/dports/biology/ncbi-cxx-toolkit/ncbi_cxx--25_2_0/include/util/bitset/ |
H A D | bmavx2.h | 173 b = _mm256_load_si256(block+0); c = _mm256_load_si256(block+1); in avx2_bit_count() 176 b = _mm256_load_si256(block+2); c = _mm256_load_si256(block+3); in avx2_bit_count() 180 b = _mm256_load_si256(block+4); c = _mm256_load_si256(block+5); in avx2_bit_count() 183 b = _mm256_load_si256(block+6); c = _mm256_load_si256(block+7); in avx2_bit_count() 188 b = _mm256_load_si256(block+8); c = _mm256_load_si256(block+9); in avx2_bit_count() 191 b = _mm256_load_si256(block+10); c = _mm256_load_si256(block+11); in avx2_bit_count() 195 b = _mm256_load_si256(block+12); c = _mm256_load_si256(block+13); in avx2_bit_count() 198 b = _mm256_load_si256(block+14); c = _mm256_load_si256(block+15); in avx2_bit_count() 640 dst0 = _mm256_load_si256(dst + 0); dst1 = _mm256_load_si256(dst + 1); in avx2_and_digest_5way() 647 dst0 = _mm256_load_si256(dst + 2); dst1 = _mm256_load_si256(dst + 3); in avx2_and_digest_5way() [all …]
|
H A D | bmavx512.h | 182 b = _mm256_load_si256(block+0); c = _mm256_load_si256(block+1); in avx2_bit_count() 185 b = _mm256_load_si256(block+2); c = _mm256_load_si256(block+3); in avx2_bit_count() 189 b = _mm256_load_si256(block+4); c = _mm256_load_si256(block+5); in avx2_bit_count() 192 b = _mm256_load_si256(block+6); c = _mm256_load_si256(block+7); in avx2_bit_count() 197 b = _mm256_load_si256(block+8); c = _mm256_load_si256(block+9); in avx2_bit_count() 200 b = _mm256_load_si256(block+10); c = _mm256_load_si256(block+11); in avx2_bit_count() 204 b = _mm256_load_si256(block+12); c = _mm256_load_si256(block+13); in avx2_bit_count() 207 b = _mm256_load_si256(block+14); c = _mm256_load_si256(block+15); in avx2_bit_count() 251 ymm0 = _mm256_load_si256(block); in avx2_bit_count_and() 258 ymm0 = _mm256_load_si256(block); in avx2_bit_count_and() [all …]
|
/dports/biology/ncbi-blast+/ncbi-blast-2.12.0+-src/c++/include/util/bitset/ |
H A D | bmavx2.h | 173 b = _mm256_load_si256(block+0); c = _mm256_load_si256(block+1); in avx2_bit_count() 176 b = _mm256_load_si256(block+2); c = _mm256_load_si256(block+3); in avx2_bit_count() 180 b = _mm256_load_si256(block+4); c = _mm256_load_si256(block+5); in avx2_bit_count() 183 b = _mm256_load_si256(block+6); c = _mm256_load_si256(block+7); in avx2_bit_count() 188 b = _mm256_load_si256(block+8); c = _mm256_load_si256(block+9); in avx2_bit_count() 191 b = _mm256_load_si256(block+10); c = _mm256_load_si256(block+11); in avx2_bit_count() 195 b = _mm256_load_si256(block+12); c = _mm256_load_si256(block+13); in avx2_bit_count() 198 b = _mm256_load_si256(block+14); c = _mm256_load_si256(block+15); in avx2_bit_count() 640 dst0 = _mm256_load_si256(dst + 0); dst1 = _mm256_load_si256(dst + 1); in avx2_and_digest_5way() 647 dst0 = _mm256_load_si256(dst + 2); dst1 = _mm256_load_si256(dst + 3); in avx2_and_digest_5way() [all …]
|
H A D | bmavx512.h | 182 b = _mm256_load_si256(block+0); c = _mm256_load_si256(block+1); in avx2_bit_count() 185 b = _mm256_load_si256(block+2); c = _mm256_load_si256(block+3); in avx2_bit_count() 189 b = _mm256_load_si256(block+4); c = _mm256_load_si256(block+5); in avx2_bit_count() 192 b = _mm256_load_si256(block+6); c = _mm256_load_si256(block+7); in avx2_bit_count() 197 b = _mm256_load_si256(block+8); c = _mm256_load_si256(block+9); in avx2_bit_count() 200 b = _mm256_load_si256(block+10); c = _mm256_load_si256(block+11); in avx2_bit_count() 204 b = _mm256_load_si256(block+12); c = _mm256_load_si256(block+13); in avx2_bit_count() 207 b = _mm256_load_si256(block+14); c = _mm256_load_si256(block+15); in avx2_bit_count() 251 ymm0 = _mm256_load_si256(block); in avx2_bit_count_and() 258 ymm0 = _mm256_load_si256(block); in avx2_bit_count_and() [all …]
