/src/libjxl/lib/jxl/dec_patch_dictionary.cc
Line | Count | Source |
1 | | // Copyright (c) the JPEG XL Project Authors. All rights reserved. |
2 | | // |
3 | | // Use of this source code is governed by a BSD-style |
4 | | // license that can be found in the LICENSE file. |
5 | | |
6 | | #include "lib/jxl/dec_patch_dictionary.h" |
7 | | |
8 | | #include <jxl/memory_manager.h> |
9 | | |
10 | | #include <algorithm> |
11 | | #include <cstdint> |
12 | | #include <cstdlib> |
13 | | #include <utility> |
14 | | #include <vector> |
15 | | |
16 | | #include "lib/jxl/base/printf_macros.h" |
17 | | #include "lib/jxl/base/status.h" |
18 | | #include "lib/jxl/blending.h" |
19 | | #include "lib/jxl/common.h" // kMaxNumReferenceFrames |
20 | | #include "lib/jxl/dec_ans.h" |
21 | | #include "lib/jxl/dec_bit_reader.h" |
22 | | #include "lib/jxl/image.h" |
23 | | #include "lib/jxl/image_bundle.h" |
24 | | #include "lib/jxl/image_metadata.h" |
25 | | #include "lib/jxl/pack_signed.h" |
26 | | #include "lib/jxl/patch_dictionary_internal.h" |
27 | | |
28 | | namespace jxl { |
29 | | |
30 | | Status PatchDictionary::Decode(JxlMemoryManager* memory_manager, BitReader* br, |
31 | | size_t xsize, size_t ysize, |
32 | | size_t num_extra_channels, |
33 | 1.73k | bool* uses_extra_channels) { |
34 | 1.73k | positions_.clear(); |
35 | 1.73k | blendings_stride_ = num_extra_channels + 1; |
36 | 1.73k | std::vector<uint8_t> context_map; |
37 | 1.73k | ANSCode code; |
38 | 1.73k | JXL_RETURN_IF_ERROR(DecodeHistograms( |
39 | 1.73k | memory_manager, br, kNumPatchDictionaryContexts, &code, &context_map)); |
40 | 3.17k | JXL_ASSIGN_OR_RETURN(ANSSymbolReader decoder, |
41 | 3.17k | ANSSymbolReader::Create(&code, br)); |
42 | | |
43 | 3.38M | auto read_num = [&](size_t context) -> size_t { |
44 | 3.38M | size_t r = decoder.ReadHybridUint(context, br, context_map); |
45 | 3.38M | return r; |
46 | 3.38M | }; |
47 | | |
48 | 3.17k | size_t num_ref_patch = read_num(kNumRefPatchContext); |
49 | | // Limit max memory usage of patches to about 66 bytes per pixel (assuming 8 |
50 | | // bytes per size_t) |
51 | 3.17k | const size_t num_pixels = xsize * ysize; |
52 | 3.17k | const size_t max_ref_patches = 1024 + num_pixels / 4; |
53 | 3.17k | const size_t max_patches = max_ref_patches * 4; |
54 | 3.17k | const size_t max_blending_infos = max_patches * 4; |
55 | 3.17k | if (num_ref_patch > max_ref_patches) { |
56 | 6 | return JXL_FAILURE("Too many patches in dictionary"); |
57 | 6 | } |
58 | | |
59 | 1.57k | size_t total_patches = 0; |
60 | 1.57k | size_t next_size = 1; |
61 | | |
62 | 117k | for (size_t id = 0; id < num_ref_patch; id++) { |
63 | 116k | PatchReferencePosition ref_pos; |
64 | 116k | ref_pos.ref = read_num(kReferenceFrameContext); |
65 | 116k | if (ref_pos.ref >= kMaxNumReferenceFrames || |
66 | 116k | reference_frames_->at(ref_pos.ref).frame->xsize() == 0) { |
67 | 40 | return JXL_FAILURE("Invalid reference frame ID"); |
68 | 40 | } |
69 | 116k | if (!reference_frames_->at(ref_pos.ref).ib_is_in_xyb) { |
70 | 3 | return JXL_FAILURE( |
71 | 3 | "Patches cannot use frames saved post color transforms"); |
72 | 3 | } |
73 | 116k | const ImageBundle& ib = *reference_frames_->at(ref_pos.ref).frame; |
74 | 116k | ref_pos.x0 = read_num(kPatchReferencePositionContext); |
