1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vhswish/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/vunary.h>
16
17
xnn_f32_vhswish_ukernel__fma3_x16(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vhswish_ukernel__fma3_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const __m256 vsixth = _mm256_load_ps(params->avx.sixth);
28 const __m256 vhalf = _mm256_load_ps(params->avx.half);
29 const __m256 vone = _mm256_load_ps(params->avx.one);
30 const __m256 vzero = _mm256_setzero_ps();
31
32 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 const __m256 vx01234567 = _mm256_loadu_ps(x);
34 const __m256 vx89ABCDEF = _mm256_loadu_ps(x + 8);
35 x += 16;
36
37 __m256 vacc01234567 = _mm256_fmadd_ps(vx01234567, vsixth, vhalf);
38 __m256 vacc89ABCDEF = _mm256_fmadd_ps(vx89ABCDEF, vsixth, vhalf);
39
40 vacc01234567 = _mm256_max_ps(vacc01234567, vzero);
41 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vzero);
42
43 vacc01234567 = _mm256_min_ps(vacc01234567, vone);
44 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vone);
45
46 vacc01234567 = _mm256_mul_ps(vacc01234567, vx01234567);
47 vacc89ABCDEF = _mm256_mul_ps(vacc89ABCDEF, vx89ABCDEF);
48
49 _mm256_storeu_ps(y, vacc01234567);
50 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
51 y += 16;
52 }
53 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
54 const __m256 vx = _mm256_loadu_ps(x);
55 x += 8;
56 __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
57 vacc = _mm256_max_ps(vacc, vzero);
58 vacc = _mm256_min_ps(vacc, vone);
59 vacc = _mm256_mul_ps(vacc, vx);
60 _mm256_storeu_ps(y, vacc);
61 y += 8;
62 }
63 if XNN_UNLIKELY(n != 0) {
64 assert(n >= 1 * sizeof(float));
65 assert(n <= 7 * sizeof(float));
66 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
67
68 const __m256 vx = _mm256_maskload_ps(x, vmask);
69 __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf);
70 vacc = _mm256_max_ps(vacc, vzero);
71 vacc = _mm256_min_ps(vacc, vone);
72 vacc = _mm256_mul_ps(vacc, vx);
73
74 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
75 if (n & (4 * sizeof(float))) {
76 _mm_storeu_ps(y, vacc_lo);
77 vacc_lo = _mm256_extractf128_ps(vacc, 1);
78 y += 4;
79 }
80 if (n & (2 * sizeof(float))) {
81 _mm_storel_pi((__m64*) y, vacc_lo);
82 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
83 y += 2;
84 }
85 if (n & (1 * sizeof(float))) {
86 _mm_store_ss(y, vacc_lo);
87 }
88 }
89 }
90