/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x16.c | 47 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() local 52 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16() 60 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 85 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x16()
|
D | neonfma-nr2fma1adj-x20.c | 49 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() local 56 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20() 65 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 96 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x20()
|
D | neonfma-nr2fma1adj-x24.c | 51 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 60 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 70 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 107 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 53 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 64 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 75 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 118 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 55 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 68 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 80 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 129 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 57 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 72 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 85 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 140 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x16.c | 62 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() local 67 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 75 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16() 86 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x16()
|
D | neonfma-nr2fma1adj-x40.c | 59 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 76 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 90 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); 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() 151 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x20.c | 67 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() local 74 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 83 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20() 97 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x20()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 72 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 81 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 91 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 108 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 77 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 88 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 99 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 119 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 82 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 95 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 107 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 130 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 87 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 102 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 115 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 141 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 92 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 109 const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 123 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 152 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|