1 // Auto-generated file. Do not edit!
2 // Template: src/f32-hswish/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_hswish_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict static1])20 void xnn_f32_hswish_ukernel__avx_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_hswish_params params[restrict static 1])
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28
29 const __m256 vsixth = _mm256_broadcast_ps((const __m128*) params->sse.sixth);
30 const __m256 vhalf = _mm256_broadcast_ps((const __m128*) params->sse.half);
31 const __m256 vone = _mm256_broadcast_ps((const __m128*) params->sse.one);
32 const __m256 vzero = _mm256_setzero_ps();
33
34 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35 const __m256 vx01234567 = _mm256_loadu_ps(x);
36 const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
37 x += 16;
38
39 __m256 vacc01234567 = _mm256_mul_ps(vx01234567, vsixth);
40 __m256 vacc89ABCDEF = _mm256_mul_ps(vx89ABCDEF, vsixth);
41
42 vacc01234567 = _mm256_add_ps(vacc01234567, vhalf);
43 vacc89ABCDEF = _mm256_add_ps(vacc89ABCDEF, vhalf);
44
45 vacc01234567 = _mm256_max_ps(vacc01234567, vzero);
46 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vzero);
47
48 vacc01234567 = _mm256_min_ps(vacc01234567, vone);
49 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vone);
50
51 vacc01234567 = _mm256_mul_ps(vacc01234567, vx01234567);
52 vacc89ABCDEF = _mm256_mul_ps(vacc89ABCDEF, vx89ABCDEF);
53
54 _mm256_storeu_ps(y, vacc01234567);
55 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
56 y += 16;
57 }
58 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
59 const __m256 vx = _mm256_loadu_ps(x);
60 x += 8;
61 __m256 vacc = _mm256_mul_ps(vx, vsixth);
62 vacc = _mm256_add_ps(vacc, vhalf);
63 vacc = _mm256_max_ps(vacc, vzero);
64 vacc = _mm256_min_ps(vacc, vone);
65 vacc = _mm256_mul_ps(vacc, vx);
66 _mm256_storeu_ps(y, vacc);
67 y += 8;
68 }
69 if XNN_UNLIKELY(n != 0) {
70 assert(n >= 1 * sizeof(float));
71 assert(n <= 7 * sizeof(float));
72 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
73
74 const __m256 vx = _mm256_maskload_ps(x, vmask);
75 __m256 vacc = _mm256_mul_ps(vx, vsixth);
76 vacc = _mm256_add_ps(vacc, vhalf);
77 vacc = _mm256_max_ps(vacc, vzero);
78 vacc = _mm256_min_ps(vacc, vone);
79 vacc = _mm256_mul_ps(vacc, vx);
80
81 // _mm256_maskstore_ps(y, vmask, vacc) could be used here, but triggers msan failures (probably an msan bug).
82 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
83 if (n & (4 * sizeof(float))) {
84 _mm_storeu_ps(y, vacc_lo);
85 vacc_lo = _mm256_extractf128_ps(vacc, 1);
86 y += 4;
87 }
88 if (n & (2 * sizeof(float))) {
89 _mm_storel_pi((__m64*) y, vacc_lo);
90 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
91 y += 2;
92 }
93 if (n & (1 * sizeof(float))) {
94 _mm_store_ss(y, vacc_lo);
95 }
96 }
97 }
98