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_x12(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_x12(
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_OOB_READS
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 >= 12 * sizeof(float); n -= 12 * 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
34 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123);
35 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567);
36 const float32x4_t vrsqrtx89AB = vrsqrteq_f32(vx89AB);
37
38 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123);
39 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf);
40 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567);
41 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf);
42 float32x4_t vsqrtx89AB = vmulq_f32(vrsqrtx89AB, vx89AB);
43 float32x4_t vhalfrsqrtx89AB = vmulq_f32(vrsqrtx89AB, vhalf);
44
45 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
46 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
47 float32x4_t vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
48
49 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
50 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
51 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
52 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
53 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
54 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
55
56 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
57 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
58 vresidual89AB = vfmsq_f32(vhalf, vsqrtx89AB, vhalfrsqrtx89AB);
59
60 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
61 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
62 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
63 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
64 vhalfrsqrtx89AB = vfmaq_f32(vhalfrsqrtx89AB, vresidual89AB, vhalfrsqrtx89AB);
65 vsqrtx89AB = vfmaq_f32(vsqrtx89AB, vresidual89AB, vsqrtx89AB);
66
67 const float32x4_t vadjustment0123 = vfmsq_f32(vx0123, vsqrtx0123, vsqrtx0123);
68 const float32x4_t vadjustment4567 = vfmsq_f32(vx4567, vsqrtx4567, vsqrtx4567);
69 const float32x4_t vadjustment89AB = vfmsq_f32(vx89AB, vsqrtx89AB, vsqrtx89AB);
70
71 const float32x4_t vy0123 = vfmaq_f32(vsqrtx0123, vhalfrsqrtx0123, vadjustment0123);
72 const float32x4_t vy4567 = vfmaq_f32(vsqrtx4567, vhalfrsqrtx4567, vadjustment4567);
73 const float32x4_t vy89AB = vfmaq_f32(vsqrtx89AB, vhalfrsqrtx89AB, vadjustment89AB);
74
75 vst1q_f32(y, vy0123); y += 4;
76 vst1q_f32(y, vy4567); y += 4;
77 vst1q_f32(y, vy89AB); y += 4;
78 }
79 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
80 const float32x4_t vx = vld1q_f32(x); x += 4;
81 const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
82 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
83 float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
84 float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
85 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
86 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
87 vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
88 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
89 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
90 const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
91 const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);
92 vst1q_f32(y, vy); y += 4;
93 }
94 if XNN_UNLIKELY(n != 0) {
95 const float32x4_t vx = vld1q_f32(x);
96 const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
97 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
98 float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
99 float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
100 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
101 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
102 vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
103 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
104 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
105 const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
106 const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);
107
108 float32x2_t vy_lo = vget_low_f32(vy);
109 if (n & (2 * sizeof(float))) {
110 vst1_f32(y, vy_lo); y += 2;
111 vy_lo = vget_high_f32(vy);
112 }
113 if (n & (1 * sizeof(float))) {
114 vst1_lane_f32(y, vy_lo, 0);
115 }
116 }
117 }
118