/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse2-p5-div-x16.c | 155 … __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x16() local 160 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x16()
|
D | sse2-p5-div-x20.c | 171 … __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20() local 177 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
|
D | sse2-p5-div-x24.c | 187 … __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24() local 194 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
|
D | neonfma-rr1-p5-div-x16.c | 146 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16() local 151 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x16()
|
D | neonfma-rr1-p5-div-x20.c | 161 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() local 167 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
D | neonfma-rr1-p5-nr2recps-x16.c | 164 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() local 169 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16()
|
D | neon-rr2-p5-nr2recps-x16.c | 172 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local 177 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
D | neonfma-rr1-p5-nr1recps1fma-x16.c | 164 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 169 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | neonfma-rr1-p5-nr2fma-x16.c | 164 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 169 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | neonfma-rr1-p5-nr1recps1fma-x20.c | 182 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 188 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | neonfma-rr1-p5-nr2recps-x20.c | 182 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local 188 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
D | neonfma-rr1-lut2048-p1-div-x16.c | 178 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local 183 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
D | neonfma-rr1-p5-div-x24.c | 176 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() local 183 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
D | neonfma-rr1-lut64-p2-div-x16.c | 184 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local 189 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
D | neon-rr2-p5-nr2recps-x20.c | 191 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local 197 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
D | neonfma-rr1-p5-nr2fma-x20.c | 182 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 188 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x16.c | 196 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 201 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 196 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 201 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
D | neon-rr2-lut64-p2-nr2recps-x16.c | 210 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() local 215 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16()
|
D | neonfma-rr1-p5-nr1recps1fma-x24.c | 200 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local 207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 202 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
D | neonfma-rr1-p5-nr2recps-x24.c | 200 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local 207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | neonfma-rr1-lut2048-p1-div-x20.c | 198 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local 204 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 202 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
D | neonfma-rr1-lut64-p2-nr2recps-x16.c | 202 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local 207 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
|