/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr1rsqrts1fma1adj-x16.c | 38 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 43 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 53 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 61 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 62 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 39 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 45 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 57 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 66 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 67 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 40 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 47 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 61 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 71 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 72 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 41 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 49 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 65 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 76 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 77 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 42 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 51 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 69 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 81 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 82 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 43 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 53 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 73 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 86 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 87 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 44 float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 55 const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 77 vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 91 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 92 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|
D | neonfma-nr2fma1adj-x16.c | 38 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 46 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 47 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 39 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 48 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 49 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 40 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 50 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 51 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 41 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 52 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 53 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 42 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 54 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 55 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 43 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 56 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 57 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 44 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 58 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 59 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|