|
/dports/graphics/openjph/OpenJPH-0.7.3/src/core/transform/ |
H A D | ojph_transform_avx2.cpp | 60 __m256i s1 = _mm256_load_si256((__m256i*)src1); in avx2_rev_vert_wvlt_fwd_predict() 61 __m256i s2 = _mm256_load_si256((__m256i*)src2); in avx2_rev_vert_wvlt_fwd_predict() 62 __m256i d = _mm256_load_si256((__m256i*)dst); in avx2_rev_vert_wvlt_fwd_predict() 76 __m256i s1 = _mm256_load_si256((__m256i*)src1); in avx2_rev_vert_wvlt_fwd_update() 78 __m256i s2 = _mm256_load_si256((__m256i*)src2); in avx2_rev_vert_wvlt_fwd_update() 80 __m256i d = _mm256_load_si256((__m256i*)dst); in avx2_rev_vert_wvlt_fwd_update() 161 __m256i s1 = _mm256_load_si256((__m256i*)src1); in avx2_rev_vert_wvlt_bwd_predict() 162 __m256i s2 = _mm256_load_si256((__m256i*)src2); in avx2_rev_vert_wvlt_bwd_predict() 163 __m256i d = _mm256_load_si256((__m256i*)dst); in avx2_rev_vert_wvlt_bwd_predict() 177 __m256i s1 = _mm256_load_si256((__m256i*)src1); in avx2_rev_vert_wvlt_bwd_update() [all …]
|
H A D | ojph_colour_avx2.cpp | 72 __m256i mr = _mm256_load_si256((__m256i*)r); in avx2_rct_forward() 73 __m256i mg = _mm256_load_si256((__m256i*)g); in avx2_rct_forward() 74 __m256i mb = _mm256_load_si256((__m256i*)b); in avx2_rct_forward() 94 __m256i my = _mm256_load_si256((__m256i*)y); in avx2_rct_backward() 95 __m256i mcb = _mm256_load_si256((__m256i*)cb); in avx2_rct_backward() 96 __m256i mcr = _mm256_load_si256((__m256i*)cr); in avx2_rct_backward()
|
/dports/multimedia/kvazaar/kvazaar-2.1.0/src/strategies/avx2/ |
H A D | dct-avx2.c | 187 _mm256_load_si256((const __m256i *)left + 0), in mul_clip_matrix_8x8_avx2() 188 _mm256_load_si256((const __m256i *)left + 1), in mul_clip_matrix_8x8_avx2() 189 _mm256_load_si256((const __m256i *)left + 2), in mul_clip_matrix_8x8_avx2() 190 _mm256_load_si256((const __m256i *)left + 3), in mul_clip_matrix_8x8_avx2() 193 _mm256_load_si256((const __m256i *)right + 0), in mul_clip_matrix_8x8_avx2() 194 _mm256_load_si256((const __m256i *)right + 1), in mul_clip_matrix_8x8_avx2() 195 _mm256_load_si256((const __m256i *)right + 2), in mul_clip_matrix_8x8_avx2() 196 _mm256_load_si256((const __m256i *)right + 3), in mul_clip_matrix_8x8_avx2() 691 _mm256_load_si256((const __m256i *)tdct + 0), in partial_butterfly_inverse_16_avx2() 692 _mm256_load_si256((const __m256i *)tdct + 1), in partial_butterfly_inverse_16_avx2() [all …]
|
/dports/biology/gmap/gmap-2020-09-12/src/ |
H A D | merge-uint8.c | 328 vOld256_0 = _mm256_load_si256((__m256i *) B); B += 4; in Merge_uint8() 329 vOld256_1 = _mm256_load_si256((__m256i *) B); B += 4; in Merge_uint8() 330 vNew256_0 = _mm256_load_si256((__m256i *) A); A += 4; in Merge_uint8() 345 vNew256_0 = _mm256_load_si256((__m256i *) A); A += 4; in Merge_uint8() 346 vNew256_1 = _mm256_load_si256((__m256i *) A); A += 4; in Merge_uint8() 349 vNew256_0 = _mm256_load_si256((__m256i *) B); B += 4; in Merge_uint8() 350 vNew256_1 = _mm256_load_si256((__m256i *) B); B += 4; in Merge_uint8() 374 vOld256 = _mm256_load_si256((__m256i *) B); B += 4; in Merge_uint8() 375 vNew256 = _mm256_load_si256((__m256i *) A); A += 4; in Merge_uint8() 377 vOld256 = _mm256_load_si256((__m256i *) A); A += 4; in Merge_uint8() [all …]
