blob: 6515daaa2b5e047ffd178bf81e78fd923f682c74 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
|
// Copyright (c) the JPEG XL Project Authors. All rights reserved.
//
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
#include "lib/jxl/simd_util.h"
#include <algorithm>
#include <cstddef>
#undef HWY_TARGET_INCLUDE
#define HWY_TARGET_INCLUDE "lib/jxl/simd_util.cc"
#include <hwy/foreach_target.h>
#include <hwy/highway.h>
#include "lib/jxl/base/common.h"
#include "lib/jxl/base/status.h"
#include "lib/jxl/cache_aligned.h"
HWY_BEFORE_NAMESPACE();
namespace jxl {
namespace HWY_NAMESPACE {
size_t MaxVectorSize() {
HWY_FULL(float) df;
return Lanes(df) * sizeof(float);
}
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace jxl
HWY_AFTER_NAMESPACE();
#if HWY_ONCE
namespace jxl {
HWY_EXPORT(MaxVectorSize);
size_t MaxVectorSize() {
// Ideally HWY framework should provide us this value.
// Less than ideal is to check all available targets and choose maximal.
// As for now, we just ask current active target, assuming it won't change.
return HWY_DYNAMIC_DISPATCH(MaxVectorSize)();
}
size_t BytesPerRow(const size_t xsize, const size_t sizeof_t) {
// Special case: we don't allow any ops -> don't need extra padding/
if (xsize == 0) {
return 0;
}
const size_t vec_size = MaxVectorSize();
size_t valid_bytes = xsize * sizeof_t;
// Allow unaligned accesses starting at the last valid value.
// Skip for the scalar case because no extra lanes will be loaded.
if (vec_size != 0) {
valid_bytes += vec_size - sizeof_t;
}
// Round up to vector and cache line size.
const size_t align = std::max(vec_size, CacheAligned::kAlignment);
size_t bytes_per_row = RoundUpTo(valid_bytes, align);
// During the lengthy window before writes are committed to memory, CPUs
// guard against read after write hazards by checking the address, but
// only the lower 11 bits. We avoid a false dependency between writes to
// consecutive rows by ensuring their sizes are not multiples of 2 KiB.
// Avoid2K prevents the same problem for the planes of an Image3.
if (bytes_per_row % CacheAligned::kAlias == 0) {
bytes_per_row += align;
}
JXL_ASSERT(bytes_per_row % align == 0);
return bytes_per_row;
}
} // namespace jxl
#endif
|
