[TILE_M * KB + i], KB, m1);
_tile_loadd(TMM3, Tile23, TILE_K);
_tile_dpbssd(TMM5, TMM3, TMM0);
_tile_dpbssd(TMM7, TMM3, TMM1);
_tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t));
_tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t));
GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc, ldc, Tile5(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1);
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1);
});
}
}
}
return;
}
template <typename TA, typename TB, typename TC, int BLOCK_K,
typename std::enable_if<is_type_qkk<TB>::value, int>::type = 0>
void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
static_assert(std::is_same<TA, block_q8_K>::value);
const int TILE_SIZE = get_tile_size<TB>();
GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N);
const TA * RESTRICT A = static_cast<const TA *>(_A);
const char * RESTRICT B = static_cast<const char *>(_B);
const int m0 = std::min(M, TILE_M);
const int m1 = std::max(M - TILE_M, 0);
//const int lda = KB * sizeof(TA);
static thread_local int8_t Tile0[TILE_N * TILE_K];
static thread_local int8_t Tile1[TILE_N * TILE_K];
static thread_local int8_t Tile23[TILE_M * TILE_K];
// mat mul result for each group
static thread_local int32_t Tile4[TILE_M * TILE_N];
static thread_local int32_t Tile5[TILE_M * TILE_N];
static thread_local int32_t Tile6[TILE_M * TILE_N];
static thread_local int32_t Tile7[TILE_M * TILE_N];
// sum of each QK_K block, contains 8 groups, int32
static thread_local int32_t Sumi4[TILE_M * TILE_N];
static thread_local int32_t Sumi5[TILE_M * TILE_N];
static thread_local int32_t Sumi6[TILE_M * TILE_N];
static thread_local int32_t Sumi7[TILE_M * TILE_N];
const int k_group_size = std::is_same<TB, block_q6_K>::value ? 16 : 32;
for (int i = 0; i < KB; ++i) {
// step 1: accumulate the quants across 8 groups, each group with 32
for (int k = 0; k < QK_K / k_group_size; ++k) {
GGML_DISPATCH_BOOL(k > 0, is_acc, [&] {
_tile_zero(TMM4);
_tile_zero(TMM6);
unpack_B<TB>(Tile0, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k);
_tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
unpack_B<TB>(Tile1, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k);
_tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
unpack_A<TB>(Tile23, &A[i], KB, k, m0);
_tile_loadd(TMM2, Tile23, TILE_K);
_tile_dpbssd(TMM4, TMM2, TMM0);
_tile_dpbssd(TMM6, TMM2, TMM1);
_tile_stored(TMM4, Tile4, TILE_N * sizeof(int32_t));
_tile_stored(TMM6, Tile6, TILE_N * sizeof(int32_t));
scale_C<TB, is_acc>(Tile4, Sumi4, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m0);
scale_C<TB, is_acc>(Tile6, Sumi6, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m0);
if (m1 != 0) {
_tile_zero(TMM5);
_tile_zero(TMM7);
unpack_A<TB>(Tile23, &A[TILE_M * KB + i], KB, k, m1);
_tile_loadd(TMM3, Tile23, TILE_K);
_tile_dpbssd(TMM5, TMM3, TMM0);
_tile_dpbssd(TMM7, TMM3, TMM1);
_tile_stored(TMM5, Tile5, TILE_N * sizeof(int32_t));
_tile_stored(TMM7, Tile7, TILE_N * sizeof(int32_t));
scale_C<TB, is_acc>(Tile5, Sumi5, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m1);
scale_C<TB, is_acc>(Tile7, Sumi7, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m1);
}
});
}
// step 2: accmulate the mins
GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
acc_C<TA, TB, is_acc>::apply(C, ldc, Sumi4, &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0);
acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Sumi6, &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0);
if (m1 != 0) {
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc, ldc, Sumi5, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1);
acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Sumi7, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1);
}
});
}
return;
}
} // anonymous namespace
// get the packed tensor size for quantized weights
size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) {
const enum ggml_type TYPE = tensor->type;
const int K = tensor->ne[0]; // ne0: in_features
const int N = tensor->ne[1]; // ne1: out_features
auto get_tensor_size = [&] {
size_t row_size_B{0};
GGML_DISPATCH_QTYPES(TYPE, [&] {
row_size_B = get_row_size<type, blck_size>(K);
});
return N * row_size_B;
};
if (qtype_has_amx_kernels(TYPE)) {
return get_tensor_size();
} else {
// for f16, bf16 we don't do packing
return ggml_nbytes(tensor);
}
}
// pack weight to vnni format
void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0 && size == ggml_nbytes(tensor)); // only full tensor conversion is supported for now
const enum ggml_type TYPE = tensor->type;
const int K = tensor->ne[0]; // ne0: in_features
const int N = tensor->ne[1]; // ne1: out_features
GGML_DISPATCH_QTYPES(TYPE, [&] {
convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K);
});
}
size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
struct ggml_tensor * src0 = dst->src[0];
const enum ggml_type TYPE = src0->type;
const bool is_floating_type = TYPE == GGML_TYPE_F16;
if (is_floating_type) {
return 0;
}
const int M = dst->ne[1];
const int K = src0->ne[0];
size_t desired_wsize = 0;
GGML_DISPATCH_QTYPES(TYPE, [&] {
const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
desired_wsize = M * row_size_A;
});
return desired_wsize;
}
// NB: mixed dtype gemm with Advanced Matrix Extensions (Intel AMX)
//
// src0: weight in shape of {N, K}, quantized
// src1: input in shape of {M, K}, float32
// dst: output in shape of {M, N}, float32
//
// the function performs: dst = src1 @ src0.T
//
void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_tensor * dst) {
struct ggml_tensor * src0 = dst->src[0];
struct ggml_tensor * src1 = dst->src[1];
const enum ggml_type TYPE = src0->type;
// f16 only has avx512 kernels for now,
// amx kernels will be added once 6th gen xeon is released.
