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
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_vsqrdiffc_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)])20 void xnn_f32_vsqrdiffc_ukernel__avx_x16(
21 size_t n,
22 const float* a,
23 const float* b,
24 float* y,
25 const union xnn_f32_default_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
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_sub_ps(va01234567, vb);
41 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb);
42
43 vy01234567 = _mm256_mul_ps(vy01234567, vy01234567);
44 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vy89ABCDEF);
45
46
47 _mm256_storeu_ps(y, vy01234567);
48 _mm256_storeu_ps(y + 8, vy89ABCDEF);
49 y += 16;
50 }
51 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
52 const __m256 va = _mm256_loadu_ps(a);
53 a += 8;
54
55 __m256 vy = _mm256_sub_ps(va, vb);
56 vy = _mm256_mul_ps(vy, vy);
57 _mm256_storeu_ps(y, vy);
58 y += 8;
59 }
60 if XNN_UNLIKELY(n != 0) {
61 assert(n >= 1 * sizeof(float));
62 assert(n <= 7 * sizeof(float));
63 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
64
65 const __m256 va = _mm256_maskload_ps(a, vmask);
66
67 __m256 vy = _mm256_sub_ps(va, vb);
68 vy = _mm256_mul_ps(vy, vy);
69
70 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
71 __m128 vy_lo = _mm256_castps256_ps128(vy);
72 if (n & (4 * sizeof(float))) {
73 _mm_storeu_ps(y, vy_lo);
74 vy_lo = _mm256_extractf128_ps(vy, 1);
75 y += 4;
76 }
77 if (n & (2 * sizeof(float))) {
78 _mm_storel_pi((__m64*) y, vy_lo);
79 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
80 y += 2;
81 }
82 if (n & (1 * sizeof(float))) {
83 _mm_store_ss(y, vy_lo);
84 }
85 }
86 }
87