/src/libjxl/lib/jxl/enc_frame.cc
Line | Count | Source (jump to first uncovered line) |
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/enc_frame.h" |
7 | | |
8 | | #include <jxl/memory_manager.h> |
9 | | |
10 | | #include <algorithm> |
11 | | #include <array> |
12 | | #include <atomic> |
13 | | #include <cmath> |
14 | | #include <cstddef> |
15 | | #include <cstdint> |
16 | | #include <memory> |
17 | | #include <numeric> |
18 | | #include <utility> |
19 | | #include <vector> |
20 | | |
21 | | #include "lib/jxl/ac_context.h" |
22 | | #include "lib/jxl/ac_strategy.h" |
23 | | #include "lib/jxl/base/bits.h" |
24 | | #include "lib/jxl/base/common.h" |
25 | | #include "lib/jxl/base/compiler_specific.h" |
26 | | #include "lib/jxl/base/data_parallel.h" |
27 | | #include "lib/jxl/base/override.h" |
28 | | #include "lib/jxl/base/printf_macros.h" |
29 | | #include "lib/jxl/base/rect.h" |
30 | | #include "lib/jxl/base/status.h" |
31 | | #include "lib/jxl/chroma_from_luma.h" |
32 | | #include "lib/jxl/coeff_order.h" |
33 | | #include "lib/jxl/coeff_order_fwd.h" |
34 | | #include "lib/jxl/color_encoding_internal.h" |
35 | | #include "lib/jxl/common.h" // kMaxNumPasses |
36 | | #include "lib/jxl/dct_util.h" |
37 | | #include "lib/jxl/dec_external_image.h" |
38 | | #include "lib/jxl/enc_ac_strategy.h" |
39 | | #include "lib/jxl/enc_adaptive_quantization.h" |
40 | | #include "lib/jxl/enc_ans.h" |
41 | | #include "lib/jxl/enc_aux_out.h" |
42 | | #include "lib/jxl/enc_bit_writer.h" |
43 | | #include "lib/jxl/enc_cache.h" |
44 | | #include "lib/jxl/enc_chroma_from_luma.h" |
45 | | #include "lib/jxl/enc_coeff_order.h" |
46 | | #include "lib/jxl/enc_context_map.h" |
47 | | #include "lib/jxl/enc_entropy_coder.h" |
48 | | #include "lib/jxl/enc_external_image.h" |
49 | | #include "lib/jxl/enc_fields.h" |
50 | | #include "lib/jxl/enc_group.h" |
51 | | #include "lib/jxl/enc_heuristics.h" |
52 | | #include "lib/jxl/enc_modular.h" |
53 | | #include "lib/jxl/enc_noise.h" |
54 | | #include "lib/jxl/enc_params.h" |
55 | | #include "lib/jxl/enc_patch_dictionary.h" |
56 | | #include "lib/jxl/enc_photon_noise.h" |
57 | | #include "lib/jxl/enc_quant_weights.h" |
58 | | #include "lib/jxl/enc_splines.h" |
59 | | #include "lib/jxl/enc_toc.h" |
60 | | #include "lib/jxl/enc_xyb.h" |
61 | | #include "lib/jxl/fields.h" |
62 | | #include "lib/jxl/frame_dimensions.h" |
63 | | #include "lib/jxl/frame_header.h" |
64 | | #include "lib/jxl/image.h" |
65 | | #include "lib/jxl/image_bundle.h" |
66 | | #include "lib/jxl/image_ops.h" |
67 | | #include "lib/jxl/jpeg/enc_jpeg_data.h" |
68 | | #include "lib/jxl/loop_filter.h" |
69 | | #include "lib/jxl/modular/options.h" |
70 | | #include "lib/jxl/quant_weights.h" |
71 | | #include "lib/jxl/quantizer.h" |
72 | | #include "lib/jxl/splines.h" |
73 | | #include "lib/jxl/toc.h" |
74 | | |
75 | | namespace jxl { |
76 | | |
77 | 52 | Status ParamsPostInit(CompressParams* p) { |
78 | 52 | if (!p->manual_noise.empty() && |
79 | 52 | p->manual_noise.size() != NoiseParams::kNumNoisePoints) { |
80 | 0 | return JXL_FAILURE("Invalid number of noise lut entries"); |
81 | 0 | } |
82 | 52 | if (!p->manual_xyb_factors.empty() && p->manual_xyb_factors.size() != 3) { |
83 | 0 | return JXL_FAILURE("Invalid number of XYB quantization factors"); |
84 | 0 | } |
