/external/XNNPACK/src/f32-vsigmoid/gen/ |
D | vsigmoid-neonfma-rr1-p5-nr2fma-x16.c | 106 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 111 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 116 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 121 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x16.c | 106 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 111 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 116 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 121 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x20.c | 119 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 125 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 131 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 137 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x16.c | 125 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() local 130 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 135 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16() 140 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x16.c | 131 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 136 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 141 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 146 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x16.c | 106 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() local 111 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 116 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16() 121 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x20.c | 119 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 125 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 131 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 137 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | vsigmoid-neonfma-rr1-p5-nr2fma-x24.c | 132 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() local 139 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 146 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24() 153 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2fma_x24()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x20.c | 150 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() local 156 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 162 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20() 168 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x20()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x20.c | 143 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() local 149 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 155 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20() 161 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x20()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x16.c | 131 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() local 136 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 141 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16() 146 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x16()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c | 125 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() local 130 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 135 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16() 140 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x16()
|
D | vsigmoid-neon-rr2-p5-nr2recps-x16.c | 112 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() local 117 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 122 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16() 127 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x16()
|
D | vsigmoid-neonfma-rr1-p5-nr1recps1fma-x24.c | 132 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() local 139 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 146 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24() 153 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x24()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2recps-x16.c | 125 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() local 130 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() 135 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16() 140 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x16()
|
D | vsigmoid-neon-rr2-p5-nr2recps-x20.c | 126 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20() local 132 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20() 138 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20() 144 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x20.c | 119 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() local 125 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 131 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20() 137 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x20()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr2fma-x24.c | 161 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() local 168 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 175 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24() 182 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x24()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2fma-x24.c | 169 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() local 176 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 183 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24() 190 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x24()
|
D | vsigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c | 143 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() local 149 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() 155 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20() 161 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x20()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr1recps1fma-x20.c | 150 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() local 156 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() 162 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20() 168 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x20()
|
D | vsigmoid-neonfma-rr1-p5-nr2recps-x24.c | 132 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() local 139 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 146 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24() 153 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_p5_nr2recps_x24()
|
D | vsigmoid-neonfma-rr1-lut64-p2-nr2recps-x16.c | 131 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() local 136 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() 141 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16() 146 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16()
|
D | vsigmoid-neon-rr2-lut2048-p1-nr2recps-x16.c | 131 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() local 136 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() 141 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16() 146 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x16()
|
D | vsigmoid-neon-rr2-lut64-p2-nr2recps-x16.c | 137 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() local 142 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() 147 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF)); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16() 152 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF); in xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x16()
|