• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsqrt/neonfma-nr1rsqrts1fma1adj.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_nr1rsqrts1fma1adj_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_nr1rsqrts1fma1adj_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     float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123);
40     float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567);
41     float32x4_t vrsqrtx89AB = vrsqrteq_f32(vx89AB);
42     float32x4_t vrsqrtxCDEF = vrsqrteq_f32(vxCDEF);
43     float32x4_t vrsqrtxGHIJ = vrsqrteq_f32(vxGHIJ);
44     float32x4_t vrsqrtxKLMN = vrsqrteq_f32(vxKLMN);
45     float32x4_t vrsqrtxOPQR = vrsqrteq_f32(vxOPQR);
46     float32x4_t vrsqrtxSTUV = vrsqrteq_f32(vxSTUV);
47 
48     const float32x4_t vrx0123 = vmulq_f32(vrsqrtx0123, vrsqrtx0123);
49     const float32x4_t vrx4567 = vmulq_f32(vrsqrtx4567, vrsqrtx4567);
50     const float32x4_t vrx89AB = vmulq_f32(vrsqrtx89AB, vrsqrtx89AB);
51     const float32x4_t vrxCDEF = vmulq_f32(vrsqrtxCDEF, vrsqrtxCDEF);
52     const float32x4_t vrxGHIJ = vmulq_f32(vrsqrtxGHIJ, vrsqrtxGHIJ);
53     const float32x4_t vrxKLMN = vmulq_f32(vrsqrtxKLMN, vrsqrtxKLMN);
54     const float32x4_t vrxOPQR = vmulq_f32(vrsqrtxOPQR, vrsqrtxOPQR);
55     const float32x4_t vrxSTUV = vmulq_f32(vrsqrtxSTUV, vrsqrtxSTUV);
56 
57     const float32x4_t vcorrection0123 = vrsqrtsq_f32(vx0123, vrx0123);
58     const float32x4_t vcorrection4567 = vrsqrtsq_f32(vx4567, vrx4567);
59     const float32x4_t vcorrection89AB = vrsqrtsq_f32(vx89AB, vrx89AB);
60     const float32x4_t vcorrectionCDEF = vrsqrtsq_f32(vxCDEF, vrxCDEF);
61     const float32x4_t vcorrectionGHIJ = vrsqrtsq_f32(vxGHIJ, vrxGHIJ);
62     const float32x4_t vcorrectionKLMN = vrsqrtsq_f32(vxKLMN, vrxKLMN);
63     const float32x4_t vcorrectionOPQR = vrsqrtsq_f32(vxOPQR, vrxOPQR);
64     const float32x4_t vcorrectionSTUV = vrsqrtsq_f32(vxSTUV, vrxSTUV);
65 
66     vrsqrtx0123 = vmulq_f32(vrsqrtx0123, vcorrection0123);
67     vrsqrtx4567 = vmulq_f32(vrsqrtx4567, vcorrection4567);
68     vrsqrtx89AB = vmulq_f32(vrsqrtx89AB, vcorrection89AB);
69     vrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vcorrectionCDEF);
70     vrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vcorrectionGHIJ);
71     vrsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vcorrectionKLMN);
72     vrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vcorrectionOPQR);
73     vrsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vcorrectionSTUV);
74 
75     float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123);
76     float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf);
77     float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567);
78     float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf);
79     float32x4_t vsqrtx89AB = vmulq_f32(vrsqrtx89AB, vx89AB);
80     float32x4_t vhalfrsqrtx89AB = vmulq_f32(vrsqrtx89AB, vhalf);
81     float32x4_t vsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vxCDEF);
82     float32x4_t vhalfrsqrtxCDEF = vmulq_f32(vrsqrtxCDEF, vhalf);
83     float32x4_t vsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vxGHIJ);
84     float32x4_t vhalfrsqrtxGHIJ = vmulq_f32(vrsqrtxGHIJ, vhalf);
85     float32x4_t vsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vxKLMN);
86     float32x4_t vhalfrsqrtxKLMN = vmulq_f32(vrsqrtxKLMN, vhalf);
87     float32x4_t vsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vxOPQR);
88     float32x4_t vhalfrsqrtxOPQR = vmulq_f32(vrsqrtxOPQR, vhalf);
89     float32x4_t vsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vxSTUV);
90     float32x4_t vhalfrsqrtxSTUV = vmulq_f32(vrsqrtxSTUV, vhalf);
91 
92     const float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