85 | 52 | if (!p->modular_mode && p->butteraugli_distance == 0.0) { |
86 | 0 | p->butteraugli_distance = kMinButteraugliDistance; |
87 | 0 | } |
88 | 52 | if (p->original_butteraugli_distance == -1.0) { |
89 | 52 | p->original_butteraugli_distance = p->butteraugli_distance; |
90 | 52 | } |
91 | 52 | if (p->resampling <= 0) { |
92 | 52 | p->resampling = 1; |
93 | | // For very low bit rates, using 2x2 resampling gives better results on |
94 | | // most photographic images, with an adjusted butteraugli score chosen to |
95 | | // give roughly the same amount of bits per pixel. |
96 | 52 | if (!p->already_downsampled && p->butteraugli_distance >= 20) { |
97 | 0 | p->resampling = 2; |
98 | 0 | p->butteraugli_distance = 6 + ((p->butteraugli_distance - 20) * 0.25); |
99 | 0 | } |
100 | 52 | } |
101 | 52 | if (p->ec_resampling <= 0) { |
102 | 52 | p->ec_resampling = p->resampling; |
103 | 52 | } |
104 | 52 | return true; |
105 | 52 | } |
106 | | |
107 | | namespace { |
108 | | |
109 | | template <typename T> |
110 | | uint32_t GetBitDepth(JxlBitDepth bit_depth, const T& metadata, |
111 | 52 | JxlPixelFormat format) { |
112 | 52 | if (bit_depth.type == JXL_BIT_DEPTH_FROM_PIXEL_FORMAT) { |
113 | 52 | return BitsPerChannel(format.data_type); |
114 | 52 | } else if (bit_depth.type == JXL_BIT_DEPTH_FROM_CODESTREAM) { |
115 | 0 | return metadata.bit_depth.bits_per_sample; |
116 | 0 | } else if (bit_depth.type == JXL_BIT_DEPTH_CUSTOM) { |
117 | 0 | return bit_depth.bits_per_sample; |
118 | 0 | } else { |
119 | 0 | return 0; |
120 | 0 | } |
121 | 52 | } enc_frame.cc:unsigned int jxl::(anonymous namespace)::GetBitDepth<jxl::ImageMetadata>(JxlBitDepth, jxl::ImageMetadata const&, JxlPixelFormat) Line | Count | Source | 111 | 52 | JxlPixelFormat format) { | 112 | 52 | if (bit_depth.type == JXL_BIT_DEPTH_FROM_PIXEL_FORMAT) { | 113 | 52 | return BitsPerChannel(format.data_type); | 114 | 52 | } else if (bit_depth.type == JXL_BIT_DEPTH_FROM_CODESTREAM) { | 115 | 0 | return metadata.bit_depth.bits_per_sample; | 116 | 0 | } else if (bit_depth.type == JXL_BIT_DEPTH_CUSTOM) { | 117 | 0 | return bit_depth.bits_per_sample; | 118 | 0 | } else { | 119 | 0 | return 0; | 120 | 0 | } | 121 | 52 | } |
Unexecuted instantiation: enc_frame.cc:unsigned int jxl::(anonymous namespace)::GetBitDepth<jxl::ExtraChannelInfo>(JxlBitDepth, jxl::ExtraChannelInfo const&, JxlPixelFormat) |
122 | | |
123 | | Status CopyColorChannels(JxlChunkedFrameInputSource input, Rect rect, |
124 | | const FrameInfo& frame_info, |
125 | | const ImageMetadata& metadata, ThreadPool* pool, |
126 | | Image3F* color, ImageF* alpha, |
127 | 52 | bool* has_interleaved_alpha) { |
128 | 52 | JxlPixelFormat format = {4, JXL_TYPE_UINT8, JXL_NATIVE_ENDIAN, 0}; |
129 | 52 | input.get_color_channels_pixel_format(input.opaque, &format); |
130 | 52 | *has_interleaved_alpha = format.num_channels == 2 || format.num_channels == 4; |
131 | 52 | size_t bits_per_sample = |
132 | 52 | GetBitDepth(frame_info.image_bit_depth, metadata, format); |
133 | 52 | size_t row_offset; |
134 | 52 | auto buffer = GetColorBuffer(input, rect.x0(), rect.y0(), rect.xsize(), |
135 | 52 | rect.ysize(), &row_offset); |
136 | 52 | if (!buffer) { |
137 | 0 | return JXL_FAILURE("no buffer for color channels given"); |
138 | 0 | } |
