/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-p5-nr2fma-x24.c | 163 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local 170 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 177 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
D | neonfma-rr1-p5-nr1recps1fma-x24.c | 163 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local 170 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 177 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x24.c | 205 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local 212 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 219 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
|
D | neonfma-rr1-lut64-p2-nr2fma-x24.c | 213 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local 220 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 227 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
|
D | neonfma-rr1-p5-nr2recps-x24.c | 163 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local 170 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 177 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 185 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | neon-rr2-p5-nr2recps-x24.c | 173 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() local 180 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() 187 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24() 195 float32x4_t vfKLMN = vmulq_f32(veKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x24()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x24.c | 213 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() local 220 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() 227 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24() 235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c | 205 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24() local 212 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24() 219 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24() 227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x24()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x24.c | 205 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() local 212 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() 219 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24() 227 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x24()
|
D | neonfma-rr1-lut64-p2-nr2recps-x24.c | 213 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() local 220 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() 227 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24() 235 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x24()
|
D | neon-rr2-lut64-p2-nr2recps-x24.c | 223 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24() local 230 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24() 237 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24() 245 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x24()
|
D | neon-rr2-lut2048-p1-nr2recps-x24.c | 215 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24() local 222 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24() 229 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24() 237 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x24()
|