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
|
// 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/jpegli/encode_finish.h"
#include <cmath>
#include <limits>
#include "lib/jpegli/error.h"
#include "lib/jpegli/memory_manager.h"
#include "lib/jpegli/quant.h"
#undef HWY_TARGET_INCLUDE
#define HWY_TARGET_INCLUDE "lib/jpegli/encode_finish.cc"
#include <hwy/foreach_target.h>
#include <hwy/highway.h>
#include "lib/jpegli/dct-inl.h"
HWY_BEFORE_NAMESPACE();
namespace jpegli {
namespace HWY_NAMESPACE {
// These templates are not found via ADL.
using hwy::HWY_NAMESPACE::GetLane;
using D = HWY_FULL(float);
using DI = HWY_FULL(int32_t);
using DI16 = Rebind<int16_t, HWY_FULL(int32_t)>;
void ReQuantizeBlock(int16_t* block, const float* qmc, float aq_strength,
const float* zero_bias_offset,
const float* zero_bias_mul) {
D d;
DI di;
DI16 di16;
const auto aq_mul = Set(d, aq_strength);
for (size_t k = 0; k < DCTSIZE2; k += Lanes(d)) {
const auto in = Load(di16, block + k);
const auto val = ConvertTo(d, PromoteTo(di, in));
const auto q = Load(d, qmc + k);
const auto qval = Mul(val, q);
const auto zb_offset = Load(d, zero_bias_offset + k);
const auto zb_mul = Load(d, zero_bias_mul + k);
const auto threshold = Add(zb_offset, Mul(zb_mul, aq_mul));
const auto nzero_mask = Ge(Abs(qval), threshold);
const auto iqval = IfThenElseZero(nzero_mask, Round(qval));
Store(DemoteTo(di16, ConvertTo(di, iqval)), di16, block + k);
}
}
float BlockError(const int16_t* block, const float* qmc, const float* iqmc,
const float aq_strength, const float* zero_bias_offset,
const float* zero_bias_mul) {
D d;
DI di;
DI16 di16;
auto err = Zero(d);
const auto scale = Set(d, 1.0 / 16);
const auto aq_mul = Set(d, aq_strength);
for (size_t k = 0; k < DCTSIZE2; k += Lanes(d)) {
const auto in = Load(di16, block + k);
const auto val = ConvertTo(d, PromoteTo(di, in));
const auto q = Load(d, qmc + k);
const auto qval = Mul(val, q);
const auto zb_offset = Load(d, zero_bias_offset + k);
const auto zb_mul = Load(d, zero_bias_mul + k);
const auto threshold = Add(zb_offset, Mul(zb_mul, aq_mul));
const auto nzero_mask = Ge(Abs(qval), threshold);
const auto iqval = IfThenElseZero(nzero_mask, Round(qval));
const auto invq = Load(d, iqmc + k);
const auto rval = Mul(iqval, invq);
const auto diff = Mul(Sub(val, rval), scale);
err = Add(err, Mul(diff, diff));
}
return GetLane(SumOfLanes(d, err));
}
void ComputeInverseWeights(const float* qmc, float* iqmc) {
for (int k = 0; k < 64; ++k) {
iqmc[k] = 1.0f / qmc[k];
}
}
float ComputePSNR(j_compress_ptr cinfo, int sampling) {
jpeg_comp_master* m = cinfo->master;
InitQuantizer(cinfo, QuantPass::SEARCH_SECOND_PASS);
double error = 0.0;
size_t num = 0;
for (int c = 0; c < cinfo->num_components; ++c) {
jpeg_component_info* comp = &cinfo->comp_info[c];
const float* qmc = m->quant_mul[c];
const int h_factor = m->h_factor[c];
const int v_factor = m->v_factor[c];
const float* zero_bias_offset = m->zero_bias_offset[c];
const float* zero_bias_mul = m->zero_bias_mul[c];
HWY_ALIGN float iqmc[64];
ComputeInverseWeights(qmc, iqmc);
for (JDIMENSION by = 0; by < comp->height_in_blocks; by += sampling) {
JBLOCKARRAY ba = GetBlockRow(cinfo, c, by);
const float* qf = m->quant_field.Row(by * v_factor);
for (JDIMENSION bx = 0; bx < comp->width_in_blocks; bx += sampling) {
error += BlockError(&ba[0][bx][0], qmc, iqmc, qf[bx * h_factor],
