1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsqrt/fma3-nr1fma1adj.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
12 #include <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x64(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28
29 const __m256 vhalf = _mm256_broadcast_ss(¶ms->fma.half);
30 for (; n >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
31 const __m256 vx0 = _mm256_loadu_ps(x);
32 const __m256 vx1 = _mm256_loadu_ps(x + 8);
33 const __m256 vx2 = _mm256_loadu_ps(x + 16);
34 const __m256 vx3 = _mm256_loadu_ps(x + 24);
35 const __m256 vx4 = _mm256_loadu_ps(x + 32);
36 const __m256 vx5 = _mm256_loadu_ps(x + 40);
37 const __m256 vx6 = _mm256_loadu_ps(x + 48);
38 const __m256 vx7 = _mm256_loadu_ps(x + 56);
39 x += 64;
40
41 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
42 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
43 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
44 const __m256 vrsqrtx3 = _mm256_rsqrt_ps(vx3);
45 const __m256 vrsqrtx4 = _mm256_rsqrt_ps(vx4);
46 const __m256 vrsqrtx5 = _mm256_rsqrt_ps(vx5);
47 const __m256 vrsqrtx6 = _mm256_rsqrt_ps(vx6);
48 const __m256 vrsqrtx7 = _mm256_rsqrt_ps(vx7);
49
50 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
51 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
52 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
53 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
54 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2);
55 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf);
56 __m256 vsqrtx3 = _mm256_mul_ps(vrsqrtx3, vx3);
57 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf);
58 __m256 vsqrtx4 = _mm256_mul_ps(vrsqrtx4, vx4);
59 __m256 vhalfrsqrtx4 = _mm256_mul_ps(vrsqrtx4, vhalf);
60 __m256 vsqrtx5 = _mm256_mul_ps(vrsqrtx5, vx5);
61 __m256 vhalfrsqrtx5 = _mm256_mul_ps(vrsqrtx5, vhalf);
62 __m256 vsqrtx6 = _mm256_mul_ps(vrsqrtx6, vx6);
63 __m256 vhalfrsqrtx6 = _mm256_mul_ps(vrsqrtx6, vhalf);
64 __m256 vsqrtx7 = _mm256_mul_ps(vrsqrtx7, vx7);
65 __m256 vhalfrsqrtx7 = _mm256_mul_ps(vrsqrtx7, vhalf);
66
67 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
68 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
69 const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
70 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
71 const __m256 vresidual4 = _mm256_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);
72 const __m256 vresidual5 = _mm256_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf);
73 const __m256 vresidual6 = _mm256_fnmadd_ps(vsqrtx6, vhalfrsqrtx6, vhalf);
74 const __m256 vresidual7 = _mm256_fnmadd_ps(vsqrtx7, vhalfrsqrtx7, vhalf);
75
76 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
77 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
78 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
79 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
80 vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
81 vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
82 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
83 vsqrtx3 = _mm256_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
84 vhalfrsqrtx4 = _mm256_fmadd_ps(vhalfrsqrtx4, vresidual4, vhalfrsqrtx4);
85 vsqrtx4 = _mm256_fmadd_ps(vsqrtx4, vresidual4, vsqrtx4);
86 vhalfrsqrtx5 = _mm256_fmadd_ps(vhalfrsqrtx5, vresidual5, vhalfrsqrtx5);
87 vsqrtx5 = _mm256_fmadd_ps(vsqrtx5, vresidual5, vsqrtx5);
88 vhalfrsqrtx6 = _mm256_fmadd_ps(vhalfrsqrtx6, vresidual6, vhalfrsqrtx6);
89 vsqrtx6 = _mm256_fmadd_ps(vsqrtx6, vresidual6, vsqrtx6);
90 vhalfrsqrtx7 = _mm256_fmadd_ps(vhalfrsqrtx7, vresidual7, vhalfrsqrtx7);
91 vsqrtx7 = _mm256_fmadd_ps(vsqrtx7, vresidual7, vsqrtx7);
92
93 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
94 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
95 const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
96 const __m256 vadjustment3 = _mm256_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
97 const __m256 vadjustment4 = _mm256_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);
98 const __m256 vadjustment5 = _mm256_fnmadd_ps(vsqrtx5, vsqrtx5, vx5);
99 const __m256 vadjustment6 = _mm256_fnmadd_ps(vsqrtx6, vsqrtx6, vx6);
100 const __m256 vadjustment7 = _mm256_fnmadd_ps(vsqrtx7, vsqrtx7, vx7);
101
102 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
103 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
104 const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
105 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
106 const __m256 vy4 = _mm256_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);
107 const __m256 vy5 = _mm256_fmadd_ps(vhalfrsqrtx5, vadjustment5, vsqrtx5);
108 const __m256 vy6 = _mm256_fmadd_ps(vhalfrsqrtx6, vadjustment6, vsqrtx6);
109 const __m256 vy7 = _mm256_fmadd_ps(vhalfrsqrtx7, vadjustment7, vsqrtx7);
110
111 _mm256_storeu_ps(y, vy0);
112 _mm256_storeu_ps(y + 8, vy1);
113 _mm256_storeu_ps(y + 16, vy2);
114 _mm256_storeu_ps(y + 24, vy3);
115 _mm256_storeu_ps(y + 32, vy4);
116 _mm256_storeu_ps(y + 40, vy5);
117 _mm256_storeu_ps(y + 48, vy6);
118 _mm256_storeu_ps(y + 56, vy7);
119 y += 64;
120 }
121 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
122 const __m256 vx = _mm256_loadu_ps(x);
123 x += 8;
124
125 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
126 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
127 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
128 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
129 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
130 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
131 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
132 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
133
134 _mm256_storeu_ps(y, vy);
135 y += 8;
136 }
137 if XNN_UNLIKELY(n != 0) {
138 assert(n >= 1 * sizeof(float));
139 assert(n <= 7 * sizeof(float));
140 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
141
142 const __m256 vx = _mm256_maskload_ps(x, vmask);
143
144 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
145 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
146 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
147 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
148 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
149 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
150 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
151 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
152
153 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
154 __m128 vy_lo = _mm256_castps256_ps128(vy);
155 if (n & (4 * sizeof(float))) {
156 _mm_storeu_ps(y, vy_lo);
157 vy_lo = _mm256_extractf128_ps(vy, 1);
158 y += 4;
159 }
160 if (n & (2 * sizeof(float))) {
161 _mm_storel_pi((__m64*) y, vy_lo);
162 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
163 y += 2;
164 }
165 if (n & (1 * sizeof(float))) {
166 _mm_store_ss(y, vy_lo);
167 }
168 }
169 }
170