139 | 52 | size_t color_channels = frame_info.ib_needs_color_transform |
140 | 52 | ? metadata.color_encoding.Channels() |
141 | 52 | : 3; |
142 | 52 | if (format.num_channels < color_channels) { |
143 | 0 | return JXL_FAILURE("Expected %" PRIuS |
144 | 0 | " color channels, received only %u channels", |
145 | 0 | color_channels, format.num_channels); |
146 | 0 | } |
147 | 52 | const uint8_t* data = reinterpret_cast<const uint8_t*>(buffer.get()); |
148 | 208 | for (size_t c = 0; c < color_channels; ++c) { |
149 | 156 | JXL_RETURN_IF_ERROR(ConvertFromExternalNoSizeCheck( |
150 | 156 | data, rect.xsize(), rect.ysize(), row_offset, bits_per_sample, format, |
151 | 156 | c, pool, &color->Plane(c))); |
152 | 156 | } |
153 | 52 | if (color_channels == 1) { |
154 | 0 | CopyImageTo(color->Plane(0), &color->Plane(1)); |
155 | 0 | CopyImageTo(color->Plane(0), &color->Plane(2)); |
156 | 0 | } |
157 | 52 | if (alpha) { |
158 | 0 | if (*has_interleaved_alpha) { |
159 | 0 | JXL_RETURN_IF_ERROR(ConvertFromExternalNoSizeCheck( |
160 | 0 | data, rect.xsize(), rect.ysize(), row_offset, bits_per_sample, format, |
161 | 0 | format.num_channels - 1, pool, alpha)); |
162 | 0 | } else { |
163 | | // if alpha is not passed, but it is expected, then assume |
164 | | // it is all-opaque |
165 | 0 | FillImage(1.0f, alpha); |
166 | 0 | } |
167 | 0 | } |
168 | 52 | return true; |
169 | 52 | } |
170 | | |
171 | | Status CopyExtraChannels(JxlChunkedFrameInputSource input, Rect rect, |
172 | | const FrameInfo& frame_info, |
173 | | const ImageMetadata& metadata, |
174 | | bool has_interleaved_alpha, ThreadPool* pool, |
175 | 52 | std::vector<ImageF>* extra_channels) { |
176 | 52 | for (size_t ec = 0; ec < metadata.num_extra_channels; ec++) { |
177 | 0 | if (has_interleaved_alpha && |
178 | 0 | metadata.extra_channel_info[ec].type == ExtraChannel::kAlpha) { |
179 | | // Skip this alpha channel, but still request additional alpha channels |
180 | | // if they exist. |
181 | 0 | has_interleaved_alpha = false; |
182 | 0 | continue; |
183 | 0 | } |
184 | 0 | JxlPixelFormat ec_format = {1, JXL_TYPE_UINT8, JXL_NATIVE_ENDIAN, 0}; |
185 | 0 | input.get_extra_channel_pixel_format(input.opaque, ec, &ec_format); |
186 | 0 | ec_format.num_channels = 1; |
187 | 0 | size_t row_offset; |
188 | 0 | auto buffer = |
189 | 0 | GetExtraChannelBuffer(input, ec, rect.x0(), rect.y0(), rect.xsize(), |
190 | 0 | rect.ysize(), &row_offset); |
191 | 0 | if (!buffer) { |
192 | 0 | return JXL_FAILURE("no buffer for extra channel given"); |
193 | 0 | } |
194 | 0 | size_t bits_per_sample = GetBitDepth( |
195 | 0 | frame_info.image_bit_depth, metadata.extra_channel_info[ec], ec_format); |
196 | 0 | if (!ConvertFromExternalNoSizeCheck( |
197 | 0 | reinterpret_cast<const uint8_t*>(buffer.get()), rect.xsize(), |
198 | 0 | rect.ysize(), row_offset, bits_per_sample, ec_format, 0, pool, |
199 | 0 | &(*extra_channels)[ec])) { |
200 | 0 | return JXL_FAILURE("Failed to set buffer for extra channel"); |
201 | 0 | } |
202 | 0 | } |
203 | 52 | return true; |
204 | 52 | } |
205 | | |
206 | | void SetProgressiveMode(const CompressParams& cparams, |
207 | 52 | ProgressiveSplitter* progressive_splitter) { |
208 | 52 | constexpr PassDefinition progressive_passes_dc_vlf_lf_full_ac[] = { |
209 | 52 | {/*num_coefficients=*/2, /*shift=*/0, |