|
/dports/net/waypipe/waypipe-a11a59c58bee3972ce9e1e18731eba28cc0a8f0a/src/ |
H A D | kernel_avx2.c | 58 __m256i m0 = _mm256_load_si256(&mod[2 * i]); 59 __m256i m1 = _mm256_load_si256(&mod[2 * i + 1]); 60 __m256i b0 = _mm256_load_si256(&base[2 * i]); 61 __m256i b1 = _mm256_load_si256(&base[2 * i + 1]); 130 __m256i m0 = _mm256_load_si256(&mod[2 * i]); 131 __m256i m1 = _mm256_load_si256(&mod[2 * i + 1]); 132 __m256i b0 = _mm256_load_si256(&base[2 * i]); 133 __m256i b1 = _mm256_load_si256(&base[2 * i + 1]);
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber1024_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); in poly_compress() 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 578 f0 = _mm256_load_si256(&a->vec[i]); in poly_add() 579 f1 = _mm256_load_si256(&b->vec[i]); in poly_add() 601 f0 = _mm256_load_si256(&a->vec[i]); in poly_sub() [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber512-90s_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); in poly_compress() 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 578 f0 = _mm256_load_si256(&a->vec[i]); in poly_add() 579 f1 = _mm256_load_si256(&b->vec[i]); in poly_add() 601 f0 = _mm256_load_si256(&a->vec[i]); in poly_sub() [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber768_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); in poly_compress() 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 578 f0 = _mm256_load_si256(&a->vec[i]); in poly_add() 579 f1 = _mm256_load_si256(&b->vec[i]); in poly_add() 601 f0 = _mm256_load_si256(&a->vec[i]); in poly_sub() [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber1024-90s_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); in poly_compress() 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 578 f0 = _mm256_load_si256(&a->vec[i]); in poly_add() 579 f1 = _mm256_load_si256(&b->vec[i]); in poly_add() 601 f0 = _mm256_load_si256(&a->vec[i]); in poly_sub() [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber512_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); in poly_compress() 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); in poly_compress() 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); in poly_compress() 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); in poly_compress() 578 f0 = _mm256_load_si256(&a->vec[i]); in poly_add() 579 f1 = _mm256_load_si256(&b->vec[i]); in poly_add() 601 f0 = _mm256_load_si256(&a->vec[i]); in poly_sub() [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/kem/kyber/pqcrystals-kyber_kyber768-90s_avx2/ |
H A D | poly.c | 42 f0 = _mm256_load_si256(&a->vec[4*i+0]); 43 f1 = _mm256_load_si256(&a->vec[4*i+1]); 44 f2 = _mm256_load_si256(&a->vec[4*i+2]); 45 f3 = _mm256_load_si256(&a->vec[4*i+3]); 125 f0 = _mm256_load_si256(&a->vec[4*i+0]); 126 f1 = _mm256_load_si256(&a->vec[4*i+1]); 127 f2 = _mm256_load_si256(&a->vec[4*i+2]); 578 f0 = _mm256_load_si256(&a->vec[i]); 579 f1 = _mm256_load_si256(&b->vec[i]); 601 f0 = _mm256_load_si256(&a->vec[i]); [all …]