75 | 116k | ref_pos.y0 = read_num(kPatchReferencePositionContext); |
76 | 116k | ref_pos.xsize = read_num(kPatchSizeContext) + 1; |
77 | 116k | ref_pos.ysize = read_num(kPatchSizeContext) + 1; |
78 | 116k | if (ref_pos.x0 + ref_pos.xsize > ib.xsize()) { |
79 | 6 | return JXL_FAILURE("Invalid position specified in reference frame"); |
80 | 6 | } |
81 | 116k | if (ref_pos.y0 + ref_pos.ysize > ib.ysize()) { |
82 | 2 | return JXL_FAILURE("Invalid position specified in reference frame"); |
83 | 2 | } |
84 | 116k | size_t id_count = read_num(kPatchCountContext); |
85 | 116k | if (id_count > max_patches) { |
86 | 4 | return JXL_FAILURE("Too many patches in dictionary"); |
87 | 4 | } |
88 | 116k | id_count++; |
89 | 116k | total_patches += id_count; |
90 | 116k | if (total_patches > max_patches) { |
91 | 3 | return JXL_FAILURE("Too many patches in dictionary"); |
92 | 3 | } |
93 | 116k | if (next_size < total_patches) { |
94 | 9.02k | next_size *= 2; |
95 | 9.02k | next_size = std::min<size_t>(next_size, max_patches); |
96 | 9.02k | } |
97 | 116k | if (next_size * blendings_stride_ > max_blending_infos) { |
98 | 0 | return JXL_FAILURE("Too many patches in dictionary"); |
99 | 0 | } |
100 | 116k | positions_.reserve(next_size); |
101 | 116k | blendings_.reserve(next_size * blendings_stride_); |
102 | 116k | bool choose_alpha = (num_extra_channels > 1); |
103 | 1.00M | for (size_t i = 0; i < id_count; i++) { |
104 | 885k | PatchPosition pos; |
105 | 885k | pos.ref_pos_idx = ref_positions_.size(); |
106 | 885k | if (i == 0) { |
107 | 116k | pos.x = read_num(kPatchPositionContext); |
108 | 116k | pos.y = read_num(kPatchPositionContext); |
109 | 769k | } else { |
110 | 769k | ptrdiff_t deltax = UnpackSigned(read_num(kPatchOffsetContext)); |
111 | 769k | if (deltax < 0 && static_cast<size_t>(-deltax) > positions_.back().x) { |
112 | 5 | return JXL_FAILURE("Invalid patch: negative x coordinate (%" PRIuS |
113 | 5 | " base x %" PRIdS " delta x)", |
114 | 5 | positions_.back().x, deltax); |
115 | 5 | } |
116 | 769k | pos.x = positions_.back().x + deltax; |
117 | 769k | ptrdiff_t deltay = UnpackSigned(read_num(kPatchOffsetContext)); |
118 | 769k | if (deltay < 0 && static_cast<size_t>(-deltay) > positions_.back().y) { |
119 | 8 | return JXL_FAILURE("Invalid patch: negative y coordinate (%" PRIuS |
120 | 8 | " base y %" PRIdS " delta y)", |
121 | 8 | positions_.back().y, deltay); |
122 | 8 | } |
123 | 769k | pos.y = positions_.back().y + deltay; |
124 | 769k | } |
125 | 885k | if (pos.x + ref_pos.xsize > xsize) { |
126 | 2 | return JXL_FAILURE("Invalid patch x: at %" PRIuS " + %" PRIuS |
127 | 2 | " > %" PRIuS, |
128 | 2 | pos.x, ref_pos.xsize, xsize); |
129 | 2 | } |
130 | 885k | if (pos.y + ref_pos.ysize > ysize) { |
131 | 9 | return JXL_FAILURE("Invalid patch y: at %" PRIuS " + %" PRIuS |
132 | 9 | " > %" PRIuS, |
133 | 9 | pos.y, ref_pos.ysize, ysize); |
134 | 9 | } |
135 | 1.78M | for (size_t j = 0; j < blendings_stride_; j++) { |
136 | 902k | uint32_t blend_mode = read_num(kPatchBlendModeContext); |
137 | 902k | if (blend_mode >= kNumPatchBlendModes) { |
138 | 8 | return JXL_FAILURE("Invalid patch blend mode: %u", blend_mode); |
139 | 8 | } |
140 | 902k | PatchBlending info; |
