ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 07:04:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

q4_1: 135 t/s -> 262 t/s

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-06-18 11:30:34 +03:00
parent 0788a71fcf
commit 8d393fccf5
2 changed files with 39 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
ggml/src/iqk/iqk_gemm_legacy_quants.cpp
View File

@@ -1674,7 +1674,8 @@ static void mul_mat_q8_1_r8_q8_2(int n, const void * vx, size_t bx, const DataIn
for (int iy = 0; iy < nrc_y; ++iy) {
auto scales = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i *)q8.y[iy][i4].d)), 16));
_mm_storeu_ps(d8 + 4*iy + 0, scales);
auto bsums4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i *)(q8.y[iy][i4].d+4))), 16));
auto bsums4 = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(_mm_loadl_epi64((const __m128i *)(q8.y[iy][i4].d+4))));
bsums4 = _mm_mul_ps(bsums4, scales);
auto bsums = _mm256_set_m128(bsums4, bsums4);
acc[iy] = _mm256_fmadd_ps(mx[0], _mm256_shuffle_ps(bsums, bsums, 0x00), acc[iy]);
acc[iy] = _mm256_fmadd_ps(mx[1], _mm256_shuffle_ps(bsums, bsums, 0x55), acc[iy]);
@@ -1775,6 +1776,39 @@ void iqk_convert_qX_q80_r8(int n, const void * vx, size_t bx, void * vy, int nrc
}
}
template <typename Block, typename Dequantizer>
void iqk_convert_qX_1_q8_1_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK8_0 == 0);
GGML_ASSERT(nrc_x%8 == 0);
int nb = n/QK8_0;
const Block * x8[8];
block_q8_1_r8 * y = (block_q8_1_r8 *)vy;
uint32_t block[8];
Dequantizer deq;
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const Block *)((const char *)vx + (ix + k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
y[i].d[k+0] = x8[k][i].d;
y[i].d[k+8] = x8[k][i].m;
_mm256_storeu_si256((__m256i *)block, deq.dequant(x8[k]+i));
auto qs = (uint32_t *)y[i].qs;
for (int l = 0; l < 4; ++l) {
qs[8*l + k + 0] = block[l + 0];
qs[8*l + k + 32] = block[l + 4];
}
}
}
y += nb;
}
}
template <typename Dequantizer> void set_functions(std::array<mul_mat_t, IQK_MAX_NY>& funcs) {
if constexpr (std::is_same_v<Dequantizer, Q4_0_Unpacker> || std::is_same_v<Dequantizer, Q5_0_Unpacker> ||
std::is_same_v<Dequantizer, Q8_0_Unpacker>) {
@@ -1801,7 +1835,9 @@ template <typename Dequantizer> void set_functions(std::array<mul_mat_t, IQK_MAX
bool iqk_convert_legacy_quants_q8_r8(int type, int n, const void * vx, size_t bx, void * vy, int nrc_x) {
switch (type) {
case GGML_TYPE_Q4_0 : iqk_convert_qX_q80_r8<block_q4_0, Q4_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q4_1 : iqk_convert_qX_1_q8_1_r8<block_q4_1, Q4_1_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q5_0 : iqk_convert_qX_q80_r8<block_q5_0, Q5_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
//case GGML_TYPE_Q5_1 : iqk_convert_qX_1_q8_1_r8<block_q5_1, Q5_1_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q6_0 : iqk_convert_qX_q80_r8<block_q6_0, Q6_0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_IQ4_NL: iqk_convert_qX_q80_r8<block_iq4_nl, IQ4_NL0_Dequantizer>(n, vx, bx, vy, nrc_x); break;
case GGML_TYPE_Q8_0 : iqk_convert_q80_q80_r8(n, vx, bx, vy, nrc_x); break;

3
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -259,6 +259,7 @@ struct MulMat {
case GGML_TYPE_IQ5_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ6_K : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_Q4_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q4_1 : return nrc_y >= 32 ? GGML_TYPE_Q8_1 : type;
case GGML_TYPE_Q5_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q6_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
@@ -409,7 +410,7 @@ bool iqk_convert_repack(int typeA, int n, const void * vx, size_t bx, void * vy,
case GGML_TYPE_IQ4_KT:
return iqk_dequantize_ktquants(typeA, n, vx, bx, vy, stride_y, nrc_x);
case GGML_TYPE_Q4_0:
//case GGML_TYPE_Q4_1:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
//case GGML_TYPE_Q5_1:
case GGML_TYPE_Q6_0: