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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * 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 x += 32;
34
35 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
36 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
37 const __m256 vrsqrtx2 = _mm256_rsqrt_ps(vx2);
38 const __m256 vrsqrtx3 = _mm256_rsqrt_ps(vx3);
39
40 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
41 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
42 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
43 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
44 __m256 vsqrtx2 = _mm256_mul_ps(vrsqrtx2, vx2);
45 __m256 vhalfrsqrtx2 = _mm256_mul_ps(vrsqrtx2, vhalf);
46 __m256 vsqrtx3 = _mm256_mul_ps(vrsqrtx3, vx3);
47 __m256 vhalfrsqrtx3 = _mm256_mul_ps(vrsqrtx3, vhalf);
48
49 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
50 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
51 const __m256 vresidual2 = _mm256_fnmadd_ps(vsqrtx2, vhalfrsqrtx2, vhalf);
52 const __m256 vresidual3 = _mm256_fnmadd_ps(vsqrtx3, vhalfrsqrtx3, vhalf);
53
54 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
55 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
56 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
57 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
58 vhalfrsqrtx2 = _mm256_fmadd_ps(vhalfrsqrtx2, vresidual2, vhalfrsqrtx2);
59 vsqrtx2 = _mm256_fmadd_ps(vsqrtx2, vresidual2, vsqrtx2);
60 vhalfrsqrtx3 = _mm256_fmadd_ps(vhalfrsqrtx3, vresidual3, vhalfrsqrtx3);
61 vsqrtx3 = _mm256_fmadd_ps(vsqrtx3, vresidual3, vsqrtx3);
62
63 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
64 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
65 const __m256 vadjustment2 = _mm256_fnmadd_ps(vsqrtx2, vsqrtx2, vx2);
66 const __m256 vadjustment3 = _mm256_fnmadd_ps(vsqrtx3, vsqrtx3, vx3);
67
68 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
69 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
70 const __m256 vy2 = _mm256_fmadd_ps(vhalfrsqrtx2, vadjustment2, vsqrtx2);
71 const __m256 vy3 = _mm256_fmadd_ps(vhalfrsqrtx3, vadjustment3, vsqrtx3);
72
73 _mm256_storeu_ps(y, vy0);
74 _mm256_storeu_ps(y + 8, vy1);
75 _mm256_storeu_ps(y + 16, vy2);
76 _mm256_storeu_ps(y + 24, vy3);
77 y += 32;
78 }
79 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
80 const __m256 vx = _mm256_loadu_ps(x);
81 x += 8;
82
83 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
84 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
85 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
86 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
87 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
88 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
89 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
90 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
91
92 _mm256_storeu_ps(y, vy);
93 y += 8;
94 }
95 if XNN_UNLIKELY(n != 0) {
96 assert(n >= 1 * sizeof(float));
97 assert(n <= 7 * sizeof(float));
98 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->fma.mask_table[7] - n));
99
100 const __m256 vx = _mm256_maskload_ps(x, vmask);
101
102 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
103 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
104 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
105 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
106 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
107 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
108 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
109 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
110
111 __m128 vy_lo = _mm256_castps256_ps128(vy);
112 if (n & (4 * sizeof(float))) {
113 _mm_storeu_ps(y, vy_lo);
114 vy_lo = _mm256_extractf128_ps(vy, 1);
115 y += 4;
116 }
117 if (n & (2 * sizeof(float))) {
118 _mm_storel_pi((__m64*) y, vy_lo);
119 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
120 y += 2;
121 }
122 if (n & (1 * sizeof(float))) {
123 _mm_store_ss(y, vy_lo);
124 }
125 }
126 }
127