141 | 902k | info.mode = static_cast<PatchBlendMode>(blend_mode); |
142 | 902k | if (UsesAlpha(info.mode)) { |
143 | 13.5k | *uses_extra_channels = true; |
144 | 13.5k | } |
145 | 902k | if (info.mode != PatchBlendMode::kNone && j > 0) { |
146 | 7 | *uses_extra_channels = true; |
147 | 7 | } |
148 | 902k | if (UsesAlpha(info.mode) && choose_alpha) { |
149 | 0 | info.alpha_channel = read_num(kPatchAlphaChannelContext); |
150 | 0 | if (info.alpha_channel >= num_extra_channels) { |
151 | 0 | return JXL_FAILURE( |
152 | 0 | "Invalid alpha channel for blending: %u out of %u\n", |
153 | 0 | info.alpha_channel, static_cast<uint32_t>(num_extra_channels)); |
154 | 0 | } |
155 | 902k | } else { |
156 | 902k | info.alpha_channel = 0; |
157 | 902k | } |
158 | 902k | if (UsesClamp(info.mode)) { |
159 | 15.7k | info.clamp = static_cast<bool>(read_num(kPatchClampContext)); |
160 | 886k | } else { |
161 | 886k | info.clamp = false; |
162 | 886k | } |
163 | 902k | blendings_.push_back(info); |
164 | 902k | } |
165 | 885k | positions_.emplace_back(pos); |
166 | 885k | } |
167 | 116k | ref_positions_.emplace_back(ref_pos); |
168 | 116k | } |
169 | 1.48k | positions_.shrink_to_fit(); |
170 | | |
171 | 1.48k | if (!decoder.CheckANSFinalState()) { |
172 | 0 | return JXL_FAILURE("ANS checksum failure."); |
173 | 0 | } |
174 | | |
175 | 1.48k | ComputePatchTree(); |
176 | 1.48k | return true; |
177 | 1.48k | } |
178 | | |
179 | 535 | int PatchDictionary::GetReferences() const { |
180 | 535 | int result = 0; |
181 | 9.24k | for (const auto& ref_pos : ref_positions_) { |
182 | 9.24k | result |= (1 << static_cast<int>(ref_pos.ref)); |
183 | 9.24k | } |
184 | 535 | return result; |
185 | 535 | } |
186 | | |
187 | | namespace { |
188 | | struct PatchInterval { |
189 | | size_t idx; |
190 | | size_t y0, y1; |
191 | | }; |
192 | | } // namespace |
193 | | |
194 | 106k | void PatchDictionary::ComputePatchTree() { |
195 | 106k | patch_tree_.clear(); |
196 | 106k | num_patches_.clear(); |
197 | 106k | sorted_patches_y0_.clear(); |
198 | 106k | sorted_patches_y1_.clear(); |
199 | 106k | if (positions_.empty()) { |
200 | 104k | return; |
201 | 104k | } |
202 | | // Create a y-interval for each patch. |
203 | 2.62k | std::vector<PatchInterval> intervals(positions_.size()); |
204 | 2.46M | for (size_t i = 0; i < positions_.size(); ++i) { |
205 | 2.46M | const auto& pos = positions_[i]; |
206 | 2.46M | intervals[i].idx = i; |
207 | 2.46M | intervals[i].y0 = pos.y; |
208 | 2.46M | intervals[i].y1 = pos.y + ref_positions_[pos.ref_pos_idx].ysize; |
209 | 2.46M | } |
210 | 325k | auto sort_by_y0 = [&intervals](size_t start, size_t end) { |
211 | 325k | std::sort(intervals.data() + start, intervals.data() + end, |
212 | 78.6M | [](const PatchInterval& i0, const PatchInterval& i1) { |
213 | 78.6M | return i0.y0 < i1.y0; |
214 | 78.6M | }); |
215 | 325k | }; |
216 | 165k | auto sort_by_y1 = [&intervals](size_t start, size_t end) { |
217 | 165k | std::sort(intervals.data() + start, intervals.data() + end, |
218 | 64.0M | [](const PatchInterval& i0, const PatchInterval& i1) { |
219 | 64.0M | return i0.y1 < i1.y1; |
220 | 64.0M | }); |
221 | 165k | }; |
222 | | // Count the number of patches for each row. |
223 | 2.62k | sort_by_y1(0, intervals.size()); |
