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_vmul_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)])20 void xnn_f32_vmul_minmax_ukernel__avx_x16(
21 size_t n,
22 const float* a,
23 const float* b,
24 float* y,
25 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
26 {
27 assert(n != 0);
28 assert(n % sizeof(float) == 0);
29 assert(a != NULL);
30 assert(b != NULL);
31 assert(y != NULL);
32
33 const __m256 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
34 const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
35
36 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
37 const __m256 va01234567 = _mm256_loadu_ps(a);
38 const __m256 va89ABCDEF = _mm256_loadu_ps(a + 8);
39 a += 16;
40
41 const __m256 vb01234567 = _mm256_loadu_ps(b);
42 const __m256 vb89ABCDEF = _mm256_loadu_ps(b + 8);
43 b += 16;
44
45 __m256 vy01234567 = _mm256_mul_ps(va01234567, vb01234567);
46 __m256 vy89ABCDEF = _mm256_mul_ps(va89ABCDEF, vb89ABCDEF);
47
48
49 vy01234567 = _mm256_max_ps(vy01234567, vy_min);
50 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min);
51
52 vy01234567 = _mm256_min_ps(vy01234567, vy_max);
53 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max);
54
55 _mm256_storeu_ps(y, vy01234567);
56 _mm256_storeu_ps(y + 8, vy89ABCDEF);
57 y += 16;
58 }
59 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
60 const __m256 va = _mm256_loadu_ps(a);
61 a += 8;
62
63 const __m256 vb = _mm256_loadu_ps(b);
64 b += 8;
65
66 __m256 vy = _mm256_mul_ps(va, vb);
67 vy = _mm256_max_ps(vy, vy_min);
68 vy = _mm256_min_ps(vy, vy_max);
69 _mm256_storeu_ps(y, vy);
70 y += 8;
71 }
72 if XNN_UNLIKELY(n != 0) {
73 assert(n >= 1 * sizeof(float));
74 assert(n <= 7 * sizeof(float));
75 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
76
77 const __m256 va = _mm256_maskload_ps(a, vmask);
78 const __m256 vb = _mm256_maskload_ps(b, vmask);
79
80 __m256 vy = _mm256_mul_ps(va, vb);
81 vy = _mm256_max_ps(vy, vy_min);
82 vy = _mm256_min_ps(vy, vy_max);
83
84 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
85 __m128 vy_lo = _mm256_castps256_ps128(vy);
86 if (n & (4 * sizeof(float))) {
87 _mm_storeu_ps(y, vy_lo);
88 vy_lo = _mm256_extractf128_ps(vy, 1);
89 y += 4;
90 }
91 if (n & (2 * sizeof(float))) {
92 _mm_storel_pi((__m64*) y, vy_lo);
93 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
94 y += 2;
95 }
96 if (n & (1 * sizeof(float))) {
97 _mm_store_ss(y, vy_lo);
98 }
99 }
100 }
101