| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx2-rr1-p5-div-x80.c | 62 const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80() local
73 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
106 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
205 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x80()
|
| D | avx2-rr1-p5-nr1fma-x80.c | 62 const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80() local
73 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
106 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
228 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x80()
|
| D | avx2-rr1-p5-nr2fma-x80.c | 62 const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80() local
73 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80()
106 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80()
238 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80()
|
| D | avx-rr2-p5-div-x80.c | 63 const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80() local
74 __m256 vn9 = _mm256_add_ps(_mm256_mul_ps(vz9, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
127 __m256 vt9 = _mm256_add_ps(_mm256_mul_ps(vn9, vminus_ln2_hi), vz9); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
237 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
|
| D | avx-rr2-p5-nr2-x80.c | 64 const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x80() local
75 __m256 vn9 = _mm256_add_ps(_mm256_mul_ps(vz9, vlog2e), vmagic_bias); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x80()
128 __m256 vt9 = _mm256_add_ps(_mm256_mul_ps(vn9, vminus_ln2_hi), vz9); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x80()
270 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x80()
|
| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx2-rr1-lut4-p4-perm-x80.c | 65 const __m256 vz9 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx9, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80() local
76 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80()
128 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x80()
|
| D | velu-avx2-rr1-lut16-p3-gather-x80.c | 64 const __m256 vz9 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx9, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80() local
75 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80()
138 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_velu_ukernel__avx2_rr1_lut16_p3_gather_x80()
|
| D | velu-avx2-rr1-p6-x80.c | 64 const __m256 vz9 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx9, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80() local
75 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
107 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_velu_ukernel__avx2_rr1_p6_x80()
|
| D | velu-avx2-rr1-lut8-p4-perm-x80.c | 64 const __m256 vz9 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx9, vprescale)); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80() local
75 __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80()
127 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9); in xnn_f32_velu_ukernel__avx2_rr1_lut8_p4_perm_x80()
|