224 | 2.62k | num_patches_.resize(intervals.back().y1); |
225 | 2.46M | for (auto iv : intervals) { |
226 | 16.1M | for (size_t y = iv.y0; y < iv.y1; ++y) num_patches_[y]++; |
227 | 2.46M | } |
228 | 2.62k | PatchTreeNode root; |
229 | 2.62k | root.start = 0; |
230 | 2.62k | root.num = intervals.size(); |
231 | 2.62k | patch_tree_.push_back(root); |
232 | 2.62k | size_t next = 0; |
233 | 165k | while (next < patch_tree_.size()) { |
234 | 162k | auto& node = patch_tree_[next]; |
235 | 162k | size_t start = node.start; |
236 | 162k | size_t end = node.start + node.num; |
237 | | // Choose the y_center for this node to be the median of interval starts. |
238 | 162k | sort_by_y0(start, end); |
239 | 162k | size_t middle_idx = start + node.num / 2; |
240 | 162k | node.y_center = intervals[middle_idx].y0; |
241 | | // Divide the intervals in [start, end) into three groups: |
242 | | // * those completely to the right of y_center: [right_start, end) |
243 | | // * those overlapping y_center: [left_end, right_start) |
244 | | // * those completely to the left of y_center: [start, left_end) |
245 | 162k | size_t right_start = middle_idx; |
246 | 1.20M | while (right_start < end && intervals[right_start].y0 == node.y_center) { |
247 | 1.04M | ++right_start; |
248 | 1.04M | } |
249 | 162k | sort_by_y1(start, right_start); |
250 | 162k | size_t left_end = right_start; |
251 | 2.62M | while (left_end > start && intervals[left_end - 1].y1 > node.y_center) { |
252 | 2.46M | --left_end; |
253 | 2.46M | } |
254 | | // Fill in sorted_patches_y0_ and sorted_patches_y1_ for the current node. |
255 | 162k | node.num = right_start - left_end; |
256 | 162k | node.start = sorted_patches_y0_.size(); |
257 | 162k | for (ptrdiff_t i = static_cast<ptrdiff_t>(right_start) - 1; |
258 | 2.62M | i >= static_cast<ptrdiff_t>(left_end); --i) { |
259 | 2.46M | sorted_patches_y1_.emplace_back(intervals[i].y1, intervals[i].idx); |
260 | 2.46M | } |
261 | 162k | sort_by_y0(left_end, right_start); |
262 | 2.62M | for (size_t i = left_end; i < right_start; ++i) { |
263 | 2.46M | sorted_patches_y0_.emplace_back(intervals[i].y0, intervals[i].idx); |
264 | 2.46M | } |
265 | | // Create the left and right nodes (if not empty). |
266 | 162k | node.left_child = node.right_child = -1; |
267 | 162k | if (left_end > start) { |
268 | 76.8k | PatchTreeNode left; |
269 | 76.8k | left.start = start; |
270 | 76.8k | left.num = left_end - left.start; |
271 | 76.8k | patch_tree_[next].left_child = patch_tree_.size(); |
272 | 76.8k | patch_tree_.push_back(left); |
273 | 76.8k | } |
274 | 162k | if (right_start < end) { |
275 | 83.1k | PatchTreeNode right; |
276 | 83.1k | right.start = right_start; |
277 | 83.1k | right.num = end - right.start; |
278 | 83.1k | patch_tree_[next].right_child = patch_tree_.size(); |
279 | 83.1k | patch_tree_.push_back(right); |
280 | 83.1k | } |
281 | 162k | ++next; |
282 | 162k | } |
283 | 2.62k | } |
284 | | |
285 | 3.49M | std::vector<size_t> PatchDictionary::GetPatchesForRow(size_t y) const { |
286 | 3.49M | std::vector<size_t> result; |
287 | 3.49M | if (y < num_patches_.size() && num_patches_[y] > 0) { |
288 | 1.85M | result.reserve(num_patches_[y]); |
289 | 13.4M | for (ptrdiff_t tree_idx = 0; tree_idx != -1;) { |
