/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr1rsqrts1fma1adj-x8.c | 34 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local 37 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 43 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 47 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 48 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8()
|
D | neonfma-nr1rsqrts1fma1adj-x12.c | 35 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local 39 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 47 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 52 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 53 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 36 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 41 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 51 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 57 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 58 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 37 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 43 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 55 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 62 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 63 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 38 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 45 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 59 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 67 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 68 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 39 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 47 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 63 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 72 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 73 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 40 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 49 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 67 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 77 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 78 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 41 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 51 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 71 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 82 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 83 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr2fma1adj-x8.c | 34 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local 38 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 39 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 42 float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 53 const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 75 vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 87 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 88 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|
D | neonfma-nr2fma1adj-x12.c | 35 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local 40 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 41 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12()
|
D | neonfma-nr2fma1adj-x16.c | 36 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 42 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 43 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 37 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 44 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 45 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 38 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 46 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 47 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 39 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 48 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 49 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 40 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 50 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 51 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 41 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 52 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 53 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 42 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 54 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 55 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|