CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : workaround for missing _mm256_setr_m128i in GCC < 8 (#1638)

Browse Source
This commit is contained in:
Xingchen Song(宋星辰)
2023-06-10 15:49:40 +08:00
committed by GitHub
parent 555275a693
commit ef3171d162
1 changed files with 9 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
ggml.c
View File

@@ -492,6 +492,8 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
// quantization // quantization
// //
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
// multiply int8_t, add results pairwise twice // multiply int8_t, add results pairwise twice
static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
@@ -551,7 +553,7 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) {
static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
{ {
const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
const __m256i bytes = _mm256_set_m128i(_mm_srli_epi16(tmp, 4), tmp); const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
const __m256i lowMask = _mm256_set1_epi8( 0xF ); const __m256i lowMask = _mm256_set1_epi8( 0xF );
return _mm256_and_si256(lowMask, bytes); return _mm256_and_si256(lowMask, bytes);
} }
@@ -624,7 +626,7 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) {
bytesh = _mm_or_si128(bytesh, bit_mask); bytesh = _mm_or_si128(bytesh, bit_mask);
bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1)); bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1)); bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
return _mm256_set_m128i(bytesh, bytesl); return MM256_SET_M128I(bytesh, bytesl);
} }
// Unpack 32 4-bit fields into 32 bytes // Unpack 32 4-bit fields into 32 bytes
@@ -637,7 +639,7 @@ static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
const __m128i lowMask = _mm_set1_epi8(0xF); const __m128i lowMask = _mm_set1_epi8(0xF);
tmpl = _mm_and_si128(lowMask, tmpl); tmpl = _mm_and_si128(lowMask, tmpl);
tmph = _mm_and_si128(lowMask, tmph); tmph = _mm_and_si128(lowMask, tmph);
return _mm256_set_m128i(tmph, tmpl); return MM256_SET_M128I(tmph, tmpl);
} }
// add int16_t pairwise and return as float vector // add int16_t pairwise and return as float vector
@@ -645,7 +647,7 @@ static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
const __m128i ones = _mm_set1_epi16(1); const __m128i ones = _mm_set1_epi16(1);
const __m128i summed_pairsl = _mm_madd_epi16(ones, xl); const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
const __m128i summed_pairsh = _mm_madd_epi16(ones, xh); const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
const __m256i summed_pairs = _mm256_set_m128i(summed_pairsh, summed_pairsl); const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
return _mm256_cvtepi32_ps(summed_pairs); return _mm256_cvtepi32_ps(summed_pairs);
} }
@@ -2350,7 +2352,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
const __m128i i32_1 = mul_sum_i8_pairs(bx, by); const __m128i i32_1 = mul_sum_i8_pairs(bx, by);
// Convert int32_t to float // Convert int32_t to float
__m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1)); __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
// Apply the scale, and accumulate // Apply the scale, and accumulate
acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
@@ -2826,7 +2828,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
__m128i bxh = _mm256_extractf128_si256(bx, 1); __m128i bxh = _mm256_extractf128_si256(bx, 1);
bxl = _mm_or_si128(bxl, bxhil); bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih); bxh = _mm_or_si128(bxh, bxhih);
bx = _mm256_set_m128i(bxh, bxl); bx = MM256_SET_M128I(bxh, bxl);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
@@ -3082,7 +3084,7 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
__m128i bxh = _mm256_extractf128_si256(bx, 1); __m128i bxh = _mm256_extractf128_si256(bx, 1);
bxl = _mm_or_si128(bxl, bxhil); bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih); bxh = _mm_or_si128(bxh, bxhih);
bx = _mm256_set_m128i(bxh, bxl); bx = MM256_SET_M128I(bxh, bxl);
const __m256 dy = _mm256_set1_ps(y[i].d); const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);