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium_5_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); in power2round_avx() 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 72 f = _mm256_load_si256(&a[i]); in decompose_avx() 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 103 f = _mm256_load_si256(&a[i]); in decompose_avx() 143 f0 = _mm256_load_si256(&a0[i]); in make_hint_avx() 144 f1 = _mm256_load_si256(&a1[i]); in make_hint_avx() 184 f = _mm256_load_si256(&a0[i]); in use_hint_avx() 185 g = _mm256_load_si256(&b[i]); in use_hint_avx() 186 h = _mm256_load_si256(&hint[i]); in use_hint_avx()
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium3-aes_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); in power2round_avx() 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 72 f = _mm256_load_si256(&a[i]); in decompose_avx() 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 103 f = _mm256_load_si256(&a[i]); in decompose_avx() 143 f0 = _mm256_load_si256(&a0[i]); in make_hint_avx() 144 f1 = _mm256_load_si256(&a1[i]); in make_hint_avx() 184 f = _mm256_load_si256(&a0[i]); in use_hint_avx() 185 g = _mm256_load_si256(&b[i]); in use_hint_avx() 186 h = _mm256_load_si256(&hint[i]); in use_hint_avx()
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium3_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); 72 f = _mm256_load_si256(&a[i]); 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); 103 f = _mm256_load_si256(&a[i]); 143 f0 = _mm256_load_si256(&a0[i]); 144 f1 = _mm256_load_si256(&a1[i]); 184 f = _mm256_load_si256(&a0[i]); 185 g = _mm256_load_si256(&b[i]); 186 h = _mm256_load_si256(&hint[i]);
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium_5-aes_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); in power2round_avx() 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 72 f = _mm256_load_si256(&a[i]); in decompose_avx() 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 103 f = _mm256_load_si256(&a[i]); in decompose_avx() 143 f0 = _mm256_load_si256(&a0[i]); in make_hint_avx() 144 f1 = _mm256_load_si256(&a1[i]); in make_hint_avx() 184 f = _mm256_load_si256(&a0[i]); in use_hint_avx() 185 g = _mm256_load_si256(&b[i]); in use_hint_avx() 186 h = _mm256_load_si256(&hint[i]); in use_hint_avx()
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium2-aes_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); in power2round_avx() 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 72 f = _mm256_load_si256(&a[i]); in decompose_avx() 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 103 f = _mm256_load_si256(&a[i]); in decompose_avx() 143 f0 = _mm256_load_si256(&a0[i]); in make_hint_avx() 144 f1 = _mm256_load_si256(&a1[i]); in make_hint_avx() 184 f = _mm256_load_si256(&a0[i]); in use_hint_avx() 185 g = _mm256_load_si256(&b[i]); in use_hint_avx() 186 h = _mm256_load_si256(&hint[i]); in use_hint_avx()
|
/dports/security/liboqs/liboqs-0.7.0/src/sig/dilithium/pqcrystals-dilithium_dilithium2_avx2/ |
H A D | rounding.c | 34 f = _mm256_load_si256(&a[i]); in power2round_avx() 63 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 72 f = _mm256_load_si256(&a[i]); in decompose_avx() 93 const __m256i q = _mm256_load_si256(&qdata.vec[_8XQ/8]); in decompose_avx() 103 f = _mm256_load_si256(&a[i]); in decompose_avx() 143 f0 = _mm256_load_si256(&a0[i]); in make_hint_avx() 144 f1 = _mm256_load_si256(&a1[i]); in make_hint_avx() 184 f = _mm256_load_si256(&a0[i]); in use_hint_avx() 185 g = _mm256_load_si256(&b[i]); in use_hint_avx() 186 h = _mm256_load_si256(&hint[i]); in use_hint_avx()
|
/dports/multimedia/librav1e/rav1e-0.5.1/src/asm/x86/dist/ |
