/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x28.c | 59 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 81 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 90 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 104 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 61 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 86 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 97 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 112 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 63 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 91 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 104 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 120 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 65 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 96 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 111 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 128 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 83 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 105 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 88 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 113 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 93 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 121 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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 | 98 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf); 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() 129 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR); 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()
|