1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrelu/avx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/vunary.h>
15 #include <xnnpack/common.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_vrelu_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vrelu_ukernel__avx_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28 assert(x != NULL);
29 assert(y != NULL);
30
31 const __m256 vzero = _mm256_setzero_ps();
32
33 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
34 __m256 vacc01234567 = _mm256_loadu_ps(x);
35 __m256 vacc89ABCDEF = _mm256_loadu_ps(x + 8);
36 x += 16;
37
38 vacc01234567 = _mm256_max_ps(vacc01234567, vzero);
39 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vzero);
40
41 _mm256_storeu_ps(y, vacc01234567);
42 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
43 y += 16;
44 }
45 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 __m256 vacc = _mm256_loadu_ps(x);
47 x += 8;
48
49 vacc = _mm256_max_ps(vacc, vzero);
50
51 _mm256_storeu_ps(y, vacc);
52 y += 8;
53 }
54 if XNN_UNLIKELY(n != 0) {
55 assert(n >= 1 * sizeof(float));
56 assert(n <= 7 * sizeof(float));
57 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
58
59 __m256 vacc = _mm256_maskload_ps(x, vmask);
60 vacc = _mm256_max_ps(vacc, vzero);
61
62 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
63 if (n & (4 * sizeof(float))) {
64 _mm_storeu_ps(y, vacc_lo);
65 vacc_lo = _mm256_extractf128_ps(vacc, 1);
66 y += 4;
67 }
68 if (n & (2 * sizeof(float))) {
69 _mm_storel_pi((__m64*) y, vacc_lo);
70 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
71 y += 2;
72 }
73 if (n & (1 * sizeof(float))) {
74 _mm_store_ss(y, vacc_lo);
75 }
76 }
77 }
78