| /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/ |
| H A D | velu-avx2-rr1-lut8-p4-perm-x56.c | 90 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x56() local
|
| H A D | velu-avx-rr2-lut4-p4-perm-x40.c | 104 __m256 vs2 = _mm256_mul_ps(vl2, ven2); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x40() local
|
| H A D | velu-avx512f-rr1-lut16-p3-perm-x112.c | 84 __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x112() local
|
| H A D | velu-avx2-rr1-p6-x80.c | 78 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/ |
| H A D | avx2-rr1-p2-x80-acc2.c | 77 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80_acc2() local
|
| H A D | avx2-rr1-p2-x80-acc5.c | 80 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80_acc5() local
|
| H A D | neonfp16arith-rr2-p2-x72.c | 73 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x72() local
|
| H A D | neonfp16arith-rr2-p2-x64-acc4.c | 73 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x64_acc4() local
|
| H A D | neonfp16arith-rr2-p2-x72-acc3.c | 75 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x72_acc3() local
|
| H A D | avx2-rr1-p2-x80.c | 76 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x80() local
|
| H A D | avx2-rr1-p2-x96.c | 82 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x96() local
|
| H A D | neonfp16arith-rr2-p2-x80.c | 76 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x80() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/ |
| H A D | vsigmoid-neonfp16arith-rr2-p2-div-x56.c | 64 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x56() local
|
| H A D | vsigmoid-avx2-rr1-p2-rcp-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_rcp_x56() local
|
| H A D | vsigmoid-neonfp16arith-rr2-p2-div-x64.c | 67 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x64() local
|
| H A D | vsigmoid-avx2-rr1-p2-rcp-x64.c | 69 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64() local
|
| H A D | vsigmoid-neonfp16arith-rr2-p2-nr1recps-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_nr1recps_x48() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-vscaleextexp/gen/ |
| H A D | avx2-p5-x80.c | 202 …const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve2, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x80() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/ |
| H A D | vsigmoid-avx-rr2-p5-div-x48.c | 70 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48() local
|
| H A D | vsigmoid-avx2-rr1-p5-div-x64.c | 69 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
| H A D | avx2-rr1-p5-x80-acc2.c | 77 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x80_acc2() local
|
| H A D | avx2-rr1-p5-x80-acc5.c | 80 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x80_acc5() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddextexp/gen/ |
| H A D | avx2-p5-x64-acc2.c | 179 …const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_… in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-raddexpminusmax/gen/ |
| H A D | avx2-p5-x80-acc2.c | 86 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x80_acc2() local
|
| H A D | avx2-p5-x96.c | 91 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96() local
|