1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsqrt/neonfma-nr2fma1adj.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11 #include <math.h>
12
13 #include <arm_neon.h>
14
15 #include <xnnpack/common.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsqrt_ukernel__neonfma_nr2fma1adj_x32(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const float32x4_t vhalf = vmovq_n_f32(0.5f);
29 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
32 const float32x4_t vx89AB = vld1q_f32(x); x += 4;
33 const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
34 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
35 const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
36 const float32x4_t vxOPQR = vld1q_f32(x); x += 4;
37 const float32x4_t vxSTUV = vld1q_f32(x); x += 4;
38
39 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123);
40 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567);
41 const float32x4_t vrsqrtx89AB = vrsqrteq_f32(vx89AB);
42 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF);
43 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ);
44 const float32x4_t vrsqrtxKLMN = vrsqrteq_f32(vxKLMN);
45 const float32x4_t vrsqrtxOPQR = vrsqrteq_f32(vxOPQR);
46 const float32x4_t vrsqrtxSTUV = vrsqrteq_f32(vxSTUV);
47
48 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123);
49 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf);
50 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567);
51 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf);
52 float32x4_t vsqrtx89AB = vmulq_f32(vrsqrtx89AB, vx89AB);
53 float32x4_t vhalfrsqrtx89AB = vmulq_f32(vrsqrtx89AB, vhalf);
54 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF);
55 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf);
56 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ);
57 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf);
58 float32x4_t vsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vxKLMN);
59 float32x4_t vhalfrsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vhalf);
60 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR);
61 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf);
62 float32x4_t vsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vxSTUV);
63 float32x4_t vhalfrsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vhalf);
64
65 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
66 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
67 float32x4_t vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
68 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF);
69 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ);
70 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN);
71 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR);
72 float32x4_t vresidualSTUV = vfmsq_f32(vhalf, vsqrtxSTUV, vhalfrsqrtxSTUV);
73
74 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
75 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
76 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
77 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
78 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
79 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
80 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF);
81 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF);
82 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ);
83 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ);
84 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN);
85 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN);
86 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR);
87 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR);
88 vhalfrsqrtxSTUV = vfmaq_f32(vhalfrsqrtxSTUV, vresidualSTUV, vhalfrsqrtxSTUV);
89 vsqrtxSTUV = vfmaq_f32(vsqrtxSTUV, vresidualSTUV, vsqrtxSTUV);
90
91 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
92 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
93 vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
94 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF);
95 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ);
96 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN);
97 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR);
98 vresidualSTUV = vfmsq_f32(vhalf, vsqrtxSTUV, vhalfrsqrtxSTUV);
99
100 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
101 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
102 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
103 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
104 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
105 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
106 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF);
107 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF);
108 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ);
109 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ);
110 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN);
111 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN);
112 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR);
113 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR);
114 vhalfrsqrtxSTUV = vfmaq_f32(vhalfrsqrtxSTUV, vresidualSTUV, vhalfrsqrtxSTUV);
115 vsqrtxSTUV = vfmaq_f32(vsqrtxSTUV, vresidualSTUV, vsqrtxSTUV);
116
117 const float32x4_t vadjustment0123 = vfmsq_f32(vx0123, vsqrtx0123, vsqrtx0123);
118 const float32x4_t vadjustment4567 = vfmsq_f32(vx4567, vsqrtx4567, vsqrtx4567);
119 const float32x4_t vadjustment89AB = vfmsq_f32(vx89AB, vsqrtx89AB, vsqrtx89AB);
120 const float32x4_t vadjustmentCDEF = vfmsq_f32(vxCDEF, vsqrtxCDEF, vsqrtxCDEF);
121 const float32x4_t vadjustmentGHIJ = vfmsq_f32(vxGHIJ, vsqrtxGHIJ, vsqrtxGHIJ);
122 const float32x4_t vadjustmentKLMN = vfmsq_f32(vxKLMN, vsqrtxKLMN, vsqrtxKLMN);
123 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR);
124 const float32x4_t vadjustmentSTUV = vfmsq_f32(vxSTUV, vsqrtxSTUV, vsqrtxSTUV);
125
126 const float32x4_t vy0123 = vfmaq_f32(vsqrtx0123, vhalfrsqrtx0123, vadjustment0123);
127 const float32x4_t vy4567 = vfmaq_f32(vsqrtx4567, vhalfrsqrtx4567, vadjustment4567);
128 const float32x4_t vy89AB = vfmaq_f32(vsqrtx89AB, vhalfrsqrtx89AB, vadjustment89AB);
129 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF);
130 const float32x4_t vyGHIJ = vfmaq_f32(vsqrtxGHIJ, vhalfrsqrtxGHIJ, vadjustmentGHIJ);
131 const float32x4_t vyKLMN = vfmaq_f32(vsqrtxKLMN, vhalfrsqrtxKLMN, vadjustmentKLMN);
132 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR);
133 const float32x4_t vySTUV = vfmaq_f32(vsqrtxSTUV, vhalfrsqrtxSTUV, vadjustmentSTUV);
134
135 vst1q_f32(y, vy0123); y += 4;
136 vst1q_f32(y, vy4567); y += 4;
137 vst1q_f32(y, vy89AB); y += 4;
138 vst1q_f32(y, vyCDEF); y += 4;
139 vst1q_f32(y, vyGHIJ); y += 4;
140 vst1q_f32(y, vyKLMN); y += 4;
141 vst1q_f32(y, vyOPQR); y += 4;
142 vst1q_f32(y, vySTUV); y += 4;
143 }
144 if XNN_UNLIKELY(n != 0) {
145 do {
146 const float vx = *x++;
147 const float vy = sqrtf(vx);
148 *y++ = vy;
149 n -= sizeof(float);
150 } while (n != 0);
151 }
152 }
153