Skip to content

Commit bd83694

Browse files
cebtenzzrecebtenzzre
authored
quants : use MM256_SET_M128I consistently to fix gcc 7 build (ggml-org#5889)
1 parent 3de3167 commit bd83694

File tree

1 file changed

+14
-13
lines changed

1 file changed

+14
-13
lines changed

‎ggml-quants.c‎

Lines changed: 14 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -51,6 +51,7 @@
5151

5252
#define UNUSED GGML_UNUSED
5353

54+
// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
5455
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
5556

5657
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
@@ -9563,7 +9564,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
95639564

95649565
const __m128i odd_bits = _mm_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
95659566
const __m128i full_sign_bits = _mm_or_si128(partial_sign_bits, odd_bits);
9566-
const __m256i full_signs = _mm256_set_m128i(full_sign_bits, full_sign_bits);
9567+
const __m256i full_signs = MM256_SET_M128I(full_sign_bits, full_sign_bits);
95679568

95689569
const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)y[i].qs);
95699570
const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)(y[i].qs+32));
@@ -9585,8 +9586,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
95859586
const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1);
95869587
const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2);
95879588

9588-
const __m256i sc1 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1));
9589-
const __m256i sc2 = _mm256_set_m128i(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1));
9589+
const __m256i sc1 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[0] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[0] & 0xf)+1));
9590+
const __m256i sc2 = MM256_SET_M128I(_mm_set1_epi16(2*(x[i].scales[1] >> 4)+1), _mm_set1_epi16(2*(x[i].scales[1] & 0xf)+1));
95909591

95919592
const __m256i sum = _mm256_add_epi32(_mm256_madd_epi16(sc1, dot1), _mm256_madd_epi16(sc2, dot2));
95929593

@@ -9653,8 +9654,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void *
96539654

96549655
const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
96559656
const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
9656-
const __m256i full_signs_1 = _mm256_set_m128i(full_signs_l, full_signs_l);
9657-
const __m256i full_signs_2 = _mm256_set_m128i(full_signs_h, full_signs_h);
9657+
const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
9658+
const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
96589659

96599660
__m256i signs;
96609661
signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
@@ -10551,10 +10552,10 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void *
1055110552
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[1].qs);
1055210553
const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[0].qs);
1055310554
const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[1].qs);
10554-
const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
10555-
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
10556-
const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
10557-
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
10555+
const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
10556+
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
10557+
const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
10558+
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
1055810559
const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
1055910560
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
1056010561
const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
@@ -10661,10 +10662,10 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
1066110662
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs); qs += 16;
1066210663
const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
1066310664
const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
10664-
const __m256i q4b_1 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
10665-
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
10666-
const __m256i q4b_2 = _mm256_set_m128i(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
10667-
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
10665+
const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
10666+
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
10667+
const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
10668+
_mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
1066810669
const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
1066910670
const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
1067010671
const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;

0 commit comments

Comments
 (0)