const bool is_floating_type = TYPE == GGML_TYPE_F16;
const int M = dst->ne[1];
const int N = dst->ne[0];
const int K = src0->ne[0];
const int ldc = dst->nb[1] / dst->nb[0];
if (is_floating_type) {
constexpr int BLOCK_M = 4;
constexpr int BLOCK_N = 6;
const int MB = div_up(M, BLOCK_M);
const int NB = div_up(N, BLOCK_N);
parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] {
for (int i = begin; i < end; ++i) {
int mb = i / NB;
int nb = i % NB;
int mb_start = mb * BLOCK_M;
int mb_size = std::min(BLOCK_M, M - mb_start);
int nb_start = nb * BLOCK_N;
int nb_size = std::min(BLOCK_N, N - nb_start);
switch (mb_size << 4 | nb_size) {
case 0x12: LAUNCH_TINYGEMM_KERNEL_AVX(1, 2); break;
case 0x14: LAUNCH_TINYGEMM_KERNEL_AVX(1, 4); break;
case 0x16: LAUNCH_TINYGEMM_KERNEL_AVX(1, 6); break;
case 0x22: LAUNCH_TINYGEMM_KERNEL_AVX(2, 2); break;
case 0x24: LAUNCH_TINYGEMM_KERNEL_AVX(2, 4); break;
case 0x26: LAUNCH_TINYGEMM_KERNEL_AVX(2, 6); break;
case 0x32: LAUNCH_TINYGEMM_KERNEL_AVX(3, 2); break;
case 0x34: LAUNCH_TINYGEMM_KERNEL_AVX(3, 4); break;
case 0x36: LAUNCH_TINYGEMM_KERNEL_AVX(3, 6); break;
case 0x42: LAUNCH_TINYGEMM_KERNEL_AVX(4, 2); break;
case 0x44: LAUNCH_TINYGEMM_KERNEL_AVX(4, 4); break;
case 0x46: LAUNCH_TINYGEMM_KERNEL_AVX(4, 6); break;
default: fprintf(stderr, "Unexpected block size!\n");
}
}
});
});
return;
}
// pointer to work space, used convert A from float to quantized type
void * wdata = params->wdata;
//TODO: performance improvement: merge quant A
if (params->ith == 0) {
GGML_DISPATCH_QTYPES(TYPE, [&] {
const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
const size_t desired_wsize = M * row_size_A;
if (params->wsize < desired_wsize) {
GGML_ABORT("insufficient work space size");
}
// Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size
// Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
const float * A_data = static_cast<const float *>(src1->data);
for (int m = 0; m < M; ++m) {
from_float<vec_dot_type>(A_data + m * K, (char *)wdata + m * row_size_A, K);
}
});
}
ggml_barrier(params->threadpool);
if (M == 1) {
// MB = 1 and handle 8 tiles in each block
constexpr int kTilesN = 4;
constexpr int BLOCK_N = TILE_N * kTilesN;
const int NB = div_up(N, BLOCK_N);
parallel_for_ggml(params, NB, [&](int begin, int end) {
GGML_DISPATCH_QTYPES(TYPE, [&] {
const int KB = K / blck_size;
const int TILE_SIZE = get_tile_size<type>();
const int row_size_A = KB * sizeof(vec_dot_type);
for (int i = begin; i < end; ++i) {
int nb = i;
int nb_start = nb * BLOCK_N;
int nb_size = std::min(BLOCK_N, N - nb_start); // 32, 64, 96
switch (nb_size) {
//case 160: LAUNCH_TINYGEMM_KERNEL_VNNI(160); break;
case 128: LAUNCH_TINYGEMM_KERNEL_VNNI(128); break;
case 96: LAUNCH_TINYGEMM_KERNEL_VNNI(96); break;
case 64: LAUNCH_TINYGEMM_KERNEL_VNNI(64); break;
case 32: LAUNCH_TINYGEMM_KERNEL_VNNI(32); break;
default: fprintf(stderr, "Unexpected n block size!\n");
}
}
});
});
return;
}
// handle 4 tiles at a tile
constexpr int BLOCK_M = TILE_M * 2;
constexpr int BLOCK_N = TILE_N * 2;
const int MB = div_up(M, BLOCK_M);
const int NB = div_up(N, BLOCK_N);
parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
// init tile config for each thread
ggml_tile_config_init();
GGML_DISPATCH_QTYPES(TYPE, [&] {
const int KB = K / blck_size;
const int TILE_SIZE = get_tile_size<type>();
const int row_size_A = KB * sizeof(vec_dot_type);
for (int i = begin; i < end; ++i) {
int mb = i / NB;
int nb = i % NB;
int mb_start = mb * BLOCK_M;
int mb_size = std::min(BLOCK_M, M - mb_start);
int nb_start = nb * BLOCK_N;
int nb_size = BLOCK_N;
tinygemm_kernel_amx<vec_dot_type, type, float, blck_size>(
mb_size, nb_size, KB,
(const char *)wdata + mb_start * row_size_A,
(const char *)src0->data + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE),
(float *) dst->data + mb_start * N + nb_start, ldc);
}
});
});
}
#endif // if defined(__AMX_INT8__) && defined(__AVX512VNNI__)