/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-p5-nr2fma-x16.c | 135 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 140 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 145 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 151 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | neonfma-rr1-p5-nr1recps1fma-x16.c | 135 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 140 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 145 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 151 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | neonfma-rr1-p5-nr2fma-x20.c | 148 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 154 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 160 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 167 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | neonfma-rr1-p5-nr2recps-x16.c | 135 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() local 140 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 145 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 151 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16()
|
D | neonfma-rr1-p5-nr1recps1fma-x20.c | 148 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 154 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 160 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 167 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 173 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 178 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 183 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-nr2fma-x16.c | 167 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local 172 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 177 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-p5-nr2fma-x24.c | 161 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local 168 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 175 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 183 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
D | neon-rr2-p5-nr2recps-x16.c | 143 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local 148 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 153 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 159 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
D | neonfma-rr1-lut64-p2-nr2fma-x20.c | 192 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 198 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 204 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-nr2fma-x20.c | 185 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local 191 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 197 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-x16.c | 167 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 172 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 177 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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 | neonfma-rr1-p5-nr1recps1fma-x24.c | 161 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local 168 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 175 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 183 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 173 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 178 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 183 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-p5-nr2recps-x20.c | 148 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local 154 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 160 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 167 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
D | neon-rr2-p5-nr2recps-x20.c | 157 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local 163 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() 169 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20() 176 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x16.c | 167 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 172 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() 177 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); 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-nr2fma-x24.c | 203 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local 210 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 217 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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 | 211 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local 218 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 225 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 192 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local 198 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() 204 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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-lut2048-p1-nr1recps1fma-x20.c | 185 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local 191 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() 197 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); 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 | neon-rr2-lut64-p2-nr2recps-x16.c | 181 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() local 186 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() 191 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); 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-p5-nr2recps-x24.c | 161 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local 168 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 175 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 183 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | neonfma-rr1-lut64-p2-nr2recps-x16.c | 173 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local 178 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() 183 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); 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 | 175 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() local 180 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() 185 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); 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()
|