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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
30 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
31 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
32
33 const float32x4_t vrsqrtx0123 = vrsqrteq_f32(vx0123);
34 const float32x4_t vrsqrtx4567 = vrsqrteq_f32(vx4567);
35
36 float32x4_t vsqrtx0123 = vmulq_f32(vrsqrtx0123, vx0123);
37 float32x4_t vhalfrsqrtx0123 = vmulq_f32(vrsqrtx0123, vhalf);
38 float32x4_t vsqrtx4567 = vmulq_f32(vrsqrtx4567, vx4567);
39 float32x4_t vhalfrsqrtx4567 = vmulq_f32(vrsqrtx4567, vhalf);
40
41 float32x4_t vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
42 float32x4_t vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
43
44 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
45 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
46 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
47 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
48
49 vresidual0123 = vfmsq_f32(vhalf, vsqrtx0123, vhalfrsqrtx0123);
50 vresidual4567 = vfmsq_f32(vhalf, vsqrtx4567, vhalfrsqrtx4567);
51
52 vhalfrsqrtx0123 = vfmaq_f32(vhalfrsqrtx0123, vresidual0123, vhalfrsqrtx0123);
53 vsqrtx0123 = vfmaq_f32(vsqrtx0123, vresidual0123, vsqrtx0123);
54 vhalfrsqrtx4567 = vfmaq_f32(vhalfrsqrtx4567, vresidual4567, vhalfrsqrtx4567);
55 vsqrtx4567 = vfmaq_f32(vsqrtx4567, vresidual4567, vsqrtx4567);
56
57 const float32x4_t vadjustment0123 = vfmsq_f32(vx0123, vsqrtx0123, vsqrtx0123);
58 const float32x4_t vadjustment4567 = vfmsq_f32(vx4567, vsqrtx4567, vsqrtx4567);
59
60 const float32x4_t vy0123 = vfmaq_f32(vsqrtx0123, vhalfrsqrtx0123, vadjustment0123);
61 const float32x4_t vy4567 = vfmaq_f32(vsqrtx4567, vhalfrsqrtx4567, vadjustment4567);
62
63 vst1q_f32(y, vy0123); y += 4;
64 vst1q_f32(y, vy4567); y += 4;
65 }
66 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
67 const float32x4_t vx = vld1q_f32(x); x += 4;
68 const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
69 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
70 float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
71 float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
72 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
73 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
74 vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
75 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
76 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
77 const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
78 const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);
79 vst1q_f32(y, vy); y += 4;
80 }
81 if XNN_UNLIKELY(n != 0) {
82 const float32x4_t vx = vld1q_f32(x);
83 const float32x4_t vrsqrtx = vrsqrteq_f32(vx);
84 float32x4_t vsqrtx = vmulq_f32(vrsqrtx, vx);
85 float32x4_t vhalfrsqrtx = vmulq_f32(vrsqrtx, vhalf);
86 float32x4_t vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
87 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
88 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
89 vresidual = vfmsq_f32(vhalf, vsqrtx, vhalfrsqrtx);
90 vhalfrsqrtx = vfmaq_f32(vhalfrsqrtx, vresidual, vhalfrsqrtx);
91 vsqrtx = vfmaq_f32(vsqrtx, vresidual, vsqrtx);
92 const float32x4_t vadjustment = vfmsq_f32(vx, vsqrtx, vsqrtx);
93 const float32x4_t vy = vfmaq_f32(vsqrtx, vhalfrsqrtx, vadjustment);
94
95 float32x2_t vy_lo = vget_low_f32(vy);
96 if (n & (2 * sizeof(float))) {
97 vst1_f32(y, vy_lo); y += 2;
98 vy_lo = vget_high_f32(vy);
99 }
100 if (n & (1 * sizeof(float))) {
101 vst1_lane_f32(y, vy_lo, 0);
102 }
103 }
104 }
105