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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
31 const __m256 vx0 = _mm256_loadu_ps(x);
32 const __m256 vx1 = _mm256_loadu_ps(x + 8);
33 x += 16;
34
35 const __m256 vrsqrtx0 = _mm256_rsqrt_ps(vx0);
36 const __m256 vrsqrtx1 = _mm256_rsqrt_ps(vx1);
37
38 __m256 vsqrtx0 = _mm256_mul_ps(vrsqrtx0, vx0);
39 __m256 vhalfrsqrtx0 = _mm256_mul_ps(vrsqrtx0, vhalf);
40 __m256 vsqrtx1 = _mm256_mul_ps(vrsqrtx1, vx1);
41 __m256 vhalfrsqrtx1 = _mm256_mul_ps(vrsqrtx1, vhalf);
42
43 const __m256 vresidual0 = _mm256_fnmadd_ps(vsqrtx0, vhalfrsqrtx0, vhalf);
44 const __m256 vresidual1 = _mm256_fnmadd_ps(vsqrtx1, vhalfrsqrtx1, vhalf);
45
46 vhalfrsqrtx0 = _mm256_fmadd_ps(vhalfrsqrtx0, vresidual0, vhalfrsqrtx0);
47 vsqrtx0 = _mm256_fmadd_ps(vsqrtx0, vresidual0, vsqrtx0);
48 vhalfrsqrtx1 = _mm256_fmadd_ps(vhalfrsqrtx1, vresidual1, vhalfrsqrtx1);
49 vsqrtx1 = _mm256_fmadd_ps(vsqrtx1, vresidual1, vsqrtx1);
50
51 const __m256 vadjustment0 = _mm256_fnmadd_ps(vsqrtx0, vsqrtx0, vx0);
52 const __m256 vadjustment1 = _mm256_fnmadd_ps(vsqrtx1, vsqrtx1, vx1);
53
54 const __m256 vy0 = _mm256_fmadd_ps(vhalfrsqrtx0, vadjustment0, vsqrtx0);
55 const __m256 vy1 = _mm256_fmadd_ps(vhalfrsqrtx1, vadjustment1, vsqrtx1);
56
57 _mm256_storeu_ps(y, vy0);
58 _mm256_storeu_ps(y + 8, vy1);
59 y += 16;
60 }
61 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
62 const __m256 vx = _mm256_loadu_ps(x);
63 x += 8;
64
65 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
66 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
67 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
68 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
69 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
70 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
71 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
72 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
73
74 _mm256_storeu_ps(y, vy);
75 y += 8;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 assert(n >= 1 * sizeof(float));
79 assert(n <= 7 * sizeof(float));
80 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
81
82 const __m256 vx = _mm256_maskload_ps(x, vmask);
83
84 const __m256 vrsqrtx = _mm256_rsqrt_ps(vx);
85 __m256 vsqrtx = _mm256_mul_ps(vrsqrtx, vx);
86 __m256 vhalfrsqrtx = _mm256_mul_ps(vrsqrtx, vhalf);
87 const __m256 vresidual = _mm256_fnmadd_ps(vsqrtx, vhalfrsqrtx, vhalf);
88 vhalfrsqrtx = _mm256_fmadd_ps(vhalfrsqrtx, vresidual, vhalfrsqrtx);
89 vsqrtx = _mm256_fmadd_ps(vsqrtx, vresidual, vsqrtx);
90 const __m256 vadjustment = _mm256_fnmadd_ps(vsqrtx, vsqrtx, vx);
91 const __m256 vy = _mm256_fmadd_ps(vhalfrsqrtx, vadjustment, vsqrtx);
92
93 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
94 __m128 vy_lo = _mm256_castps256_ps128(vy);
95 if (n & (4 * sizeof(float))) {
96 _mm_storeu_ps(y, vy_lo);
97 vy_lo = _mm256_extractf128_ps(vy, 1);
98 y += 4;
99 }
100 if (n & (2 * sizeof(float))) {
101 _mm_storel_pi((__m64*) y, vy_lo);
102 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
103 y += 2;
104 }
105 if (n & (1 * sizeof(float))) {
106 _mm_store_ss(y, vy_lo);
107 }
108 }
109 }
110