| /aosp_15_r20/external/XNNPACK/src/f32-vscaleexpminusmax/gen/ |
| H A D | avx2-p5-x48.c | 75 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x48() local
|
| H A D | avx2-p5-x56.c | 78 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_x56() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/ |
| H A D | vsigmoid-avx2-rr1-p5-nr1fma-x32.c | 57 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x32() local
|
| H A D | vsigmoid-avx2-rr1-p5-nr2fma-x32.c | 57 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x32() local
|
| H A D | vsigmoid-avx2-rr1-p5-div-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_div_x40() local
|
| H A D | vsigmoid-avx-rr2-p5-nr2-x24.c | 62 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x24() local
|
| H A D | vsigmoid-avx-rr2-p5-div-x32.c | 64 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x32() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-vsigmoid/gen/ |
| H A D | vsigmoid-avx2-rr1-p2-rcp-x32.c | 57 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x32() local
|
| H A D | vsigmoid-neonfp16arith-rr2-p2-nr1fma-x24.c | 52 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x24() local
|
| H A D | vsigmoid-neonfp16arith-rr2-p2-div-x32.c | 55 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_div_x32() local
|
| H A D | vsigmoid-avx2-rr1-p2-div-x40.c | 60 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x40() local
|
| /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/ |
| H A D | velu-wasm-rr2-lut16-p3-x5.c | 82 float vs2 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx2] + ven2); in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x5() local
|
| H A D | velu-avx2-rr1-lut4-p4-perm-x32.c | 72 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x32() local
|
| H A D | velu-scalar-rr2-lut16-p3-x5.c | 82 float vs2 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx2] + ven2); in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x5() local
|
| H A D | velu-avx-rr2-p6-x32.c | 76 __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1); in xnn_f32_velu_ukernel__avx_rr2_p6_x32() local
|
| H A D | velu-avx512f-rr1-p6-x96.c | 68 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23)); in xnn_f32_velu_ukernel__avx512f_rr1_p6_x96() local
|
| H A D | velu-avx2-rr1-p6-x48.c | 66 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x48() local
|
| H A D | velu-avx-rr2-lut4-p4-perm-x24.c | 84 __m256 vs2 = _mm256_mul_ps(vl2, ven2); in xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x24() local
|
| H A D | velu-avx2-rr1-lut16-p3-gather-x32.c | 78 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x32() local
|
| /aosp_15_r20/external/XNNPACK/src/f16-raddstoreexpminusmax/gen/ |
| H A D | neonfp16arith-rr2-p2-x40.c | 61 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40() local
|
| H A D | avx2-rr1-p2-x48-acc2.c | 65 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc2() local
|
| H A D | avx2-rr1-p2-x48.c | 64 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48() local
|
| H A D | neonfp16arith-rr2-p2-x40-acc2.c | 62 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40_acc2() local
|
| H A D | avx2-rr1-p2-x48-acc3.c | 66 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); in xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x48_acc3() local
|
| H A D | neonfp16arith-rr2-p2-x40-acc5.c | 65 const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10)); in xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40_acc5() local
|