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

ggml : IQ4_NL sgemm + Q4_0 AVX optimization (#9422)

Browse source 
* squashed

readd my iq4_nl sgemm PR https://github.com/ggerganov/llama.cpp/pull/8049

have ggml_vec_dot_q4_0 do two blocks per loop for avx

try out f16c ggml_vec_dot_iq4_nl, but it's not really faster. as per https://github.com/ggerganov/llama.cpp/pull/8549 we can calculate several blocks at a time with no issue

* shuffle

* remove f16c iq4_nl as i cant make it faster than before
This commit is contained in:
Eve 2024-09-16 06:48:24 +00:00 • committed by GitHub
parent 0aadac10c7
commit 5c3d0f1824
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 71 additions and 36 deletions
Download patch file Download diff file Expand all files Collapse all files

38
ggml/src/llamafile/sgemm.cpp
View file

@ -235,6 +235,14 @@ template <> inline __m512 load(const ggml_fp16_t *p) {
}
#endif // __AVX512F__
////////////////////////////////////////////////////////////////////////////////////////////////////
// CONSTANTS
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// FLOATING POINT MATRIX MULTIPLICATION
@ -933,6 +941,20 @@ class tinyBLAS_Q0_AVX {
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
}
inline __m256i load(const block_iq4_nl *b) {
return MM256_SET_M128I(load1(b), load0(b));
}
inline __m128i load0(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
}
inline __m128i load1(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
}
inline __m256 updot(__m256i u, __m256i s) {
__m256i res;
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
@ -1159,6 +1181,22 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
#endif
}
case GGML_TYPE_IQ4_NL: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
k, (const block_iq4_nl *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
default:
return false;
}