ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-08 07:20:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

iqk_mul_mat: use block_q8_0_x4 also for AVX2

Browse Source
This commit is contained in:
Kawrakow
2024-06-08 08:20:26 +03:00
parent 30a0bf30fa
commit 5b19e5e4a9
2 changed files with 50 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
ggml-quants.c
View File

@@ -948,7 +948,15 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
}
}
#elif defined(__AVX2__) || defined(__AVX__)
block_q8_0_x4 * y4 = (block_q8_0_x4 *)vy;
int nb4 = 4*(nb/4);
#ifdef __AVX2__
const bool pack = true;
#else
const bool pack = false;
#endif
for (int i = 0; i < nb; i++) {
int i4 = i/4, ir = i%4;
// Load elements into 4 AVX vectors
__m256 v0 = _mm256_loadu_ps( x );
__m256 v1 = _mm256_loadu_ps( x + 8 );
@@ -970,7 +978,11 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = GGML_FP32_TO_FP16(d);
if (pack && i < nb4) {
y4[i4].d[ir] = GGML_FP32_TO_FP16(d);
} else {
y[i].d = GGML_FP32_TO_FP16(d);
}
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
@@ -1005,7 +1017,11 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
i0 = _mm256_permutevar8x32_epi32( i0, perm );
_mm256_storeu_si256((__m256i *)y[i].qs, i0);
if (i < nb4) {
_mm256_storeu_si256((__m256i *)y4[i4].qs + ir, i0);
} else {
_mm256_storeu_si256((__m256i *)y[i].qs, i0);
}
#else
// Since we don't have in AVX some necessary functions,
// we split the registers in half and call AVX2 analogs from SSE

41
iqk_mul_mat.cpp
View File

@@ -1749,16 +1749,39 @@ struct UnsignedDot {
template <typename Q8, typename Dot> struct Sum4 {
Dot dot;
inline __m256i compute(const __m256i * qx, const Q8 * y) const {
const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y[0].qs));
const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y[1].qs));
const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y[2].qs));
const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y[3].qs));
const __m256i p01 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p0, p1)); // 0,0, 1,1, 0,0, 1,1
const __m256i p23 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p2, p3)); // 2,2, 3,3, 2,2, 3,3
return _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p01, p23)); // 0,1,2,3, 0,1,2,3
if constexpr (std::is_same_v<Q8, block_q8_0>) {
const block_q8_0_x4 * y4 = (const block_q8_0_x4 *)y;
const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y4->qs+0));
const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y4->qs+1));
const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y4->qs+2));
const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y4->qs+3));
const __m256i p01 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p0, p1)); // 0,0, 1,1, 0,0, 1,1
const __m256i p23 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p2, p3)); // 2,2, 3,3, 2,2, 3,3
return _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p01, p23)); // 0,1,2,3, 0,1,2,3
} else {
const __m256i p0 = dot.compute(qx[0], _mm256_loadu_si256((const __m256i *)y[0].qs));
const __m256i p1 = dot.compute(qx[1], _mm256_loadu_si256((const __m256i *)y[1].qs));
const __m256i p2 = dot.compute(qx[2], _mm256_loadu_si256((const __m256i *)y[2].qs));
const __m256i p3 = dot.compute(qx[3], _mm256_loadu_si256((const __m256i *)y[3].qs));
const __m256i p01 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p0, p1)); // 0,0, 1,1, 0,0, 1,1
const __m256i p23 = _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p2, p3)); // 2,2, 3,3, 2,2, 3,3
return _mm256_madd_epi16(dot.helper.m1, _mm256_packs_epi32(p01, p23)); // 0,1,2,3, 0,1,2,3
}
}
};
struct ScaleHelperQ8_0 {
inline __m128 prepare4(const block_q8_0 * y) {
const block_q8_0_x4 * y4 = (const block_q8_0_x4 *)y;
return _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)y4->d));
}
inline __m128 prepare4(__m128 other_scales, const block_q8_0 * y) {
return _mm_mul_ps(other_scales, prepare4(y));
}
template <typename Q> inline float prepare1(const Q * y) const { return GGML_FP16_TO_FP32(y->d); }
template <typename Q> inline float prepare1(float d, const Q * y) const { return d*prepare1(y); }
};
struct ScaleHelperQ_0 {
ggml_half scales8[4];
template <typename Q>
@@ -1893,11 +1916,11 @@ void mul_mat_qX_0_q8_0_T(int n, const void * vx, size_t bx, const DataInfo& info
Q8<nrc_y, block_q8_0> q8(info);
int nb = n/Unpacker::block_size();
if (nb%4 == 0) {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, true>, ScaleHelperQ_0, block_q8_0, nrc_y>(
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, true>, ScaleHelperQ8_0, block_q8_0, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x
);
} else {
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, false>, ScaleHelperQ_0, block_q8_0, nrc_y>(
mul_mat_qX_q8_Helper<Unpacker, Sum4Type0, AccumType0<nrc_y, false>, ScaleHelperQ8_0, block_q8_0, nrc_y>(
nb, vx, bx, info, q8.y, nrc_x
);
}