H A D | hbd.rs | 74 let row_src = _mm256_load_si256(input1.as_ptr().offset(y * sizei) as *const __m256i); 75 let row_dst = _mm256_load_si256(input2.as_ptr().offset(y * sizei) as *const __m256i); 85 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 86 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 93 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 94 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 95 let row3 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(16) as *const __m256i)); 96 let row4 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(24) as *const __m256i)); 97 let row5 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(32) as *const __m256i)); 98 let row6 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(40) as *const __m256i)); [all …]
|
/dports/multimedia/rav1e/rav1e-0.5.1/src/asm/x86/dist/ |
H A D | hbd.rs | 74 let row_src = _mm256_load_si256(input1.as_ptr().offset(y * sizei) as *const __m256i); 75 let row_dst = _mm256_load_si256(input2.as_ptr().offset(y * sizei) as *const __m256i); 85 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 86 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 93 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 94 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 95 let row3 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(16) as *const __m256i)); 96 let row4 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(24) as *const __m256i)); 97 let row5 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(32) as *const __m256i)); 98 let row6 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(40) as *const __m256i)); [all …]
|
/dports/multimedia/gstreamer1-plugins-rust/gst-plugins-rs-d0466b3eee114207f851b37cae0015c0e718f021/cargo-crates/rav1e-0.4.1/src/asm/x86/dist/ |
H A D | hbd.rs | 74 let row_src = _mm256_load_si256(input1.as_ptr().offset(y * sizei) as *const __m256i); 75 let row_dst = _mm256_load_si256(input2.as_ptr().offset(y * sizei) as *const __m256i); 85 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 86 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 93 let row1 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(0) as *const __m256i)); 94 let row2 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(8) as *const __m256i)); 95 let row3 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(16) as *const __m256i)); 96 let row4 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(24) as *const __m256i)); 97 let row5 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(32) as *const __m256i)); 98 let row6 = _mm256_abs_epi32(_mm256_load_si256(buf.as_ptr().add(40) as *const __m256i)); [all …]
|
/dports/devel/folly/folly-2021.12.27.00/folly/experimental/crypto/detail/ |
H A D | MathOperation_AVX2.cpp | 76 __m256i v1 = _mm256_load_si256(v1p + i); in add() 77 __m256i v2 = _mm256_load_si256(v2p + i); in add() 92 __m256i v1 = _mm256_load_si256(v1p + i); in add() 93 __m256i v2 = _mm256_load_si256(v2p + i); in add() 133 __m256i v1 = _mm256_load_si256(v1p + i); in sub() 134 __m256i v2 = _mm256_load_si256(v2p + i); in sub() 150 __m256i v1 = _mm256_load_si256(v1p + i); in sub() 151 __m256i v2 = _mm256_load_si256(v2p + i); in sub() 184 results[i] = _mm256_and_si256(_mm256_load_si256(p + i), mask); in clearPaddingBits() 201 _mm256_load_si256(reinterpret_cast<const __m256i*>(buf.data() + pos)); in checkPaddingBits()
|