1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vclamp/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/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vclamp_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vclamp_ukernel__avx_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m256 vy_min = _mm256_load_ps(params->avx.min);
30 const __m256 vy_max = _mm256_load_ps(params->avx.max);
31
32 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 __m256 vacc01234567 = _mm256_loadu_ps(x);
34 __m256 vacc89ABCDEF = _mm256_loadu_ps(x + 8);
35 x += 16;
36
37 vacc01234567 = _mm256_max_ps(vacc01234567, vy_min);
38 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vy_min);
39
40 vacc01234567 = _mm256_min_ps(vacc01234567, vy_max);
41 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vy_max);
42
43 _mm256_storeu_ps(y, vacc01234567);
44 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
45 y += 16;
46 }
47 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
48 __m256 vacc = _mm256_loadu_ps(x);
49 x += 8;
50
51 vacc = _mm256_max_ps(vacc, vy_min);
52 vacc = _mm256_min_ps(vacc, vy_max);
53
54 _mm256_storeu_ps(y, vacc);
55 y += 8;
56 }
57 if XNN_UNLIKELY(n != 0) {
58 assert(n >= 1 * sizeof(float));
59 assert(n <= 7 * sizeof(float));
60 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n));
61
62 __m256 vacc = _mm256_maskload_ps(x, vmask);
63 vacc = _mm256_max_ps(vacc, vy_min);
64 vacc = _mm256_min_ps(vacc, vy_max);
65
66 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
67 if (n & (4 * sizeof(float))) {
68 _mm_storeu_ps(y, vacc_lo);
69 vacc_lo = _mm256_extractf128_ps(vacc, 1);
70 y += 4;
71 }
72 if (n & (2 * sizeof(float))) {
73 _mm_storel_pi((__m64*) y, vacc_lo);
74 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
75 y += 2;
76 }
77 if (n & (1 * sizeof(float))) {
78 _mm_store_ss(y, vacc_lo);
79 }
80 }
81 }
82