1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vop-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
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_vmax_ukernel__avx_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_output_params params[restrict static1])20 void xnn_f32_vmax_ukernel__avx_x16(
21 size_t n,
22 const float* a,
23 const float* b,
24 float* y,
25 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(n != 0);
28 assert(n % sizeof(float) == 0);
29
30 const __m256 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
31 const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
32
33 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
34 const __m256 va01234567 = _mm256_loadu_ps(a);
35 const __m256 va89ABCDEF = _mm256_loadu_ps(a + 8);
36 a += 16;
37
38 const __m256 vb01234567 = _mm256_loadu_ps(b);
39 const __m256 vb89ABCDEF = _mm256_loadu_ps(b + 8);
40 b += 16;
41
42 __m256 vy01234567 = _mm256_max_ps(va01234567, vb01234567);
43 __m256 vy89ABCDEF = _mm256_max_ps(va89ABCDEF, vb89ABCDEF);
44
45 vy01234567 = _mm256_max_ps(vy01234567, vy_min);
46 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min);
47
48 vy01234567 = _mm256_min_ps(vy01234567, vy_max);
49 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max);
50
51 _mm256_storeu_ps(y, vy01234567);
52 _mm256_storeu_ps(y + 8, vy89ABCDEF);
53 y += 16;
54 }
55 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
56 const __m256 va = _mm256_loadu_ps(a);
57 a += 8;
58
59 const __m256 vb = _mm256_loadu_ps(b);
60 b += 8;
61
62 __m256 vy = _mm256_max_ps(va, vb);
63 vy = _mm256_max_ps(vy, vy_min);
64 vy = _mm256_min_ps(vy, vy_max);
65 _mm256_storeu_ps(y, vy);
66 y += 8;
67 }
68 if XNN_UNLIKELY(n != 0) {
69 assert(n >= 1 * sizeof(float));
70 assert(n <= 7 * sizeof(float));
71 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
72
73 const __m256 va = _mm256_maskload_ps(a, vmask);
74 const __m256 vb = _mm256_maskload_ps(b, vmask);
75
76 __m256 vy = _mm256_max_ps(va, vb);
77 vy = _mm256_max_ps(vy, vy_min);
78 vy = _mm256_min_ps(vy, vy_max);
79
80 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
81 __m128 vy_lo = _mm256_castps256_ps128(vy);
82 if (n & (4 * sizeof(float))) {
83 _mm_storeu_ps(y, vy_lo);
84 vy_lo = _mm256_extractf128_ps(vy, 1);
85 y += 4;
86 }
87 if (n & (2 * sizeof(float))) {
88 _mm_storel_pi((__m64*) y, vy_lo);
89 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
90 y += 2;
91 }
92 if (n & (1 * sizeof(float))) {
93 _mm_store_ss(y, vy_lo);
94 }
95 }
96 }
97