/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-lut2048-p1-div-x16.c | 153 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local 159 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() 165 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
D | neonfma-rr1-lut64-p2-div-x16.c | 159 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local 165 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() 171 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x16.c | 153 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 159 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() 183 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 153 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 159 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 183 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
D | neon-rr2-lut64-p2-nr2recps-x16.c | 167 const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() local 173 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() 197 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16()
|
D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 159 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 165 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 189 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
D | neonfma-rr1-lut2048-p1-div-x20.c | 169 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local 176 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() 183 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 159 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 165 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 189 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
D | neonfma-rr1-lut64-p2-nr2recps-x16.c | 159 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local 165 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() 189 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
|
D | neon-rr2-lut2048-p1-nr2recps-x16.c | 161 const float32x4_t vyCDEF = vmlaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() local 167 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() 191 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x16.c | 153 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local 159 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 183 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
|
D | neonfma-rr1-lut64-p2-div-x20.c | 176 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local 183 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() 190 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 176 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local 183 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() 211 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
|
D | neonfma-rr1-lut64-p2-nr2fma-x20.c | 176 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 183 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 211 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x20.c | 169 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local 176 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() 204 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-div-x24.c | 193 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() local 201 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24() 209 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x24()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x20.c | 169 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local 176 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 204 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c | 169 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local 176 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() 204 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
|
D | neonfma-rr1-lut2048-p1-div-x24.c | 185 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() local 193 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24() 201 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x24()
|
D | neon-rr2-lut64-p2-nr2recps-x20.c | 185 const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20() local 192 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20() 220 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x20()
|
D | neon-rr2-lut2048-p1-nr2recps-x20.c | 178 const float32x4_t vyCDEF = vmlaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() local 185 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20() 213 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-nr2recps-x20.c | 176 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() local 183 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20() 211 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x24.c | 193 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() local 201 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() 233 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x24.c | 185 const float32x4_t vyCDEF = vfmaq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local 193 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 225 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
|
D | neonfma-rr1-lut64-p2-nr2fma-x24.c | 193 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local 201 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 233 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
|