/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x8.c | 41 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local 44 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 45 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 49 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 52 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 53 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8()
|
D | neonfma-nr2fma1adj-x12.c | 45 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local 49 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 50 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 56 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 60 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 61 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12()
|
D | neonfma-nr2fma1adj-x16.c | 49 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 54 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 55 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 63 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 68 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 69 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 53 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 59 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 60 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 70 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 76 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 77 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 57 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 64 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 65 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 77 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 84 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 85 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 61 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 69 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 70 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 84 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 92 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 93 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 65 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 74 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 75 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 91 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 100 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 101 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 69 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 79 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 80 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 98 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 108 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 109 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 73 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 84 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 85 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 105 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 116 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 117 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x8.c | 50 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local 53 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 54 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8()
|
D | neonfma-nr1rsqrts1fma1adj-x12.c | 57 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local 61 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 62 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 64 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 69 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 70 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 71 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 77 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 78 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 78 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 85 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 86 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 85 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 93 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 94 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 92 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 101 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 102 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 99 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 109 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 110 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 106 const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 117 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 118 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|