/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x16.c | 52 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 60 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 61 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 66 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 74 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 75 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 56 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 65 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 66 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 73 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 82 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 83 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 60 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 70 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 71 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 80 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 90 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 91 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 64 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 75 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 76 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 87 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 98 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 99 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 68 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 80 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 81 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 94 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 106 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 107 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 72 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 85 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 86 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 101 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 114 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 115 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 76 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 90 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 91 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 108 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 122 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 123 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 67 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 75 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 76 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 74 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 83 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 84 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 81 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 91 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 92 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 88 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 99 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 100 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 95 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 107 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 108 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 102 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 115 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 116 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 109 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 123 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 124 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|