93     const float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
94     const float32x4_t vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
95     const float32x4_t vresidualCDEF = vfmsq_f32(vhalf, vsqrtxCDEF, vhalfrsqrtxCDEF);
96     const float32x4_t vresidualGHIJ = vfmsq_f32(vhalf, vsqrtxGHIJ, vhalfrsqrtxGHIJ);
97     const float32x4_t vresidualKLMN = vfmsq_f32(vhalf, vsqrtxKLMN, vhalfrsqrtxKLMN);
98     const float32x4_t vresidualOPQR = vfmsq_f32(vhalf, vsqrtxOPQR, vhalfrsqrtxOPQR);
99     const float32x4_t vresidualSTUV = vfmsq_f32(vhalf, vsqrtxSTUV, vhalfrsqrtxSTUV);
100 
101     vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
102     vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
103     vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
104     vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
105     vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
106     vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
107     vhalfrsqrtxCDEF = vfmaq_f32(vhalfrsqrtxCDEF, vresidualCDEF, vhalfrsqrtxCDEF);
108     vsqrtxCDEF = vfmaq_f32(vsqrtxCDEF, vresidualCDEF, vsqrtxCDEF);
109     vhalfrsqrtxGHIJ = vfmaq_f32(vhalfrsqrtxGHIJ, vresidualGHIJ, vhalfrsqrtxGHIJ);
110     vsqrtxGHIJ = vfmaq_f32(vsqrtxGHIJ, vresidualGHIJ, vsqrtxGHIJ);
111     vhalfrsqrtxKLMN = vfmaq_f32(vhalfrsqrtxKLMN, vresidualKLMN, vhalfrsqrtxKLMN);
112     vsqrtxKLMN = vfmaq_f32(vsqrtxKLMN, vresidualKLMN, vsqrtxKLMN);
113     vhalfrsqrtxOPQR = vfmaq_f32(vhalfrsqrtxOPQR, vresidualOPQR, vhalfrsqrtxOPQR);
114     vsqrtxOPQR = vfmaq_f32(vsqrtxOPQR, vresidualOPQR, vsqrtxOPQR);
115     vhalfrsqrtxSTUV = vfmaq_f32(vhalfrsqrtxSTUV, vresidualSTUV, vhalfrsqrtxSTUV);
116     vsqrtxSTUV = vfmaq_f32(vsqrtxSTUV, vresidualSTUV, vsqrtxSTUV);
117 
118     const float32x4_t vadjustment0123 = vfmsq_f32(vx0123, vsqrtx0123, vsqrtx0123);
119     const float32x4_t vadjustment4567 = vfmsq_f32(vx4567, vsqrtx4567, vsqrtx4567);
120     const float32x4_t vadjustment89AB = vfmsq_f32(vx89AB, vsqrtx89AB, vsqrtx89AB);
121     const float32x4_t vadjustmentCDEF = vfmsq_f32(vxCDEF, vsqrtxCDEF, vsqrtxCDEF);
122     const float32x4_t vadjustmentGHIJ = vfmsq_f32(vxGHIJ, vsqrtxGHIJ, vsqrtxGHIJ);
123     const float32x4_t vadjustmentKLMN = vfmsq_f32(vxKLMN, vsqrtxKLMN, vsqrtxKLMN);
124     const float32x4_t vadjustmentOPQR = vfmsq_f32(vxOPQR, vsqrtxOPQR, vsqrtxOPQR);
125     const float32x4_t vadjustmentSTUV = vfmsq_f32(vxSTUV, vsqrtxSTUV, vsqrtxSTUV);
126 
127     const float32x4_t vy0123 = vfmaq_f32(vsqrtx0123, vhalfrsqrtx0123, vadjustment0123);
128     const float32x4_t vy4567 = vfmaq_f32(vsqrtx4567, vhalfrsqrtx4567, vadjustment4567);
129     const float32x4_t vy89AB = vfmaq_f32(vsqrtx89AB, vhalfrsqrtx89AB, vadjustment89AB);
130     const float32x4_t vyCDEF = vfmaq_f32(vsqrtxCDEF, vhalfrsqrtxCDEF, vadjustmentCDEF);
131     const float32x4_t vyGHIJ = vfmaq_f32(vsqrtxGHIJ, vhalfrsqrtxGHIJ, vadjustmentGHIJ);
132     const float32x4_t vyKLMN = vfmaq_f32(vsqrtxKLMN, vhalfrsqrtxKLMN, vadjustmentKLMN);
133     const float32x4_t vyOPQR = vfmaq_f32(vsqrtxOPQR, vhalfrsqrtxOPQR, vadjustmentOPQR);
134     const float32x4_t vySTUV = vfmaq_f32(vsqrtxSTUV, vhalfrsqrtxSTUV, vadjustmentSTUV);
135 
136     vst1q_f32(y, vy0123); y += 4;
137     vst1q_f32(y, vy4567); y += 4;
138     vst1q_f32(y, vy89AB); y += 4;
139     vst1q_f32(y, vyCDEF); y += 4;
140     vst1q_f32(y, vyGHIJ); y += 4;
141     vst1q_f32(y, vyKLMN); y += 4;
142     vst1q_f32(y, vyOPQR); y += 4;
143     vst1q_f32(y, vySTUV); y += 4;
144   }
145   if XNN_UNLIKELY(n != 0) {
146     do {
147       const float vx = *x++;
148       const float vy = sqrtf(vx);
149       *y++ = vy;
150       n -= sizeof(float);
151     } while (n != 0);
152   }
153 }
154