/external/XNNPACK/src/f32-sigmoid/gen/ |
D | neonfma-rr1-p5-nr2fma-x8.c | 78 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8() local 81 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8() 84 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8() 87 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x8()
|
D | neonfma-rr1-p5-nr1recps1fma-x8.c | 78 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() local 81 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() 84 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8() 87 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x8()
|
D | neonfma-rr1-p5-nr2fma-x12.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() local 95 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() 99 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12() 103 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x12()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x8.c | 87 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8() local 90 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8() 93 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8() 96 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8()
|
D | neonfma-rr1-lut64-p2-nr2fma-x8.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8() local 94 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8() 97 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8() 100 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x8()
|
D | neon-rr2-p5-nr2recps-x8.c | 82 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() local 85 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() 88 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8() 91 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x8()
|
D | neonfma-rr1-p5-nr2recps-x8.c | 78 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() local 81 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() 84 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8() 87 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x8()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x8.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8() local 94 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8() 97 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8() 100 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x8()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c | 87 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8() local 90 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8() 93 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8() 96 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x8()
|
D | neonfma-rr1-p5-nr1recps1fma-x12.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() local 95 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() 99 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12() 103 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x12()
|
D | neonfma-rr1-lut2048-p1-nr2fma-x12.c | 105 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() local 109 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() 113 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12() 117 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x12()
|
D | neonfma-rr1-p5-nr2fma-x16.c | 104 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() local 109 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 114 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16() 119 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x16()
|
D | neonfma-rr1-lut64-p2-nr2fma-x12.c | 110 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() local 114 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() 118 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12() 122 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x12()
|
D | neonfma-rr1-lut64-p2-nr2recps-x8.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8() local 94 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8() 97 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8() 100 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x8()
|
D | neon-rr2-lut64-p2-nr2recps-x8.c | 95 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8() local 98 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8() 101 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8() 104 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8()
|
D | neon-rr2-p5-nr2recps-x12.c | 96 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() local 100 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() 104 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12() 108 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neon_rr2_p5_nr2recps_x12()
|
D | neonfma-rr1-p5-nr2recps-x12.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() local 95 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() 99 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12() 103 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2recps_x12()
|
D | neon-rr2-lut2048-p1-nr2recps-x8.c | 91 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() local 94 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() 97 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8() 100 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neon_rr2_lut2048_p1_nr2recps_x8()
|
D | neonfma-rr1-lut2048-p1-nr2recps-x8.c | 87 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() local 90 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() 93 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8() 96 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2recps_x8()
|
D | neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c | 105 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() local 109 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() 113 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12() 117 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr1recps1fma_x12()
|
D | neonfma-rr1-p5-nr1recps1fma-x16.c | 104 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() local 109 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 114 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16() 119 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x16()
|
D | neonfma-rr1-p5-nr2fma-x20.c | 117 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() local 123 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 129 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20() 135 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr2fma_x20()
|
D | neonfma-rr1-lut64-p2-nr1recps1fma-x12.c | 110 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() local 114 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() 118 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12() 122 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x12()
|
D | neonfma-rr1-p5-nr1recps1fma-x20.c | 117 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() local 123 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 129 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20() 135 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_p5_nr1recps1fma_x20()
|
D | neonfma-rr1-lut64-p2-nr2fma-x16.c | 129 float32x4_t vr4567 = vrecpeq_f32(vd4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() local 134 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 139 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567)); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16() 144 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567); in xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2fma_x16()
|