| /external/XNNPACK/src/f32-vscaleextexp/gen/ |
| D | avx2-p5-x56.c | 147 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56() local
159 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
170 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x56()
|
| D | avx2-p5-x64.c | 157 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x64() local
170 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x64()
182 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x72.c | 167 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x72() local
181 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x72()
194 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x72()
|
| D | avx2-p5-x80.c | 177 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x80() local
192 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x80()
206 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x80()
|
| D | avx2-p5-x88.c | 187 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88() local
203 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88()
218 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x88()
|
| D | avx2-p5-x96.c | 197 __m256 ve6 = _mm256_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96() local
214 ve6 = _mm256_max_ps(ve6, vmin_exponent); in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
230 …const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(ve6, vm… in xnn_f32_vscaleextexp_ukernel__avx2_p5_x96()
|
| D | avx512f-p5-scalef-x112.c | 141 const __m512 ve6 = _mm512_add_ps(vn6, vscalee); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112() local
150 vf6 = _mm512_scalef_ps(vf6, ve6); in xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x112()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x112.c | 128 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112() local
136 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
144 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112()
|
| D | avx512f-rr1-p5-scalef-div-x112.c | 125 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112() local
133 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
141 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c | 134 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112() local
142 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
150 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x112()
|
| D | avx512f-rr1-p5-scalef-div-x128.c | 135 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128() local
144 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
153 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x128()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c | 134 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112() local
142 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
166 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x112()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c | 128 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112() local
136 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
160 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x112()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x112.c | 125 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112() local
133 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
157 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c | 144 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128() local
153 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
162 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x128()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-div-x128.c | 138 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128() local
147 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
156 __m512 vf6 = _mm512_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x128()
|
| D | avx2-rr1-p5-div-x56.c | 136 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56() local
144 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56()
152 __m256 vf6 = _mm256_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x56()
|
| D | avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c | 138 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128() local
147 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
174 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128()
|
| D | avx2-rr1-p5-div-x64.c | 147 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64() local
156 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64()
165 __m256 vf6 = _mm256_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x64()
|
| D | avx512f-rr1-p5-scalef-nr1fma-x128.c | 135 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128() local
144 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
171 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128()
|
| D | avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c | 144 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128() local
153 const __m512 vd6 = _mm512_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
180 __m512 vf6 = _mm512_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_nr1fma_x128()
|
| D | avx2-rr1-p5-nr1fma-x56.c | 136 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56() local
144 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
169 __m256 vf6 = _mm256_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr1fma_x56()
|
| D | avx-rr2-p5-div-x56.c | 159 const __m256 ve6 = _mm256_add_ps(_mm256_mul_ps(vt6, vp6), vs6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56() local
167 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
175 __m256 vf6 = _mm256_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
|
| D | avx2-rr1-p5-nr2fma-x56.c | 136 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56() local
144 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
176 __m256 vf6 = _mm256_mul_ps(ve6, vr6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x56()
|
| D | avx2-rr1-p5-div-x72.c | 158 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72() local
168 const __m256 vd6 = _mm256_add_ps(ve6, vone); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72()
178 __m256 vf6 = _mm256_div_ps(ve6, vd6); in xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x72()
|