| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-scalar-rr2-lut16-p3-x6.c | 55 const float vz5 = vx5 * vprescale; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6() local
62 float vn5 = vz5 * vlog2e + vmagic_bias; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
93 float vt5 = vn5 * vminus_ln2_hi + vz5; in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
122 if XNN_UNPREDICTABLE(vz5 <= vsat_cutoff) { in xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x6()
|
| D | velu-scalar-rr2-p6-x6.c | 55 const float vz5 = vx5 * vprescale; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6() local
62 float vn5 = vz5 * vlog2e + vmagic_bias; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
82 float vt5 = vn5 * vminus_ln2_hi + vz5; in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
111 if XNN_UNPREDICTABLE(vz5 <= vsat_cutoff) { in xnn_f32_velu_ukernel__scalar_rr2_p6_x6()
|
| D | velu-wasm-rr2-lut16-p3-x6.c | 55 …const float vz5 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx5 * vprescale, vsat_cutoff), 0.0… in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6() local
62 float vn5 = vz5 * vlog2e + vmagic_bias; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6()
93 float vt5 = vn5 * vminus_ln2_hi + vz5; in xnn_f32_velu_ukernel__wasm_rr2_lut16_p3_x6()
|
| D | velu-wasm-rr2-p6-x6.c | 55 …const float vz5 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx5 * vprescale, vsat_cutoff), 0.0… in xnn_f32_velu_ukernel__wasm_rr2_p6_x6() local
62 float vn5 = vz5 * vlog2e + vmagic_bias; in xnn_f32_velu_ukernel__wasm_rr2_p6_x6()
82 float vt5 = vn5 * vminus_ln2_hi + vz5; in xnn_f32_velu_ukernel__wasm_rr2_p6_x6()
|
| /external/XNNPACK/src/f16-vsigmoid/gen/ |
| D | vsigmoid-avx2-rr1-p2-rcp-x48.c | 52 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x48() local
59 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x48()
80 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x48()
129 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x48()
|
| D | vsigmoid-avx2-rr1-p2-div-x48.c | 52 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48() local
59 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48()
80 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48()
122 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x48()
|
| D | vsigmoid-avx2-rr1-p2-div-x56.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56() local
61 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56()
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56()
133 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x56()
|
| D | vsigmoid-avx2-rr1-p2-rcp-x56.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x56() local
61 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x56()
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x56()
141 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x56()
|
| D | vsigmoid-avx2-rr1-p2-div-x64.c | 54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x64() local
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x64()
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x64()
144 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_div_x64()
|
| D | vsigmoid-avx2-rr1-p2-rcp-x64.c | 54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64() local
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64()
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64()
153 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x64()
|
| /external/XNNPACK/src/f32-vsigmoid/gen/ |
| D | vsigmoid-avx2-rr1-p5-div-x48.c | 52 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x48() local
59 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x48()
80 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x48()
143 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x48()
|
| D | vsigmoid-avx2-rr1-p5-nr1fma-x48.c | 52 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48() local
59 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48()
80 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48()
158 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x48()
|
| D | vsigmoid-avx2-rr1-p5-div-x56.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56() local
61 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56()
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56()
157 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x56()
|
| D | vsigmoid-avx512f-rr1-p5-scalef-div-x96.c | 51 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x96() local
58 __m512 vn5 = _mm512_mul_ps(vz5, vlog2e); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x96()
72 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x96()
|
| D | vsigmoid-avx512f-rr1-lut16-p3-perm-scalef-div-x96.c | 50 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96() local
57 __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
78 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96()
|
| D | vsigmoid-avx2-rr1-p5-nr1fma-x56.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x56() local
61 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
174 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
|
| D | vsigmoid-avx2-rr1-p5-nr2fma-x48.c | 52 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x48() local
59 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x48()
80 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x48()
164 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x48()
|
| D | vsigmoid-avx2-rr1-p5-div-x64.c | 54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64() local
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64()
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64()
171 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x64()
|
| D | vsigmoid-avx2-rr1-p5-div-x72.c | 55 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x72() local
65 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x72()
95 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x72()
185 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x72()
|
| D | vsigmoid-avx2-rr1-p5-nr1fma-x64.c | 54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x64() local
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
190 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x64()
|
| D | vsigmoid-avx-rr2-p5-div-x48.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48() local
60 __m256 vn5 = _mm256_add_ps(_mm256_mul_ps(vz5, vlog2e), vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48()
93 __m256 vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_hi), vz5); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48()
163 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x48()
|
| D | vsigmoid-avx2-rr1-p5-nr2fma-x56.c | 53 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x56() local
61 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
85 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
181 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
|
| D | vsigmoid-avx-rr2-p5-div-x56.c | 54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x56() local
62 __m256 vn5 = _mm256_add_ps(_mm256_mul_ps(vz5, vlog2e), vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x56()
100 __m256 vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_hi), vz5); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x56()
180 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5); in xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x56()
|
| D | vsigmoid-avx512f-rr1-p5-scalef-div-x112.c | 52 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112() local
60 __m512 vn5 = _mm512_mul_ps(vz5, vlog2e); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
76 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5); in xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
|
| D | vsigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c | 52 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask)); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96() local
59 __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
80 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vz5); in xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96()
|