1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vopc-avx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/vbinary.h>
16
17
xnn_f32_vrdivc_minmax_ukernel__avx_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vrdivc_minmax_ukernel__avx_x16(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
23 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(a != NULL);
28 assert(b != NULL);
29 assert(y != NULL);
30
31 const __m256 vy_min = _mm256_load_ps(params->avx.min);
32 const __m256 vy_max = _mm256_load_ps(params->avx.max);
33
34 const __m256 vb = _mm256_broadcast_ss(b);
35 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
36 const __m256 va01234567 = _mm256_loadu_ps(a);
37 const __m256 va89ABCDEF = _mm256_loadu_ps(a + 8);
38 a += 16;
39
40 __m256 vy01234567 = _mm256_div_ps(vb, va01234567);
41 __m256 vy89ABCDEF = _mm256_div_ps(vb, va89ABCDEF);
42
43
44 vy01234567 = _mm256_max_ps(vy01234567, vy_min);
45 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min);
46
47 vy01234567 = _mm256_min_ps(vy01234567, vy_max);
48 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max);
49
50 _mm256_storeu_ps(y, vy01234567);
51 _mm256_storeu_ps(y + 8, vy89ABCDEF);
52 y += 16;
53 }
54 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
55 const __m256 va = _mm256_loadu_ps(a);
56 a += 8;
57
58 __m256 vy = _mm256_div_ps(vb, va);
59 vy = _mm256_max_ps(vy, vy_min);
60 vy = _mm256_min_ps(vy, vy_max);
61 _mm256_storeu_ps(y, vy);
62 y += 8;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 assert(n >= 1 * sizeof(float));
66 assert(n <= 7 * sizeof(float));
67 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
68
69 const __m256 va = _mm256_maskload_ps(a, vmask);
70
71 __m256 vy = _mm256_div_ps(vb, va);
72 vy = _mm256_max_ps(vy, vy_min);
73 vy = _mm256_min_ps(vy, vy_max);
74
75 __m128 vy_lo = _mm256_castps256_ps128(vy);
76 if (n & (4 * sizeof(float))) {
77 _mm_storeu_ps(y, vy_lo);
78 vy_lo = _mm256_extractf128_ps(vy, 1);
79 y += 4;
80 }
81 if (n & (2 * sizeof(float))) {
82 _mm_storel_pi((__m64*) y, vy_lo);
83 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
84 y += 2;
85 }
86 if (n & (1 * sizeof(float))) {
87 _mm_store_ss(y, vy_lo);
88 }
89 }
90 }
91