| /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/ |
| H A D | vsigmoid-avx2-rr1-p5-nr2fma-x40.c | 60 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x40() local
|
| H A D | vsigmoid-avx-rr2-p5-div-x40.c | 67 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x40() local
|
| H A D | vsigmoid-avx2-rr1-p5-div-x56.c | 66 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56() local
|
| H A D | vsigmoid-avx2-rr1-p5-nr1fma-x48.c | 63 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| H A D | avx2-p5-x64-acc2.c | 80 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x64_acc2() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/ |
| H A D | velu-avx2-rr1-p6-x56.c | 69 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x56() local
|
| H A D | velu-avx2-rr1-lut16-p3-gather-x40.c | 85 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x40() local
|
| H A D | velu-avx-rr2-lut4-p4-perm-x32.c | 94 __m256 vs2 = _mm256_mul_ps(vl2, ven2); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x32() local
|
| H A D | velu-avx2-rr1-p6-x64.c | 72 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x64() local
|
| H A D | velu-avx2-rr1-lut4-p4-perm-x48.c | 84 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x48() local
|
| H A D | velu-avx512f-rr1-p6-x128.c | 74 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x128() local
|
| H A D | velu-avx2-rr1-lut8-p4-perm-x48.c | 84 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x48() local
|
| H A D | velu-avx-rr2-p6-x48.c | 88 __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_p6_x48() local
|
| H A D | velu-avx512f-rr1-lut16-p3-perm-x96.c | 79 __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x96() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/ |
| H A D | avx2-rr1-p2-x64-acc2.c | 71 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x64_acc2() local
|
| H A D | neonfp16arith-rr2-p2-x64-acc2.c | 71 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x64_acc2() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| H A D | avx2-rr1-p5-x64-acc2.c | 71 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x64_acc2() local
|
| H A D | avx2-rr1-p5-x64-acc4.c | 73 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x64_acc4() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/ |
| H A D | vsigmoid-avx2-rr1-p2-div-x48.c | 63 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48() local
|
| H A D | vsigmoid-avx2-rr1-p2-div-x56.c | 66 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56() local
|
| H A D | vsigmoid-neonfp16arith-rr2-p2-div-x48.c | 61 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x48() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddextexp/gen/ |
| H A D | avx2-p5-x64.c | 175 …const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_… in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local
|
| H A D | avx2-p5-x72.c | 186 …const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_… in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/ |
| H A D | avx2-p5-x72.c | 190 …const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x72() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| H A D | avx2-p5-x72.c | 84 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x72() local
|