210 | 52 | /*suitable_for_downsampling_of_at_least=*/4}, |
211 | 52 | {/*num_coefficients=*/3, /*shift=*/0, |
212 | 52 | /*suitable_for_downsampling_of_at_least=*/2}, |
213 | 52 | {/*num_coefficients=*/8, /*shift=*/0, |
214 | 52 | /*suitable_for_downsampling_of_at_least=*/0}, |
215 | 52 | }; |
216 | 52 | constexpr PassDefinition progressive_passes_dc_quant_ac_full_ac[] = { |
217 | 52 | {/*num_coefficients=*/8, /*shift=*/1, |
218 | 52 | /*suitable_for_downsampling_of_at_least=*/2}, |
219 | 52 | {/*num_coefficients=*/8, /*shift=*/0, |
220 | 52 | /*suitable_for_downsampling_of_at_least=*/0}, |
221 | 52 | }; |
222 | 52 | bool progressive_mode = ApplyOverride(cparams.progressive_mode, false); |
223 | 52 | bool qprogressive_mode = ApplyOverride(cparams.qprogressive_mode, false); |
224 | 52 | if (cparams.custom_progressive_mode) { |
225 | 0 | progressive_splitter->SetProgressiveMode(*cparams.custom_progressive_mode); |
226 | 52 | } else if (qprogressive_mode) { |
227 | 0 | progressive_splitter->SetProgressiveMode( |
228 | 0 | ProgressiveMode{progressive_passes_dc_quant_ac_full_ac}); |
229 | 52 | } else if (progressive_mode) { |
230 | 0 | progressive_splitter->SetProgressiveMode( |
231 | 0 | ProgressiveMode{progressive_passes_dc_vlf_lf_full_ac}); |
232 | 0 | } |
233 | 52 | } |
234 | | |
235 | 52 | uint64_t FrameFlagsFromParams(const CompressParams& cparams) { |
236 | 52 | uint64_t flags = 0; |
237 | | |
238 | 52 | const float dist = cparams.butteraugli_distance; |
239 | | |
240 | | // We don't add noise at low butteraugli distances because the original |
241 | | // noise is stored within the compressed image and adding noise makes things |
242 | | // worse. |
243 | 52 | if (ApplyOverride(cparams.noise, dist >= kMinButteraugliForNoise) || |
244 | 52 | cparams.photon_noise_iso > 0 || |
245 | 52 | cparams.manual_noise.size() == NoiseParams::kNumNoisePoints) { |
246 | 0 | flags |= FrameHeader::kNoise; |
247 | 0 | } |
248 | | |
249 | 52 | if (cparams.progressive_dc > 0 && cparams.modular_mode == false) { |
250 | 0 | flags |= FrameHeader::kUseDcFrame; |
251 | 0 | } |
252 | | |
253 | 52 | return flags; |
254 | 52 | } |
255 | | |
256 | | Status LoopFilterFromParams(const CompressParams& cparams, bool streaming_mode, |
257 | 52 | FrameHeader* JXL_RESTRICT frame_header) { |
258 | 52 | LoopFilter* loop_filter = &frame_header->loop_filter; |
259 | | |
260 | | // Gaborish defaults to enabled in Hare or slower. |
261 | 52 | loop_filter->gab = ApplyOverride( |
262 | 52 | cparams.gaborish, cparams.speed_tier <= SpeedTier::kHare && |
263 | 52 | frame_header->encoding == FrameEncoding::kVarDCT && |
264 | 52 | cparams.decoding_speed_tier < 4 && |
265 | 52 | !cparams.disable_percepeptual_optimizations); |
266 | | |
267 | 52 | if (cparams.epf != -1) { |
268 | 0 | loop_filter->epf_iters = cparams.epf; |
269 | 52 | } else if (cparams.disable_percepeptual_optimizations) { |
270 | 0 | loop_filter->epf_iters = 0; |
271 | 0 | return true; |
272 | 52 | } else { |
273 | 52 | if (frame_header->encoding == FrameEncoding::kModular) { |
274 | 0 | loop_filter->epf_iters = 0; |
275 | 52 | } else { |
276 | 52 | constexpr float kThresholds[3] = {0.7, 1.5, 4.0}; |
277 | 52 | loop_filter->epf_iters = 0; |
278 | 52 | if (cparams.decoding_speed_tier < 3) { |
279 | 208 | for (size_t i = cparams.decoding_speed_tier == 2 ? 1 : 0; i < 3; i++) { |
