/external/XNNPACK/src/f32-vsigmoid/gen/ |
D | vsigmoid-neonfma-rr1-p5-nr2fma-x12.c | 92 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() local 96 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() 100 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() 104 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x12.c | 92 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() local 96 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() 100 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() 104 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x16.c | 105 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 110 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 115 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 120 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x12.c | 106 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() local 110 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() 114 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() 118 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x12.c | 111 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() local 115 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() 119 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() 123 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x12.c | 92 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() local 96 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() 100 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() 104 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c | 105 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 110 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 115 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 120 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c | 106 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() local 110 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() 114 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() 118 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x20.c | 118 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 124 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 130 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 136 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | vsigmoid-neon-rr2-p5-nr2recps-x12.c | 97 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local 101 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x12() 105 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x12() 109 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x12.c | 111 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() local 115 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() 119 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() 123 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x16.c | 124 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local 129 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 134 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 139 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x16.c | 130 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 135 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 140 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 145 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x16.c | 105 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() local 110 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 115 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 120 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x12.c | 106 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() local 110 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() 114 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12() 118 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x12()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c | 118 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 124 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 130 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 136 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x24.c | 131 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local 138 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 145 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 152 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c | 149 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 155 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 161 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 167 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x20.c | 142 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local 148 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 154 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 160 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 130 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 135 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 140 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 145 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 124 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 129 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 134 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 139 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
D | vsigmoid-neon-rr2-lut2048-p1-nr2recps-x12.c | 111 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12() local 115 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12() 119 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12() 123 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x12()
|
D | vsigmoid-neon-rr2-p5-nr2recps-x16.c | 111 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local 116 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 121 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 126 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x12.c | 111 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() local 115 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() 119 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12() 123 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x12()
|
D | vsigmoid-neon-rr2-lut64-p2-nr2recps-x12.c | 116 float32x4_t vr89AB = vrecpeq_f32(vd89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12() local 120 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12() 124 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12() 128 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x12()
|