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_x36(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_x36(
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 >= 36 * sizeof(float); n -= 36 * 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 const float32x4_t vxWXYZ = vld1q_f32(x); x += 4;
39
40 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123);
41 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567);
42 const float32x4_t vrsqrtx89AB = vrsqrteq_f32(vx89AB);
43 const float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF);
44 const float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ);
45 const float32x4_t vrsqrtxKLMN = vrsqrteq_f32(vxKLMN);
46 const float32x4_t vrsqrtxOPQR = vrsqrteq_f32(vxOPQR);
47 const float32x4_t vrsqrtxSTUV = vrsqrteq_f32(vxSTUV);
48 const float32x4_t vrsqrtxWXYZ = vrsqrteq_f32(vxWXYZ);
49
50 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123);
51 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf);
52 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567);
53 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf);
54 float32x4_t vsqrtx89AB = vmulq_f32(vrsqrtx89AB, vx89AB);
55 float32x4_t vhalfrsqrtx89AB = vmulq_f32(vrsqrtx89AB, vhalf);
56 float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF);
57 float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf);
58 float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ);
59 float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf);
60 float32x4_t vsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vxKLMN);
61 float32x4_t vhalfrsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vhalf);
62 float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR);
63 float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf);
64 float32x4_t vsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vxSTUV);
65 float32x4_t vhalfrsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vhalf);
66 float32x4_t vsqrtxWXYZ = vmulq_f32(vrsqrtxWXYZ, vxWXYZ);
67 float32x4_t vhalfrsqrtxWXYZ = vmulq_f32(vrsqrtxWXYZ, vhalf);
68
69 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
70 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
71 float32x4_t vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
72 float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF);
73 float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ);
74 float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN);
75 float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR);
76 float32x4_t vresidualSTUV = vfmsq_f32(vhalf, vsqrtxSTUV, vhalfrsqrtxSTUV);
77 float32x4_t vresidualWXYZ = vfmsq_f32(vhalf, vsqrtxWXYZ, vhalfrsqrtxWXYZ);
78
79 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
80 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
81 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
82 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
83 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
84 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
85 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF);
86 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF);
87 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ);
88 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ);
89 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN);
90 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN);
91 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR);
92 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR);
93 vhalfrsqrtxSTUV = vfmaq_f32(vhalfrsqrtxSTUV, vresidualSTUV, vhalfrsqrtxSTUV);
94 vsqrtxSTUV = vfmaq_f32(vsqrtxSTUV, vresidualSTUV, vsqrtxSTUV);
95 vhalfrsqrtxWXYZ = vfmaq_f32(vhalfrsqrtxWXYZ, vresidualWXYZ, vhalfrsqrtxWXYZ);
96 vsqrtxWXYZ = vfmaq_f32(vsqrtxWXYZ, vresidualWXYZ, vsqrtxWXYZ);
97
98 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
99 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
100 vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
101 vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF);
102 vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ);
103 vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN);
104 vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR);
105 vresidualSTUV = vfmsq_f32(vhalf, vsqrtxSTUV, vhalfrsqrtxSTUV);
106 vresidualWXYZ = vfmsq_f32(vhalf, vsqrtxWXYZ, vhalfrsqrtxWXYZ);
107
108 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
109 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
110 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
111 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
112 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
113 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
114 vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF);
115 vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF);
116 vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ);
117 vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ);
118 vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN);
119 vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN);
120 vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR);
121 vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR);
122 vhalfrsqrtxSTUV = vfmaq_f32(vhalfrsqrtxSTUV, vresidualSTUV, vhalfrsqrtxSTUV);
123 vsqrtxSTUV = vfmaq_f32(vsqrtxSTUV, vresidualSTUV, vsqrtxSTUV);
124 vhalfrsqrtxWXYZ = vfmaq_f32(vhalfrsqrtxWXYZ, vresidualWXYZ, vhalfrsqrtxWXYZ);
125 vsqrtxWXYZ = vfmaq_f32(vsqrtxWXYZ, vresidualWXYZ, vsqrtxWXYZ);
126
127 const float32x4_t vadjustment0123 = vfmsq_f32(vx0123, vsqrtx0123, vsqrtx0123);
128 const float32x4_t vadjustment4567 = vfmsq_f32(vx4567, vsqrtx4567, vsqrtx4567);
129 const float32x4_t vadjustment89AB = vfmsq_f32(vx89AB, vsqrtx89AB, vsqrtx89AB);
130 const float32x4_t vadjustmentCDEF = vfmsq_f32(vxCDEF, vsqrtxCDEF, vsqrtxCDEF);
131 const float32x4_t vadjustmentGHIJ = vfmsq_f32(vxGHIJ, vsqrtxGHIJ, vsqrtxGHIJ);
132 const float32x4_t vadjustmentKLMN = vfmsq_f32(vxKLMN, vsqrtxKLMN, vsqrtxKLMN);
133 const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR);
134 const float32x4_t vadjustmentSTUV = vfmsq_f32(vxSTUV, vsqrtxSTUV, vsqrtxSTUV);
135 const float32x4_t vadjustmentWXYZ = vfmsq_f32(vxWXYZ, vsqrtxWXYZ, vsqrtxWXYZ);
136
137 const float32x4_t vy0123 = vfmaq_f32(vsqrtx0123, vhalfrsqrtx0123, vadjustment0123);
138 const float32x4_t vy4567 = vfmaq_f32(vsqrtx4567, vhalfrsqrtx4567, vadjustment4567);
139 const float32x4_t vy89AB = vfmaq_f32(vsqrtx89AB, vhalfrsqrtx89AB, vadjustment89AB);
140 const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF);
141 const float32x4_t vyGHIJ = vfmaq_f32(vsqrtxGHIJ, vhalfrsqrtxGHIJ, vadjustmentGHIJ);
142 const float32x4_t vyKLMN = vfmaq_f32(vsqrtxKLMN, vhalfrsqrtxKLMN, vadjustmentKLMN);
143 const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR);
144 const float32x4_t vySTUV = vfmaq_f32(vsqrtxSTUV, vhalfrsqrtxSTUV, vadjustmentSTUV);
145 const float32x4_t vyWXYZ = vfmaq_f32(vsqrtxWXYZ, vhalfrsqrtxWXYZ, vadjustmentWXYZ);
146
147 vst1q_f32(y, vy0123); y += 4;
148 vst1q_f32(y, vy4567); y += 4;
149 vst1q_f32(y, vy89AB); y += 4;
150 vst1q_f32(y, vyCDEF); y += 4;
151 vst1q_f32(y, vyGHIJ); y += 4;
152 vst1q_f32(y, vyKLMN); y += 4;
153 vst1q_f32(y, vyOPQR); y += 4;
154 vst1q_f32(y, vySTUV); y += 4;
155 vst1q_f32(y, vyWXYZ); y += 4;
156 }
157 if XNN_UNLIKELY(n != 0) {
158 do {
159 const float vx = *x++;
160 const float vy = sqrtf(vx);
161 *y++ = vy;
162 n -= sizeof(float);
163 } while (n != 0);
164 }
165 }
166