/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x20.c | 57 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 67 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 68 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 74 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 84 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 85 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 61 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 72 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 73 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 81 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 92 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 93 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 65 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 77 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 78 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 88 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 100 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 101 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 69 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 82 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 83 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 95 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 108 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 109 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 73 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 87 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 88 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 102 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 116 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 117 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 77 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 92 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 93 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 109 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 124 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 125 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 75 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 85 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 86 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 82 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 93 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 94 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 89 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 101 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 102 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 96 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 109 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 110 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 103 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 117 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 118 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 110 const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 125 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 126 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|