/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x8.c | 42 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() local 46 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 47 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 50 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 54 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8() 55 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x8()
|
D | neonfma-nr2fma1adj-x12.c | 46 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() local 51 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 52 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 57 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 62 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12() 63 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x12()
|
D | neonfma-nr2fma1adj-x16.c | 50 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 56 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 57 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 64 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 70 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 71 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 54 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 61 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 62 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 71 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 78 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 79 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 58 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 66 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 67 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 78 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 86 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 87 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 62 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 71 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 72 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 85 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 94 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 95 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x8.c | 51 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() local 55 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8() 56 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x8()
|
D | neonfma-nr2fma1adj-x32.c | 66 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 76 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 77 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 92 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 102 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 103 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 70 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 81 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 82 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 99 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 110 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 111 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x12.c | 58 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() local 63 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12() 64 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x12()
|
D | neonfma-nr2fma1adj-x40.c | 74 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 86 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 87 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 106 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 118 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 119 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 65 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 71 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 72 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 72 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 79 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 80 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 79 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 87 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 88 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 86 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 95 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 96 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 93 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 103 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 104 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 100 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 111 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 112 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 107 const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 119 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 120 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|