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_vaddc_ukernel__avx_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_output_params params[restrict static1])20 void xnn_f32_vaddc_ukernel__avx_x16(
21 size_t n,
22 const float* a,
23 const float* b,
24 float* y,
25 const union xnn_f32_output_params params[restrict static 1])
26 {
27 assert(n != 0);
28 assert(n % sizeof(float) == 0);
29
30 const __m256 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
31 const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
32
33 const __m256 vb = _mm256_broadcast_ss(b);
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 __m256 vy01234567 = _mm256_add_ps(va01234567, vb);
40 __m256 vy89ABCDEF = _mm256_add_ps(va89ABCDEF, vb);
41
42 vy01234567 = _mm256_max_ps(vy01234567, vy_min);
43 vy89ABCDEF = _mm256_max_ps(vy89ABCDEF, vy_min);
44
45 vy01234567 = _mm256_min_ps(vy01234567, vy_max);
46 vy89ABCDEF = _mm256_min_ps(vy89ABCDEF, vy_max);
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 __m256 vy = _mm256_add_ps(va, vb);
57 vy = _mm256_max_ps(vy, vy_min);
58 vy = _mm256_min_ps(vy, vy_max);
59 _mm256_storeu_ps(y, vy);
60 y += 8;
61 }
62 if XNN_UNLIKELY(n != 0) {
63 assert(n >= 1 * sizeof(float));
64 assert(n <= 7 * sizeof(float));
65 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
66
67 const __m256 va = _mm256_maskload_ps(a, vmask);
68
69 __m256 vy = _mm256_add_ps(va, vb);
70 vy = _mm256_max_ps(vy, vy_min);
71 vy = _mm256_min_ps(vy, vy_max);
72
73 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug).
74 __m128 vy_lo = _mm256_castps256_ps128(vy);
75 if (n & (4 * sizeof(float))) {
76 _mm_storeu_ps(y, vy_lo);
77 vy_lo = _mm256_extractf128_ps(vy, 1);
78 y += 4;
79 }
80 if (n & (2 * sizeof(float))) {
81 _mm_storel_pi((__m64*) y, vy_lo);
82 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
83 y += 2;
84 }
85 if (n & (1 * sizeof(float))) {
86 _mm_store_ss(y, vy_lo);
87 }
88 }
89 }
90