280 | 156 | if (cparams.butteraugli_distance >= kThresholds[i]) { |
281 | 52 | loop_filter->epf_iters++; |
282 | 52 | } |
283 | 156 | } |
284 | 52 | } |
285 | 52 | } |
286 | 52 | } |
287 | | // Strength of EPF in modular mode. |
288 | 52 | if (frame_header->encoding == FrameEncoding::kModular && |
289 | 52 | !cparams.IsLossless()) { |
290 | | // TODO(veluca): this formula is nonsense. |
291 | 0 | loop_filter->epf_sigma_for_modular = |
292 | 0 | std::max(cparams.butteraugli_distance, 1.0f); |
293 | 0 | } |
294 | 52 | if (frame_header->encoding == FrameEncoding::kModular && |
295 | 52 | cparams.lossy_palette) { |
296 | 0 | loop_filter->epf_sigma_for_modular = 1.0f; |
297 | 0 | } |
298 | | |
299 | 52 | return true; |
300 | 52 | } |
301 | | |
302 | | Status MakeFrameHeader(size_t xsize, size_t ysize, |
303 | | const CompressParams& cparams, |
304 | | const ProgressiveSplitter& progressive_splitter, |
305 | | const FrameInfo& frame_info, |
306 | | const jpeg::JPEGData* jpeg_data, bool streaming_mode, |
307 | 52 | FrameHeader* JXL_RESTRICT frame_header) { |
308 | 52 | frame_header->nonserialized_is_preview = frame_info.is_preview; |
309 | 52 | frame_header->is_last = frame_info.is_last; |
310 | 52 | frame_header->save_before_color_transform = |
311 | 52 | frame_info.save_before_color_transform; |
312 | 52 | frame_header->frame_type = frame_info.frame_type; |
313 | 52 | frame_header->name = frame_info.name; |
314 | | |
315 | 52 | progressive_splitter.InitPasses(&frame_header->passes); |
316 | | |
317 | 52 | if (cparams.modular_mode) { |
318 | 0 | frame_header->encoding = FrameEncoding::kModular; |
319 | 0 | if (cparams.modular_group_size_shift == -1) { |
320 | 0 | frame_header->group_size_shift = 1; |
321 | | // no point using groups when only one group is full and the others are |
322 | | // less than half full: multithreading will not really help much, while |
323 | | // compression does suffer |
324 | 0 | if (xsize <= 400 && ysize <= 400) { |
325 | 0 | frame_header->group_size_shift = 2; |
326 | 0 | } |
327 | 0 | } else { |
328 | 0 | frame_header->group_size_shift = cparams.modular_group_size_shift; |
329 | 0 | } |
330 | 0 | } |
331 | | |
332 | 52 | if (jpeg_data) { |
333 | | // we are transcoding a JPEG, so we don't get to choose |
334 | 0 | frame_header->encoding = FrameEncoding::kVarDCT; |
335 | 0 | frame_header->x_qm_scale = 2; |
336 | 0 | frame_header->b_qm_scale = 2; |
337 | 0 | JXL_RETURN_IF_ERROR(SetChromaSubsamplingFromJpegData( |
338 | 0 | *jpeg_data, &frame_header->chroma_subsampling)); |
339 | 0 | JXL_RETURN_IF_ERROR(SetColorTransformFromJpegData( |
340 | 0 | *jpeg_data, &frame_header->color_transform)); |
341 | 52 | } else { |
342 | 52 | frame_header->color_transform = cparams.color_transform; |
343 | 52 | if (!cparams.modular_mode && |
344 | 52 | (frame_header->chroma_subsampling.MaxHShift() != 0 || |
345 | 52 | frame_header->chroma_subsampling.MaxVShift() != 0)) { |
346 | 0 | return JXL_FAILURE( |
347 | 0 | "Chroma subsampling is not supported in VarDCT mode when not " |
348 | 0 | "recompressing JPEGs"); |
349 | 0 | } |
350 | 52 | } |
351 | 52 | if (frame_header->color_transform != ColorTransform::kYCbCr && |
352 | 52 | (frame_header->chroma_subsampling.MaxHShift() != 0 || |
353 | 52 | frame_header->chroma_subsampling.MaxVShift() != 0)) { |
