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_x40(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_x40(
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 >= 40 * sizeof(float); n -= 40 * 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 x += 40;
37
38 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
39 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
40 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
41 const __m256 vrsqrtx3 = _mm256_rsqrt_ps(vx3);
42 const __m256 vrsqrtx4 = _mm256_rsqrt_ps(vx4);
43
44 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
45 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
46 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
47 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
48 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2);
49 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf);
50 __m256 vsqrtx3 = _mm256_mul_ps(vrsqrtx3, vx3);
51 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf);
52 __m256 vsqrtx4 = _mm256_mul_ps(vrsqrtx4, vx4);
53 __m256 vhalfrsqrtx4 = _mm256_mul_ps(vrsqrtx4, vhalf);
54
55 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
56 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
57 const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
58 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
59 const __m256 vresidual4 = _mm256_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);
60
61 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
62 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
63 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
64 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
65 vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
66 vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
67 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
68 vsqrtx3 = _mm256_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
69 vhalfrsqrtx4 = _mm256_fmadd_ps(vhalfrsqrtx4, vresidual4, vhalfrsqrtx4);
70 vsqrtx4 = _mm256_fmadd_ps(vsqrtx4, vresidual4, vsqrtx4);
71
72 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
73 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
74 const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
75 const __m256 vadjustment3 = _mm256_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
76 const __m256 vadjustment4 = _mm256_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);
77
78 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
79 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
80 const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
81 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
82 const __m256 vy4 = _mm256_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);
83
84 _mm256_storeu_ps(y, vy0);
85 _mm256_storeu_ps(y + 8, vy1);
86 _mm256_storeu_ps(y + 16, vy2);
87 _mm256_storeu_ps(y + 24, vy3);
88 _mm256_storeu_ps(y + 32, vy4);
89 y += 40;
90 }
91 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
92 const __m256 vx = _mm256_loadu_ps(x);
93 x += 8;
94
95 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
96 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
97 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
98 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
99 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
100 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
101 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
102 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
103
104 _mm256_storeu_ps(y, vy);
105 y += 8;
106 }
107 if XNN_UNLIKELY(n != 0) {
108 assert(n >= 1 * sizeof(float));
109 assert(n <= 7 * sizeof(float));
110 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
111
112 const __m256 vx = _mm256_maskload_ps(x, vmask);
113
114 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
115 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
116 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
117 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
118 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
119 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
120 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
121 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
122
123 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
124 __m128 vy_lo = _mm256_castps256_ps128(vy);
125 if (n & (4 * sizeof(float))) {
126 _mm_storeu_ps(y, vy_lo);
127 vy_lo = _mm256_extractf128_ps(vy, 1);
128 y += 4;
129 }
130 if (n & (2 * sizeof(float))) {
131 _mm_storel_pi((__m64*) y, vy_lo);
132 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
133 y += 2;
134 }
135 if (n & (1 * sizeof(float))) {
136 _mm_store_ss(y, vy_lo);
137 }
138 }
139 }
140