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
xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48(size_t n,const float * x,float * y,const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsqrt_ukernel__fma3_nr1fma1adj_x48(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const __m256 vhalf = _mm256_load_ps(params->fma.half);
28 for (; n >= 48 * sizeof(float); n -= 48 * sizeof(float)) {
29 const __m256 vx0 = _mm256_loadu_ps(x);
30 const __m256 vx1 = _mm256_loadu_ps(x + 8);
31 const __m256 vx2 = _mm256_loadu_ps(x + 16);
32 const __m256 vx3 = _mm256_loadu_ps(x + 24);
33 const __m256 vx4 = _mm256_loadu_ps(x + 32);
34 const __m256 vx5 = _mm256_loadu_ps(x + 40);
35 x += 48;
36
37 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
38 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
39 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
40 const __m256 vrsqrtx3 = _mm256_rsqrt_ps(vx3);
41 const __m256 vrsqrtx4 = _mm256_rsqrt_ps(vx4);
42 const __m256 vrsqrtx5 = _mm256_rsqrt_ps(vx5);
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 __m256 vsqrtx5 = _mm256_mul_ps(vrsqrtx5, vx5);
55 __m256 vhalfrsqrtx5 = _mm256_mul_ps(vrsqrtx5, vhalf);
56
57 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
58 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
59 const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
60 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
61 const __m256 vresidual4 = _mm256_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);
62 const __m256 vresidual5 = _mm256_fnmadd_ps(vsqrtx5, vhalfrsqrtx5, vhalf);
63
64 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
65 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
66 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
67 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
68 vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
69 vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
70 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
71 vsqrtx3 = _mm256_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
72 vhalfrsqrtx4 = _mm256_fmadd_ps(vhalfrsqrtx4, vresidual4, vhalfrsqrtx4);
73 vsqrtx4 = _mm256_fmadd_ps(vsqrtx4, vresidual4, vsqrtx4);
74 vhalfrsqrtx5 = _mm256_fmadd_ps(vhalfrsqrtx5, vresidual5, vhalfrsqrtx5);
75 vsqrtx5 = _mm256_fmadd_ps(vsqrtx5, vresidual5, vsqrtx5);
76
77 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
78 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
79 const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
80 const __m256 vadjustment3 = _mm256_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
81 const __m256 vadjustment4 = _mm256_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);
82 const __m256 vadjustment5 = _mm256_fnmadd_ps(vsqrtx5, vsqrtx5, vx5);
83
84 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
85 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
86 const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
87 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
88 const __m256 vy4 = _mm256_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);
89 const __m256 vy5 = _mm256_fmadd_ps(vhalfrsqrtx5, vadjustment5, vsqrtx5);
90
91 _mm256_storeu_ps(y, vy0);
92 _mm256_storeu_ps(y + 8, vy1);
93 _mm256_storeu_ps(y + 16, vy2);
94 _mm256_storeu_ps(y + 24, vy3);
95 _mm256_storeu_ps(y + 32, vy4);
96 _mm256_storeu_ps(y + 40, vy5);
97 y += 48;
98 }
99 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
100 const __m256 vx = _mm256_loadu_ps(x);
101 x += 8;
102
103 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
104 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
105 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
106 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
107 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
108 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
109 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
110 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
111
112 _mm256_storeu_ps(y, vy);
113 y += 8;
114 }
115 if XNN_UNLIKELY(n != 0) {
116 assert(n >= 1 * sizeof(float));
117 assert(n <= 7 * sizeof(float));
118 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->fma.mask_table[7] - n));
119
120 const __m256 vx = _mm256_maskload_ps(x, vmask);
121
122 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
123 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
124 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
125 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
126 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
127 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
128 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
129 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
130
131 __m128 vy_lo = _mm256_castps256_ps128(vy);
132 if (n & (4 * sizeof(float))) {
133 _mm_storeu_ps(y, vy_lo);
134 vy_lo = _mm256_extractf128_ps(vy, 1);
135 y += 4;
136 }
137 if (n & (2 * sizeof(float))) {
138 _mm_storel_pi((__m64*) y, vy_lo);
139 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
140 y += 2;
141 }
142 if (n & (1 * sizeof(float))) {
143 _mm_store_ss(y, vy_lo);
144 }
145 }
146 }
147