290 | 11.6M | JXL_DASSERT(tree_idx < static_cast<ptrdiff_t>(patch_tree_.size())); |
291 | 11.6M | const auto& node = patch_tree_[tree_idx]; |
292 | 11.6M | if (y <= node.y_center) { |
293 | 21.4M | for (size_t i = 0; i < node.num; ++i) { |
294 | 20.6M | const auto& p = sorted_patches_y0_[node.start + i]; |
295 | 20.6M | if (y < p.first) break; |
296 | 15.6M | result.push_back(p.second); |
297 | 15.6M | } |
298 | 5.85M | tree_idx = y < node.y_center ? node.left_child : -1; |
299 | 5.85M | } else { |
300 | 17.7M | for (size_t i = 0; i < node.num; ++i) { |
301 | 17.0M | const auto& p = sorted_patches_y1_[node.start + i]; |
302 | 17.0M | if (y >= p.first) break; |
303 | 12.0M | result.push_back(p.second); |
304 | 12.0M | } |
305 | 5.77M | tree_idx = node.right_child; |
306 | 5.77M | } |
307 | 11.6M | } |
308 | | // Ensure that he relative order of patches that affect the same pixels is |
309 | | // preserved. This is important for patches that have a blend mode |
310 | | // different from kAdd. |
311 | 1.85M | std::sort(result.begin(), result.end()); |
312 | 1.85M | } |
313 | 3.49M | return result; |
314 | 3.49M | } |
315 | | |
316 | | // Adds patches to a segment of `xsize` pixels, starting at `inout`, assumed |
317 | | // to be located at position (x0, y) in the frame. |
318 | | Status PatchDictionary::AddOneRow( |
319 | | float* const* inout, size_t y, size_t x0, size_t xsize, |
320 | 2.56M | const std::vector<ExtraChannelInfo>& extra_channel_info) const { |
321 | 2.56M | size_t num_ec = extra_channel_info.size(); |
322 | 2.56M | JXL_ENSURE(num_ec + 1 <= blendings_stride_); |
323 | 2.56M | std::vector<const float*> fg_ptrs(3 + num_ec); |
324 | 23.6M | for (size_t pos_idx : GetPatchesForRow(y)) { |
325 | 23.6M | const size_t blending_idx = pos_idx * blendings_stride_; |
326 | 23.6M | const PatchPosition& pos = positions_[pos_idx]; |
327 | 23.6M | const PatchReferencePosition& ref_pos = ref_positions_[pos.ref_pos_idx]; |
328 | 23.6M | size_t by = pos.y; |
329 | 23.6M | size_t bx = pos.x; |
330 | 23.6M | size_t patch_xsize = ref_pos.xsize; |
331 | 23.6M | JXL_ENSURE(y >= by); |
332 | 23.6M | JXL_ENSURE(y < by + ref_pos.ysize); |
333 | 23.6M | size_t iy = y - by; |
334 | 23.6M | size_t ref = ref_pos.ref; |
335 | 23.6M | if (bx >= x0 + xsize) continue; |
336 | 18.2M | if (bx + patch_xsize < x0) continue; |
337 | 12.3M | size_t patch_x0 = std::max(bx, x0); |
338 | 12.3M | size_t patch_x1 = std::min(bx + patch_xsize, x0 + xsize); |
339 | 49.5M | for (size_t c = 0; c < 3; c++) { |
340 | 37.1M | fg_ptrs[c] = reference_frames_->at(ref).frame->color()->ConstPlaneRow( |
341 | 37.1M | c, ref_pos.y0 + iy) + |
342 | 37.1M | ref_pos.x0 + x0 - bx; |
343 | 37.1M | } |
344 | 12.3M | for (size_t i = 0; i < num_ec; i++) { |
345 | 0 | fg_ptrs[3 + i] = |
346 | 0 | reference_frames_->at(ref).frame->extra_channels()[i].ConstRow( |
347 | 0 | ref_pos.y0 + iy) + |
348 | 0 | ref_pos.x0 + x0 - bx; |
349 | 0 | } |
350 | 12.3M | JXL_RETURN_IF_ERROR(PerformBlending( |
351 | 12.3M | memory_manager_, inout, fg_ptrs.data(), inout, patch_x0 - x0, |
352 | 12.3M | patch_x1 - patch_x0, blendings_[blending_idx], |
353 | 12.3M | blendings_.data() + blending_idx + 1, extra_channel_info)); |
354 | 12.3M | } |
355 | 2.56M | return true; |
356 | 2.56M | } |
357 | | } // namespace jxl |