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
xnn_f32_vsub_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_vsub_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 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35 const __m256 va01234567 = _mm256_loadu_ps(a);
36 const __m256 va89ABCDEF = _mm256_loadu_ps(a + 8);
37 a += 16;
38
39 const __m256 vb01234567 = _mm256_loadu_ps(b);
40 const __m256 vb89ABCDEF = _mm256_loadu_ps(b + 8);
41 b += 16;
42
43 __m256 vy01234567 = _mm256_sub_ps(va01234567, vb01234567);
44 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb89ABCDEF);
45
46
47 vy01234567 = _mm256_max_ps(vy01234567, vy_min);
48 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min);
49
50 vy01234567 = _mm256_min_ps(vy01234567, vy_max);
51 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max);
52
53 _mm256_storeu_ps(y, vy01234567);
54 _mm256_storeu_ps(y + 8, vy89ABCDEF);
55 y += 16;
56 }
57 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
58 const __m256 va = _mm256_loadu_ps(a);
59 a += 8;
60
61 const __m256 vb = _mm256_loadu_ps(b);
62 b += 8;
63
64 __m256 vy = _mm256_sub_ps(va, vb);
65 vy = _mm256_max_ps(vy, vy_min);
66 vy = _mm256_min_ps(vy, vy_max);
67 _mm256_storeu_ps(y, vy);
68 y += 8;
69 }
70 if XNN_UNLIKELY(n != 0) {
71 assert(n >= 1 * sizeof(float));
72 assert(n <= 7 * sizeof(float));
73 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
74
75 const __m256 va = _mm256_maskload_ps(a, vmask);
76 const __m256 vb = _mm256_maskload_ps(b, vmask);
77
78 __m256 vy = _mm256_sub_ps(va, vb);
79 vy = _mm256_max_ps(vy, vy_min);
80 vy = _mm256_min_ps(vy, vy_max);
81
82 __m128 vy_lo = _mm256_castps256_ps128(vy);
83 if (n & (4 * sizeof(float))) {
84 _mm_storeu_ps(y, vy_lo);
85 vy_lo = _mm256_extractf128_ps(vy, 1);
86 y += 4;
87 }
88 if (n & (2 * sizeof(float))) {
89 _mm_storel_pi((__m64*) y, vy_lo);
90 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
91 y += 2;
92 }
93 if (n & (1 * sizeof(float))) {
94 _mm_store_ss(y, vy_lo);
95 }
96 }
97 }
98