/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr1rsqrts1fma1adj-x20.c | 40 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 46 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 58 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 68 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 69 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 41 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 48 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 62 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 73 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 74 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 42 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 50 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 66 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 78 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 79 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 43 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 52 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 70 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 83 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 84 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 44 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 54 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 74 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 88 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 89 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 45 float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 56 const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 78 vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 93 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 94 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|
D | neonfma-nr2fma1adj-x20.c | 40 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 50 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 51 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 41 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 52 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 53 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 42 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 54 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 55 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 43 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 56 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 57 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 44 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 58 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 59 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 45 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 60 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 61 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|