/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-lut2048-p1-div-x12.c | 136 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12() local 141 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12() 146 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x12()
|
D | neonfma-rr1-lut64-p2-div-x12.c | 141 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12() local 146 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12() 151 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x12()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c | 136 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() local 141 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() 161 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x12.c | 136 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() local 141 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() 161 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x12.c | 136 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() local 141 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() 161 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12()
|
D | neonfma-rr1-lut2048-p1-div-x16.c | 152 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() local 158 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16() 164 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x16()
|
D | neonfma-rr1-lut64-p2-nr2fma-x12.c | 141 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() local 146 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() 166 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x12.c | 141 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() local 146 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() 166 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
|
D | neonfma-rr1-lut64-p2-div-x16.c | 158 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() local 164 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16() 170 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x16()
|
D | neon-rr2-lut2048-p1-nr2recps-x12.c | 143 const float32x4_t vy89AB = vmlaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12() local 148 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12() 168 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12()
|
D | neonfma-rr1-lut64-p2-nr2recps-x12.c | 141 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() local 146 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() 166 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
|
D | neon-rr2-lut64-p2-nr2recps-x12.c | 148 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12() local 153 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12() 173 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x16.c | 152 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 158 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() 182 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 152 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 158 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 182 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
D | neon-rr2-lut64-p2-nr2recps-x16.c | 166 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() local 172 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() 196 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16()
|
D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 158 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 164 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 188 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
D | neonfma-rr1-lut2048-p1-div-x20.c | 168 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() local 175 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20() 182 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_div_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 158 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 164 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 188 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
D | neonfma-rr1-lut64-p2-nr2recps-x16.c | 158 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local 164 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() 188 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
|
D | neon-rr2-lut2048-p1-nr2recps-x16.c | 160 const float32x4_t vy89AB = vmlaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() local 166 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() 190 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x16.c | 152 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local 158 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 182 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
|
D | neonfma-rr1-lut64-p2-div-x20.c | 175 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() local 182 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20() 189 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_div_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 175 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local 182 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() 210 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
|
D | neonfma-rr1-lut64-p2-nr2fma-x20.c | 175 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 182 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 210 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x20.c | 168 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() local 175 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20() 203 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x20()
|