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
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
|
// 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.
#if defined(LIB_JXL_FAST_DCT_INL_H_) == defined(HWY_TARGET_TOGGLE)
#ifdef LIB_JXL_FAST_DCT_INL_H_
#undef LIB_JXL_FAST_DCT_INL_H_
#else
#define LIB_JXL_FAST_DCT_INL_H_
#endif
#include <cmath>
#include <hwy/aligned_allocator.h>
#include <hwy/highway.h>
#include "lib/jxl/base/status.h"
HWY_BEFORE_NAMESPACE();
namespace jxl {
namespace HWY_NAMESPACE {
namespace {
#if HWY_TARGET == HWY_NEON
HWY_NOINLINE void FastTransposeBlock(const int16_t* JXL_RESTRICT data_in,
size_t stride_in, size_t N, size_t M,
int16_t* JXL_RESTRICT data_out,
size_t stride_out) {
JXL_DASSERT(N % 8 == 0);
JXL_DASSERT(M % 8 == 0);
for (size_t i = 0; i < N; i += 8) {
for (size_t j = 0; j < M; j += 8) {
// TODO(veluca): one could optimize the M==8, stride_in==8 case further
// with vld4.
// This code is about 40% faster for N == M == stride_in ==
// stride_out == 8
// Using loads + stores to reshuffle things to be able to
// use vld4 doesn't help.
/*
auto a0 = vld4q_s16(data_in); auto a1 = vld4q_s16(data_in + 32);
int16x8x4_t out0;
int16x8x4_t out1;
out0.val[0] = vuzp1q_s16(a0.val[0], a1.val[0]);
out0.val[1] = vuzp1q_s16(a0.val[1], a1.val[1]);
out0.val[2] = vuzp1q_s16(a0.val[2], a1.val[2]);
out0.val[3] = vuzp1q_s16(a0.val[3], a1.val[3]);
out1.val[0] = vuzp2q_s16(a0.val[0], a1.val[0]);
out1.val[1] = vuzp2q_s16(a0.val[1], a1.val[1]);
out1.val[2] = vuzp2q_s16(a0.val[2], a1.val[2]);
out1.val[3] = vuzp2q_s16(a0.val[3], a1.val[3]);
vst1q_s16_x4(data_out, out0);
vst1q_s16_x4(data_out + 32, out1);
*/
auto a0 = vld1q_s16(data_in + i * stride_in + j);
auto a1 = vld1q_s16(data_in + (i + 1) * stride_in + j);
auto a2 = vld1q_s16(data_in + (i + 2) * stride_in + j);
auto a3 = vld1q_s16(data_in + (i + 3) * stride_in + j);
auto a01 = vtrnq_s16(a0, a1);
auto a23 = vtrnq_s16(a2, a3);
auto four0 = vtrnq_s32(vreinterpretq_s32_s16(a01.val[0]),
vreinterpretq_s32_s16(a23.val[0]));
auto four1 = vtrnq_s32(vreinterpretq_s32_s16(a01.val[1]),
vreinterpretq_s32_s16(a23.val[1]));
auto a4 = vld1q_s16(data_in + (i + 4) * stride_in + j);
auto a5 = vld1q_s16(data_in + (i + 5) * stride_in + j);
auto a6 = vld1q_s16(data_in + (i + 6) * stride_in + j);
auto a7 = vld1q_s16(data_in + (i + 7) * stride_in + j);
auto a45 = vtrnq_s16(a4, a5);
auto a67 = vtrnq_s16(a6, a7);
auto four2 = vtrnq_s32(vreinterpretq_s32_s16(a45.val[0]),
vreinterpretq_s32_s16(a67.val[0]));
auto four3 = vtrnq_s32(vreinterpretq_s32_s16(a45.val[1]),
vreinterpretq_s32_s16(a67.val[1]));
auto out0 =
vcombine_s32(vget_low_s32(four0.val[0]), vget_low_s32(four2.val[0]));
auto out1 =
vcombine_s32(vget_low_s32(four1.val[0]), vget_low_s32(four3.val[0]));
auto out2 =
vcombine_s32(vget_low_s32(four0.val[1]), vget_low_s32(four2.val[1]));
auto out3 =
vcombine_s32(vget_low_s32(four1.val[1]), vget_low_s32(four3.val[1]));
auto out4 = vcombine_s32(vget_high_s32(four0.val[0]),
vget_high_s32(four2.val[0]));
auto out5 = vcombine_s32(vget_high_s32(four1.val[0]),
vget_high_s32(four3.val[0]));
auto out6 = vcombine_s32(vget_high_s32(four0.val[1]),
vget_high_s32(four2.val[1]));
auto out7 = vcombine_s32(vget_high_s32(four1.val[1]),
vget_high_s32(four3.val[1]));
vst1q_s16(data_out + j * stride_out + i, vreinterpretq_s16_s32(out0));
vst1q_s16(data_out + (j + 1) * stride_out + i,
vreinterpretq_s16_s32(out1));
vst1q_s16(data_out + (j + 2) * stride_out + i,
vreinterpretq_s16_s32(out2));
vst1q_s16(data_out + (j + 3) * stride_out + i,
vreinterpretq_s16_s32(out3));
vst1q_s16(data_out + (j + 4) * stride_out + i,
vreinterpretq_s16_s32(out4));
vst1q_s16(data_out + (j + 5) * stride_out + i,
vreinterpretq_s16_s32(out5));
vst1q_s16(data_out + (j + 6) * stride_out + i,
vreinterpretq_s16_s32(out6));
vst1q_s16(data_out + (j + 7) * stride_out + i,
vreinterpretq_s16_s32(out7));
}
}
}
template <size_t N>
struct FastDCTTag {};
#include "lib/jxl/fast_dct128-inl.h"
#include "lib/jxl/fast_dct16-inl.h"
#include "lib/jxl/fast_dct256-inl.h"
#include "lib/jxl/fast_dct32-inl.h"
#include "lib/jxl/fast_dct64-inl.h"
#include "lib/jxl/fast_dct8-inl.h"
template <size_t ROWS, size_t COLS>
struct ComputeFastScaledIDCT {
// scratch_space must be aligned, and should have space for ROWS*COLS
// int16_ts.
