1 // Auto-generated file. Do not edit!
2 // Template: src/f32-clamp/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/clamp.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_clamp_ukernel__avx_x16(size_t n,const float * x,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_clamp_ukernel__avx_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_minmax_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 vy_min = _mm256_broadcast_ps((const __m128*) params->sse.min);
32 const __m256 vy_max = _mm256_broadcast_ps((const __m128*) params->sse.max);
33
34 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35 __m256 vacc01234567 = _mm256_loadu_ps(x);
36 __m256 vacc89ABCDEF = _mm256_loadu_ps(x + 8);
37 x += 16;
38
39 vacc01234567 = _mm256_max_ps(vacc01234567, vy_min);
40 vacc89ABCDEF = _mm256_max_ps(vacc89ABCDEF, vy_min);
41
42 vacc01234567 = _mm256_min_ps(vacc01234567, vy_max);
43 vacc89ABCDEF = _mm256_min_ps(vacc89ABCDEF, vy_max);
44
45 _mm256_storeu_ps(y, vacc01234567);
46 _mm256_storeu_ps(y + 8, vacc89ABCDEF);
47 y += 16;
48 }
49 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
50 __m256 vacc = _mm256_loadu_ps(x);
51 x += 8;
52
53 vacc = _mm256_max_ps(vacc, vy_min);
54 vacc = _mm256_min_ps(vacc, vy_max);
55
56 _mm256_storeu_ps(y, vacc);
57 y += 8;
58 }
59 if XNN_UNLIKELY(n != 0) {
60 assert(n >= 1 * sizeof(float));
61 assert(n <= 7 * sizeof(float));
62 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
63
64 __m256 vacc = _mm256_maskload_ps(x, vmask);
65 vacc = _mm256_max_ps(vacc, vy_min);
66 vacc = _mm256_min_ps(vacc, vy_max);
67
68 // _mm256_maskstore_ps(y, vmask, vacc) could be used here, but triggers msan failures (probably an msan bug).
69 __m128 vacc_lo = _mm256_castps256_ps128(vacc);
70 if (n & (4 * sizeof(float))) {
71 _mm_storeu_ps(y, vacc_lo);
72 vacc_lo = _mm256_extractf128_ps(vacc, 1);
73 y += 4;
74 }
75 if (n & (2 * sizeof(float))) {
76 _mm_storel_pi((__m64*) y, vacc_lo);
77 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo);
78 y += 2;
79 }
80 if (n & (1 * sizeof(float))) {
81 _mm_store_ss(y, vacc_lo);
82 }
83 }
84 }
85