/external/XNNPACK/src/f32-sigmoid/gen/ |
D | sse2-p5-div-x20.c | 172 … __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20() local 178 vfGHIJ = _mm_or_ps(_mm_and_ps(vfGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, _mm_sub_ps(vone, vfGHIJ))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x20()
|
D | sse2-p5-div-x24.c | 188 … __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24() local 195 vfGHIJ = _mm_or_ps(_mm_and_ps(vfGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, _mm_sub_ps(vone, vfGHIJ))); in xnn_f32_sigmoid_ukernel__sse2_p5_div_x24()
|
D | neonfma-rr1-p5-div-x20.c | 162 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20() local 168 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x20()
|
D | neonfma-rr1-p5-nr1recps1fma-x20.c | 183 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 189 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | neonfma-rr1-p5-nr2recps-x20.c | 183 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local 189 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
D | neonfma-rr1-p5-div-x24.c | 177 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24() local 184 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_div_x24()
|
D | neon-rr2-p5-nr2recps-x20.c | 192 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local 198 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
D | neonfma-rr1-p5-nr2fma-x20.c | 183 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 189 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | neonfma-rr1-p5-nr1recps1fma-x24.c | 201 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local 208 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | neonfma-rr1-p5-nr2recps-x24.c | 201 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local 208 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | neonfma-rr1-lut2048-p1-div-x20.c | 199 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local 205 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
D | neonfma-rr1-p5-nr2fma-x24.c | 201 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local 208 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
D | neonfma-rr1-lut64-p2-div-x20.c | 206 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local 212 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
|
D | neon-rr2-p5-nr2recps-x24.c | 211 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local 218 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 227 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local 233 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
|
D | neonfma-rr1-lut64-p2-nr2fma-x20.c | 227 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 233 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x20.c | 220 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local 226 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-div-x24.c | 227 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local 234 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x20.c | 220 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local 226 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c | 220 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local 226 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
|
D | neonfma-rr1-lut2048-p1-div-x24.c | 219 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local 226 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
|
D | neon-rr2-lut64-p2-nr2recps-x20.c | 236 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20() local 242 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20()
|
D | neon-rr2-lut2048-p1-nr2recps-x20.c | 229 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() local 235 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-nr2recps-x20.c | 227 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() local 233 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x24.c | 251 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() local 258 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
|