/external/XNNPACK/src/f32-vsqrt/gen/ |
D | neonfma-nr2fma1adj-x24.c | 62 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() local 74 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 75 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 82 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 94 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24() 95 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x24()
|
D | neonfma-nr2fma1adj-x28.c | 66 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() local 79 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 80 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 89 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 102 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28() 103 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x28()
|
D | neonfma-nr2fma1adj-x32.c | 70 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() local 84 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 85 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 96 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 110 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32() 111 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32()
|
D | neonfma-nr2fma1adj-x36.c | 74 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() local 89 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 90 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 103 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 118 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36() 119 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x36()
|
D | neonfma-nr2fma1adj-x40.c | 78 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() local 94 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 95 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 110 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 126 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40() 127 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x40()
|
D | neonfma-nr1rsqrts1fma1adj-x24.c | 83 const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() local 95 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24() 96 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x24()
|
D | neonfma-nr1rsqrts1fma1adj-x28.c | 90 const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() local 103 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28() 104 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x28()
|
D | neonfma-nr1rsqrts1fma1adj-x32.c | 97 const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() local 111 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32() 112 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x32()
|
D | neonfma-nr1rsqrts1fma1adj-x36.c | 104 const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() local 119 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36() 120 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x36()
|
D | neonfma-nr1rsqrts1fma1adj-x40.c | 111 const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() local 127 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40() 128 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN); in xnn_f32_vsqrt_ukernel__neonfma_nr1rsqrts1fma1adj_x40()
|