354 | 0 | return JXL_FAILURE( |
355 | 0 | "Chroma subsampling is not supported when color transform is not " |
356 | 0 | "YCbCr"); |
357 | 0 | } |
358 | | |
359 | 52 | frame_header->flags = FrameFlagsFromParams(cparams); |
360 | | // Non-photon noise is not supported in the Modular encoder for now. |
361 | 52 | if (frame_header->encoding != FrameEncoding::kVarDCT && |
362 | 52 | cparams.photon_noise_iso == 0 && cparams.manual_noise.empty()) { |
363 | 0 | frame_header->UpdateFlag(false, FrameHeader::Flags::kNoise); |
364 | 0 | } |
365 | | |
366 | 52 | JXL_RETURN_IF_ERROR( |
367 | 52 | LoopFilterFromParams(cparams, streaming_mode, frame_header)); |
368 | | |
369 | 52 | frame_header->dc_level = frame_info.dc_level; |
370 | 52 | if (frame_header->dc_level > 2) { |
371 | | // With 3 or more progressive_dc frames, the implementation does not yet |
372 | | // work, see enc_cache.cc. |
373 | 0 | return JXL_FAILURE("progressive_dc > 2 is not yet supported"); |
374 | 0 | } |
375 | 52 | if (cparams.progressive_dc > 0 && |
376 | 52 | (cparams.ec_resampling != 1 || cparams.resampling != 1)) { |
377 | 0 | return JXL_FAILURE("Resampling not supported with DC frames"); |
378 | 0 | } |
379 | 52 | if (cparams.resampling != 1 && cparams.resampling != 2 && |
380 | 52 | cparams.resampling != 4 && cparams.resampling != 8) { |
381 | 0 | return JXL_FAILURE("Invalid resampling factor"); |
382 | 0 | } |
383 | 52 | if (cparams.ec_resampling != 1 && cparams.ec_resampling != 2 && |
384 | 52 | cparams.ec_resampling != 4 && cparams.ec_resampling != 8) { |
385 | 0 | return JXL_FAILURE("Invalid ec_resampling factor"); |
386 | 0 | } |
387 | | // Resized frames. |
388 | 52 | if (frame_info.frame_type != FrameType::kDCFrame) { |
389 | 52 | frame_header->frame_origin = frame_info.origin; |
390 | 52 | size_t ups = 1; |
391 | 52 | if (cparams.already_downsampled) ups = cparams.resampling; |
392 | | |
393 | | // TODO(lode): this is not correct in case of odd original image sizes in |
394 | | // combination with cparams.already_downsampled. Likely these values should |
395 | | // be set to respectively frame_header->default_xsize() and |
396 | | // frame_header->default_ysize() instead, the original (non downsampled) |
397 | | // intended decoded image dimensions. But it may be more subtle than that |
398 | | // if combined with crop. This issue causes custom_size_or_origin to be |
399 | | // incorrectly set to true in case of already_downsampled with odd output |
400 | | // image size when no cropping is used. |
401 | 52 | frame_header->frame_size.xsize = xsize * ups; |
402 | 52 | frame_header->frame_size.ysize = ysize * ups; |
403 | 52 | if (frame_info.origin.x0 != 0 || frame_info.origin.y0 != 0 || |
404 | 52 | frame_header->frame_size.xsize != frame_header->default_xsize() || |
405 | 52 | frame_header->frame_size.ysize != frame_header->default_ysize()) { |
406 | 0 | frame_header->custom_size_or_origin = true; |
407 | 0 | } |
408 | 52 | } |
409 | | // Upsampling. |
410 | 52 | frame_header->upsampling = cparams.resampling; |
411 | 52 | const std::vector<ExtraChannelInfo>& extra_channels = |
412 | 52 | frame_header->nonserialized_metadata->m.extra_channel_info; |
413 | 52 | frame_header->extra_channel_upsampling.clear(); |
