/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr1rsqrts1fma1adj-x8.c | 33 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local 36 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 42 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 45 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 46 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8()
|
D | neonfma-nr1rsqrts1fma1adj-x12.c | 34 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local 38 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 46 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 50 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 51 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 35 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 40 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 50 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 55 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 56 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 36 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 42 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 54 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 60 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 61 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 37 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 44 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 58 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 65 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 66 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 38 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 46 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 62 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 70 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 71 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 39 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 48 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 66 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 75 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 76 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 40 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 50 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 70 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 80 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 81 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr2fma1adj-x8.c | 33 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local 36 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 37 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 41 float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 52 const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 74 vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 85 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 86 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|
D | neonfma-nr2fma1adj-x12.c | 34 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local 38 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 39 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12()
|
D | neonfma-nr2fma1adj-x16.c | 35 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 40 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 41 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 36 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 42 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 43 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 37 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 44 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 45 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 38 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 46 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 47 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 39 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 48 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 49 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 40 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 50 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 51 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 41 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 52 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 53 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|