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_vsqrdiff_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_vsqrdiff_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 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 const __m256 va01234567 = _mm256_loadu_ps(a);
34 const __m256 va89ABCDEF = _mm256_loadu_ps(a + 8);
35 a += 16;
36
37 const __m256 vb01234567 = _mm256_loadu_ps(b);
38 const __m256 vb89ABCDEF = _mm256_loadu_ps(b + 8);
39 b += 16;
40
41 __m256 vy01234567 = _mm256_sub_ps(va01234567, vb01234567);
42 __m256 vy89ABCDEF = _mm256_sub_ps(va89ABCDEF, vb89ABCDEF);
43
44 vy01234567 = _mm256_mul_ps(vy01234567, vy01234567);
45 vy89ABCDEF = _mm256_mul_ps(vy89ABCDEF, vy89ABCDEF);
46
47
48 _mm256_storeu_ps(y, vy01234567);
49 _mm256_storeu_ps(y + 8, vy89ABCDEF);
50 y += 16;
51 }
52 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
53 const __m256 va = _mm256_loadu_ps(a);
54 a += 8;
55
56 const __m256 vb = _mm256_loadu_ps(b);
57 b += 8;
58
59 __m256 vy = _mm256_sub_ps(va, vb);
60 vy = _mm256_mul_ps(vy, vy);
61 _mm256_storeu_ps(y, vy);
62 y += 8;
63 }
64 if XNN_UNLIKELY(n != 0) {
65 assert(n >= 1 * sizeof(float));
66 assert(n <= 7 * sizeof(float));
67 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
68
69 const __m256 va = _mm256_maskload_ps(a, vmask);
70 const __m256 vb = _mm256_maskload_ps(b, vmask);
71
72 __m256 vy = _mm256_sub_ps(va, vb);
73 vy = _mm256_mul_ps(vy, vy);
74
75 __m128 vy_lo = _mm256_castps256_ps128(vy);
76 if (n & (4 * sizeof(float))) {
77 _mm_storeu_ps(y, vy_lo);
78 vy_lo = _mm256_extractf128_ps(vy, 1);
79 y += 4;
80 }
81 if (n & (2 * sizeof(float))) {
82 _mm_storel_pi((__m64*) y, vy_lo);
83 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
84 y += 2;
85 }
86 if (n & (1 * sizeof(float))) {
87 _mm_store_ss(y, vy_lo);
88 }
89 }
90 }
91