414 | 52 | frame_header->extra_channel_upsampling.resize(extra_channels.size(), |
415 | 52 | cparams.ec_resampling); |
416 | 52 | frame_header->save_as_reference = frame_info.save_as_reference; |
417 | | |
418 | | // Set blending-related information. |
419 | 52 | if (frame_info.blend || frame_header->custom_size_or_origin) { |
420 | | // Set blend_channel to the first alpha channel. These values are only |
421 | | // encoded in case a blend mode involving alpha is used and there are more |
422 | | // than one extra channels. |
423 | 0 | size_t index = 0; |
424 | 0 | if (frame_info.alpha_channel == -1) { |
425 | 0 | if (extra_channels.size() > 1) { |
426 | 0 | for (size_t i = 0; i < extra_channels.size(); i++) { |
427 | 0 | if (extra_channels[i].type == ExtraChannel::kAlpha) { |
428 | 0 | index = i; |
429 | 0 | break; |
430 | 0 | } |
431 | 0 | } |
432 | 0 | } |
433 | 0 | } else { |
434 | 0 | index = static_cast<size_t>(frame_info.alpha_channel); |
435 | 0 | JXL_ASSERT(index == 0 || index < extra_channels.size()); |
436 | 0 | } |
437 | 0 | frame_header->blending_info.alpha_channel = index; |
438 | 0 | frame_header->blending_info.mode = |
439 | 0 | frame_info.blend ? frame_info.blendmode : BlendMode::kReplace; |
440 | 0 | frame_header->blending_info.source = frame_info.source; |
441 | 0 | frame_header->blending_info.clamp = frame_info.clamp; |
442 | 0 | const auto& extra_channel_info = frame_info.extra_channel_blending_info; |
443 | 0 | for (size_t i = 0; i < extra_channels.size(); i++) { |
444 | 0 | if (i < extra_channel_info.size()) { |
445 | 0 | frame_header->extra_channel_blending_info[i] = extra_channel_info[i]; |
446 | 0 | } else { |
447 | 0 | frame_header->extra_channel_blending_info[i].alpha_channel = index; |
448 | 0 | BlendMode default_blend = frame_info.blendmode; |
449 | 0 | if (extra_channels[i].type != ExtraChannel::kBlack && i != index) { |
450 | | // K needs to be blended, spot colors and other stuff gets added |
451 | 0 | default_blend = BlendMode::kAdd; |
452 | 0 | } |
453 | 0 | frame_header->extra_channel_blending_info[i].mode = |
454 | 0 | frame_info.blend ? default_blend : BlendMode::kReplace; |
455 | 0 | frame_header->extra_channel_blending_info[i].source = 1; |
456 | 0 | } |
457 | 0 | } |
458 | 0 | } |
459 | | |
460 | 52 | frame_header->animation_frame.duration = frame_info.duration; |
461 | 52 | frame_header->animation_frame.timecode = frame_info.timecode; |
462 | | |
463 | 52 | if (jpeg_data) { |
464 | 0 | frame_header->UpdateFlag(false, FrameHeader::kUseDcFrame); |
465 | 0 | frame_header->UpdateFlag(true, FrameHeader::kSkipAdaptiveDCSmoothing); |
466 | 0 | } |
467 | | |
468 | 52 | return true; |
469 | 52 | } |
470 | | |
471 | | // Invisible (alpha = 0) pixels tend to be a mess in optimized PNGs. |
472 | | // Since they have no visual impact whatsoever, we can replace them with |
473 | | // something that compresses better and reduces artifacts near the edges. This |
474 | | // does some kind of smooth stuff that seems to work. |
475 | | // Replace invisible pixels with a weighted average of the pixel to the left, |
476 | | // the pixel to the topright, and non-invisible neighbours. |
477 | | // Produces downward-blurry smears, with in the upwards direction only a 1px |
478 | | // edge duplication but not more. It would probably be better to smear in all |
