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_vmaxc_ukernel__avx_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vmaxc_ukernel__avx_x16(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
23 const union xnn_f32_default_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
32 const __m256 vb = _mm256_broadcast_ss(b);
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 __m256 vy01234567 = _mm256_max_ps(va01234567, vb);
39 __m256 vy89ABCDEF = _mm256_max_ps(va89ABCDEF, vb);
40
41
42
43 _mm256_storeu_ps(y, vy01234567);
44 _mm256_storeu_ps(y + 8, vy89ABCDEF);
45 y += 16;
46 }
47 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
48 const __m256 va = _mm256_loadu_ps(a);
49 a += 8;
50
51 __m256 vy = _mm256_max_ps(va, vb);
52 _mm256_storeu_ps(y, vy);
53 y += 8;
54 }
55 if XNN_UNLIKELY(n != 0) {
56 assert(n >= 1 * sizeof(float));
57 assert(n <= 7 * sizeof(float));
58 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
59
60 const __m256 va = _mm256_maskload_ps(a, vmask);
61
62 __m256 vy = _mm256_max_ps(va, vb);
63
64 __m128 vy_lo = _mm256_castps256_ps128(vy);
65 if (n & (4 * sizeof(float))) {
66 _mm_storeu_ps(y, vy_lo);
67 vy_lo = _mm256_extractf128_ps(vy, 1);
68 y += 4;
69 }
70 if (n & (2 * sizeof(float))) {
71 _mm_storel_pi((__m64*) y, vy_lo);
72 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
73 y += 2;
74 }
75 if (n & (1 * sizeof(float))) {
76 _mm_store_ss(y, vy_lo);
77 }
78 }
79 }
80