HWY_MAYBE_UNUSED void operator()(int16_t* JXL_RESTRICT from, int16_t* to,
size_t to_stride,
int16_t* JXL_RESTRICT scratch_space) {
// Reverse the steps done in ComputeScaledDCT.
if (ROWS < COLS) {
FastTransposeBlock(from, COLS, ROWS, COLS, scratch_space, ROWS);
FastIDCT(FastDCTTag<COLS>(), scratch_space, ROWS, from, ROWS, ROWS);
FastTransposeBlock(from, ROWS, COLS, ROWS, scratch_space, COLS);
FastIDCT(FastDCTTag<ROWS>(), scratch_space, COLS, to, to_stride, COLS);
} else {
FastIDCT(FastDCTTag<COLS>(), from, ROWS, scratch_space, ROWS, ROWS);
FastTransposeBlock(scratch_space, ROWS, COLS, ROWS, from, COLS);
FastIDCT(FastDCTTag<ROWS>(), from, COLS, to, to_stride, COLS);
}
}
};
#endif
template <size_t N, size_t M>
HWY_NOINLINE void TestFastIDCT() {
#if HWY_TARGET == HWY_NEON
auto pixels_mem = hwy::AllocateAligned<float>(N * M);
float* pixels = pixels_mem.get();
auto dct_mem = hwy::AllocateAligned<float>(N * M);
float* dct = dct_mem.get();
auto dct_i_mem = hwy::AllocateAligned<int16_t>(N * M);
int16_t* dct_i = dct_i_mem.get();
auto dct_in_mem = hwy::AllocateAligned<int16_t>(N * M);
int16_t* dct_in = dct_in_mem.get();
auto idct_mem = hwy::AllocateAligned<int16_t>(N * M);
int16_t* idct = idct_mem.get();
const HWY_FULL(float) df;
auto scratch_space_mem = hwy::AllocateAligned<float>(
N * M * 2 + 3 * std::max(N, M) * MaxLanes(df));
float* scratch_space = scratch_space_mem.get();
auto scratch_space_i_mem = hwy::AllocateAligned<int16_t>(N * M * 2);
int16_t* scratch_space_i = scratch_space_i_mem.get();
Rng rng(0);
for (size_t i = 0; i < N * M; i++) {
pixels[i] = rng.UniformF(-1, 1);
}
ComputeScaledDCT<M, N>()(DCTFrom(pixels, N), dct, scratch_space);
size_t integer_bits = std::max(FastIDCTIntegerBits(FastDCTTag<N>()),
FastIDCTIntegerBits(FastDCTTag<M>()));
// Enough range for [-2, 2] output values.
JXL_ASSERT(integer_bits <= 14);
float scale = (1 << (14 - integer_bits));
for (size_t i = 0; i < N * M; i++) {
dct_i[i] = std::round(dct[i] * scale);
}
for (size_t j = 0; j < 40000000 / (M * N); j++) {
memcpy(dct_in, dct_i, sizeof(*dct_i) * N * M);
ComputeFastScaledIDCT<M, N>()(dct_in, idct, N, scratch_space_i);
}
float max_error = 0;
for (size_t i = 0; i < M * N; i++) {
float err = std::abs(idct[i] * (1.0f / scale) - pixels[i]);
if (std::abs(err) > max_error) {
max_error = std::abs(err);
}
}
printf("max error: %f mantissa bits: %d\n", max_error,
14 - (int)integer_bits);
#endif
}
template <size_t N, size_t M>
HWY_NOINLINE void TestFloatIDCT() {
auto pixels_mem = hwy::AllocateAligned<float>(N * M);
float* pixels = pixels_mem.get();
auto dct_mem = hwy::AllocateAligned<float>(N * M);
float* dct = dct_mem.get();
auto idct_mem = hwy::AllocateAligned<float>(N * M);
float* idct = idct_mem.get();
auto dct_in_mem = hwy::AllocateAligned<float>(N * M);
float* dct_in = dct_mem.get();
auto scratch_space_mem = hwy::AllocateAligned<float>(N * M * 5);
float* scratch_space = scratch_space_mem.get();
Rng rng(0);
for (size_t i = 0; i < N * M; i++) {
pixels[i] = rng.UniformF(-1, 1);
}
ComputeScaledDCT<M, N>()(DCTFrom(pixels, N), dct, scratch_space);
for (size_t j = 0; j < 40000000 / (M * N); j++) {
memcpy(dct_in, dct, sizeof(*dct) * N * M);
ComputeScaledIDCT<M, N>()(dct_in, DCTTo(idct, N), scratch_space);
}
float max_error = 0;
for (size_t i = 0; i < M * N; i++) {
float err = std::abs(idct[i] - pixels[i]);
if (std::abs(err) > max_error) {
max_error = std::abs(err);
}
}
printf("max error: %e\n", max_error);
}
} // namespace
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace jxl
HWY_AFTER_NAMESPACE();
#endif // LIB_JXL_FAST_DCT_INL_H_
|
