/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x28.c | 58 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 67 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 82 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 90 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 105 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 113 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 121 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 60 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 71 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 87 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 97 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 113 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 123 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 132 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 62 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 75 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 92 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 104 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 121 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 133 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 143 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 64 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 79 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 97 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 111 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 129 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 143 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 154 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 82 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 91 const float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 106 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 114 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 122 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 87 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 98 const float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 114 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 124 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 133 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 92 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 105 const float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 122 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 134 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 144 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 97 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 112 const float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 130 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 144 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 155 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|