zero_bias_offset, zero_bias_mul);
num += DCTSIZE2;
}
}
}
return 4.3429448f * log(num / (error / 255. / 255.));
}
void ReQuantizeCoeffs(j_compress_ptr cinfo) {
jpeg_comp_master* m = cinfo->master;
InitQuantizer(cinfo, QuantPass::SEARCH_SECOND_PASS);
for (int c = 0; c < cinfo->num_components; ++c) {
jpeg_component_info* comp = &cinfo->comp_info[c];
const float* qmc = m->quant_mul[c];
const int h_factor = m->h_factor[c];
const int v_factor = m->v_factor[c];
const float* zero_bias_offset = m->zero_bias_offset[c];
const float* zero_bias_mul = m->zero_bias_mul[c];
for (JDIMENSION by = 0; by < comp->height_in_blocks; ++by) {
JBLOCKARRAY ba = GetBlockRow(cinfo, c, by);
const float* qf = m->quant_field.Row(by * v_factor);
for (JDIMENSION bx = 0; bx < comp->width_in_blocks; ++bx) {
ReQuantizeBlock(&ba[0][bx][0], qmc, qf[bx * h_factor], zero_bias_offset,
zero_bias_mul);
}
}
}
}
// NOLINTNEXTLINE(google-readability-namespace-comments)
} // namespace HWY_NAMESPACE
} // namespace jpegli
HWY_AFTER_NAMESPACE();
#if HWY_ONCE
namespace jpegli {
namespace {
HWY_EXPORT(ComputePSNR);
HWY_EXPORT(ReQuantizeCoeffs);
void ReQuantizeCoeffs(j_compress_ptr cinfo) {
HWY_DYNAMIC_DISPATCH(ReQuantizeCoeffs)(cinfo);
}
float ComputePSNR(j_compress_ptr cinfo, int sampling) {
return HWY_DYNAMIC_DISPATCH(ComputePSNR)(cinfo, sampling);
}
void UpdateDistance(j_compress_ptr cinfo, float distance) {
float distances[NUM_QUANT_TBLS] = {distance, distance, distance};
SetQuantMatrices(cinfo, distances, /*add_two_chroma_tables=*/true);
}
float Clamp(float val, float minval, float maxval) {
return std::max(minval, std::min(maxval, val));
}
#define PSNR_SEARCH_DBG 0
float FindDistanceForPSNR(j_compress_ptr cinfo) {
constexpr int kMaxIters = 20;
const float psnr_target = cinfo->master->psnr_target;
const float tolerance = cinfo->master->psnr_tolerance;
const float min_dist = cinfo->master->min_distance;
const float max_dist = cinfo->master->max_distance;
float d = Clamp(1.0f, min_dist, max_dist);
for (int sampling : {4, 1}) {
float best_diff = std::numeric_limits<float>::max();
float best_distance = 0.0f;
float best_psnr = 0.0;
float dmin = min_dist;
float dmax = max_dist;
bool found_lower_bound = false;
bool found_upper_bound = false;
for (int i = 0; i < kMaxIters; ++i) {
UpdateDistance(cinfo, d);
float psnr = ComputePSNR(cinfo, sampling);
if (psnr > psnr_target) {
dmin = d;
found_lower_bound = true;
} else {
dmax = d;
found_upper_bound = true;
}
#if (PSNR_SEARCH_DBG > 1)
printf("sampling %d iter %2d d %7.4f psnr %.2f", sampling, i, d, psnr);
if (found_upper_bound && found_lower_bound) {
printf(" d-interval: [ %7.4f .. %7.4f ]", dmin, dmax);
}
printf("\n");
#endif
float diff = std::abs(psnr - psnr_target);
if (diff < best_diff) {
best_diff = diff;
best_distance = d;
best_psnr = psnr;
}
if (diff < tolerance * psnr_target || dmin == dmax) {
break;
}
if (!found_lower_bound || !found_upper_bound) {
d *= std::exp(0.15f * (psnr - psnr_target));
} else {
d = 0.5f * (dmin + dmax);
}
d = Clamp(d, min_dist, max_dist);
}
d = best_distance;
if (sampling == 1 && PSNR_SEARCH_DBG) {
printf("Final PSNR %.2f at distance %.4f\n", best_psnr, d);
} else {
(void)best_psnr;
}
}
return d;
}
} // namespace
void QuantizetoPSNR(j_compress_ptr cinfo) {
float distance = FindDistanceForPSNR(cinfo);
UpdateDistance(cinfo, distance);
ReQuantizeCoeffs(cinfo);
}
} // namespace jpegli
#endif // HWY_ONCE
|
