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_x40(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_x40(
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 >= 40 * sizeof(float); n -= 40 * 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 x += 40;
35
36 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
37 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
38 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
39 const __m256 vrsqrtx3 = _mm256_rsqrt_ps(vx3);
40 const __m256 vrsqrtx4 = _mm256_rsqrt_ps(vx4);
41
42 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
43 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
44 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
45 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
46 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2);
47 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf);
48 __m256 vsqrtx3 = _mm256_mul_ps(vrsqrtx3, vx3);
49 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf);
50 __m256 vsqrtx4 = _mm256_mul_ps(vrsqrtx4, vx4);
51 __m256 vhalfrsqrtx4 = _mm256_mul_ps(vrsqrtx4, vhalf);
52
53 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
54 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
55 const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
56 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
57 const __m256 vresidual4 = _mm256_fnmadd_ps(vsqrtx4, vhalfrsqrtx4, vhalf);
58
59 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
60 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
61 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
62 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
63 vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
64 vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
65 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
66 vsqrtx3 = _mm256_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
67 vhalfrsqrtx4 = _mm256_fmadd_ps(vhalfrsqrtx4, vresidual4, vhalfrsqrtx4);
68 vsqrtx4 = _mm256_fmadd_ps(vsqrtx4, vresidual4, vsqrtx4);
69
70 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
71 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
72 const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
73 const __m256 vadjustment3 = _mm256_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
74 const __m256 vadjustment4 = _mm256_fnmadd_ps(vsqrtx4, vsqrtx4, vx4);
75
76 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
77 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
78 const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
79 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
80 const __m256 vy4 = _mm256_fmadd_ps(vhalfrsqrtx4, vadjustment4, vsqrtx4);
81
82 _mm256_storeu_ps(y, vy0);
83 _mm256_storeu_ps(y + 8, vy1);
84 _mm256_storeu_ps(y + 16, vy2);
85 _mm256_storeu_ps(y + 24, vy3);
86 _mm256_storeu_ps(y + 32, vy4);
87 y += 40;
88 }
89 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
90 const __m256 vx = _mm256_loadu_ps(x);
91 x += 8;
92
93 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
94 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
95 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
96 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
97 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
98 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
99 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
100 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
101
102 _mm256_storeu_ps(y, vy);
103 y += 8;
104 }
105 if XNN_UNLIKELY(n != 0) {
106 assert(n >= 1 * sizeof(float));
107 assert(n <= 7 * sizeof(float));
108 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->fma.mask_table[7] - n));
109
110 const __m256 vx = _mm256_maskload_ps(x, vmask);
111
112 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
113 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
114 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
115 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
116 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
117 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
118 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
119 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
120
121 __m128 vy_lo = _mm256_castps256_ps128(vy);
122 if (n & (4 * sizeof(float))) {
123 _mm_storeu_ps(y, vy_lo);
124 vy_lo = _mm256_extractf128_ps(vy, 1);
125 y += 4;
126 }
127 if (n & (2 * sizeof(float))) {
128 _mm_storel_pi((__m64*) y, vy_lo);
129 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
130 y += 2;
131 }
132 if (n & (1 * sizeof(float))) {
133 _mm_store_ss(y, vy_lo);
134 }
135 }
136 }
137