ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-13 07:20:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable next_128() also on AVX2

Browse Source 
Despite having just 16 vector registers it is still faster.
This commit is contained in:
Iwan Kawrakow
2025-06-20 08:34:45 +03:00
parent 1287a66627
commit 5b677c3caf
2 changed files with 19 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
ggml/src/iqk/iqk_gemm_ktquants.cpp
View File

@@ -171,8 +171,7 @@ struct Trellis3 {
}
}
IQK_ALWAYS_INLINE inline void next_128(const uint32_t * val, __m256i * result) const {
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
// On AVX2 we don't have enough vector registers to do this
// Even though we only have 16 vector registers nn AVX2, this is still faster
__m256i aux[16];
auto perm = _mm256_setr_epi32(0, 2, 4, 6, 1, 3, 5, 7);
for (int k = 0; k < 4; ++k) {
@@ -191,9 +190,16 @@ struct Trellis3 {
aux[i] = _mm256_and_si256(aux[i], mask);
}
auto offset = _mm256_set1_epi32(-126);
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
auto m1 = _mm256_set1_epi32(0x01010101);
#endif
for (int i = 0; i < 16; ++i) {
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
aux[i] = _mm256_dpbusd_epi32(offset, aux[i], m1);
#else
auto dot = _mm256_maddubs_epi16(aux[i], _mm256_set1_epi32(0x01010101));
aux[i] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot, _mm256_set1_epi16(1)));
#endif
}
for (int k = 0; k < 4; ++k) {
auto v1 = _mm256_packs_epi32(aux[4*k+0], aux[4*k+1]);
@@ -205,13 +211,9 @@ struct Trellis3 {
result[k] = _mm256_sign_epi8(result[k], result[k]);
}
}
#else
for (int k = 0; k < 4; ++k) result[k] = next32(val + 8*k);
#endif
}
IQK_ALWAYS_INLINE inline void next_128(const uint16_t * val, uint32_t v0, __m256i * result) const {
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
// On AVX2 we don't have enough vector registers to do this
// Even though we only have 16 vector registers nn AVX2, this is still faster
__m256i aux[16];
for (int k = 0; k < 4; ++k) {
auto v128 = _mm_add_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i *)(val + 4*k))), _mm_set1_epi32(v0));
@@ -229,9 +231,16 @@ struct Trellis3 {
aux[i] = _mm256_and_si256(aux[i], mask);
}
auto offset = _mm256_set1_epi32(-126);
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
auto m1 = _mm256_set1_epi32(0x01010101);
#endif
for (int i = 0; i < 16; ++i) {
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
aux[i] = _mm256_dpbusd_epi32(offset, aux[i], m1);
#else
auto dot = _mm256_maddubs_epi16(aux[i], _mm256_set1_epi32(0x01010101));
aux[i] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot, _mm256_set1_epi16(1)));
#endif
}
for (int k = 0; k < 4; ++k) {
auto v1 = _mm256_packs_epi32(aux[4*k+0], aux[4*k+1]);
@@ -243,9 +252,6 @@ struct Trellis3 {
result[k] = _mm256_sign_epi8(result[k], result[k]);
}
}
#else
for (int k = 0; k < 4; ++k) result[k] = next32(val + 4*k, v0);
#endif
}
inline __m256i next32(const uint16_t * val, uint32_t v0) const {
const __m256i offset = _mm256_set1_epi32(-126);

6
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -264,9 +264,9 @@ struct MulMat {
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;
case GGML_TYPE_Q8_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ2_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ2_KT : return nrc_y >= 16 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ3_KT : return nrc_y >= 16 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_KT : return nrc_y >= 24 ? GGML_TYPE_Q8_0_R8 : type;
default: break;
}
#else