479 | | // directions. That requires an alpha-weighed convolution with a large enough |
480 | | // kernel though, which might be overkill... |
481 | 0 | void SimplifyInvisible(Image3F* image, const ImageF& alpha, bool lossless) { |
482 | 0 | for (size_t c = 0; c < 3; ++c) { |
483 | 0 | for (size_t y = 0; y < image->ysize(); ++y) { |
484 | 0 | float* JXL_RESTRICT row = image->PlaneRow(c, y); |
485 | 0 | const float* JXL_RESTRICT prow = |
486 | 0 | (y > 0 ? image->PlaneRow(c, y - 1) : nullptr); |
487 | 0 | const float* JXL_RESTRICT nrow = |
488 | 0 | (y + 1 < image->ysize() ? image->PlaneRow(c, y + 1) : nullptr); |
489 | 0 | const float* JXL_RESTRICT a = alpha.Row(y); |
490 | 0 | const float* JXL_RESTRICT pa = (y > 0 ? alpha.Row(y - 1) : nullptr); |
491 | 0 | const float* JXL_RESTRICT na = |
492 | 0 | (y + 1 < image->ysize() ? alpha.Row(y + 1) : nullptr); |
493 | 0 | for (size_t x = 0; x < image->xsize(); ++x) { |
494 | 0 | if (a[x] == 0) { |
495 | 0 | if (lossless) { |
496 | 0 | row[x] = 0; |
497 | 0 | continue; |
498 | 0 | } |
499 | 0 | float d = 0.f; |
500 | 0 | row[x] = 0; |
501 | 0 | if (x > 0) { |
502 | 0 | row[x] += row[x - 1]; |
503 | 0 | d++; |
504 | 0 | if (a[x - 1] > 0.f) { |
505 | 0 | row[x] += row[x - 1]; |
506 | 0 | d++; |
507 | 0 | } |
508 | 0 | } |
509 | 0 | if (x + 1 < image->xsize()) { |
510 | 0 | if (y > 0) { |
511 | 0 | row[x] += prow[x + 1]; |
512 | 0 | d++; |
513 | 0 | } |
514 | 0 | if (a[x + 1] > 0.f) { |
515 | 0 | row[x] += 2.f * row[x + 1]; |
516 | 0 | d += 2.f; |
517 | 0 | } |
518 | 0 | if (y > 0 && pa[x + 1] > 0.f) { |
519 | 0 | row[x] += 2.f * prow[x + 1]; |
520 | 0 | d += 2.f; |
521 | 0 | } |
522 | 0 | if (y + 1 < image->ysize() && na[x + 1] > 0.f) { |
523 | 0 | row[x] += 2.f * nrow[x + 1]; |
524 | 0 | d += 2.f; |
525 | 0 | } |
526 | 0 | } |
527 | 0 | if (y > 0 && pa[x] > 0.f) { |
528 | 0 | row[x] += 2.f * prow[x]; |
529 | 0 | d += 2.f; |
530 | 0 | } |
531 | 0 | if (y + 1 < image->ysize() && na[x] > 0.f) { |
532 | 0 | row[x] += 2.f * nrow[x]; |
533 | 0 | d += 2.f; |
534 | 0 | } |
535 | 0 | if (d > 1.f) row[x] /= d; |
536 | 0 | } |
537 | 0 | } |
538 | 0 | } |
539 | 0 | } |
540 | 0 | } |
541 | | |
542 | | struct PixelStatsForChromacityAdjustment { |
543 | | float dx = 0; |
544 | | float db = 0; |
545 | | float exposed_blue = 0; |
546 | 52 | static float CalcPlane(const ImageF* JXL_RESTRICT plane, const Rect& rect) { |
547 | 52 | float xmax = 0; |
548 | 52 | float ymax = 0; |
549 | 416 | for (size_t ty = 1; ty < rect.ysize(); ++ty) { |
550 | 2.91k | for (size_t tx = 1; tx < rect.xsize(); ++tx) { |
551 | 2.54k | float cur = rect.Row(plane, ty)[tx]; |
552 | 2.54k | float prev_row = rect.Row(plane, ty - 1)[tx]; |
553 | 2.54k | float prev = rect.Row(plane, ty)[tx - 1]; |
554 | 2.54k | xmax = std::max(xmax, std::abs(cur - prev)); |
555 | 2.54k | ymax = std::max(ymax, std::abs(cur - prev_row)); |
556 | 2.54k | } |
557 | 364 | } |
558 | 52 | return std::max(xmax, ymax); |
559 | 52 | } |
560 | | void CalcExposedBlue(const ImageF* JXL_RESTRICT plane_y, |
561 | 52 | const ImageF* JXL_RESTRICT plane_b, const Rect& rect) { |
562 | 52 | float eb = 0; |
563 | 52 | float xmax = 0; |
564 | 52 | float ymax = 0; |
565 | 416 | for (size_t ty = 1; ty < rect.ysize(); ++ty) { |
566 | 2.91k | for (size_t tx = 1; tx < rect.xsize(); ++tx) { |
567 | 2.54k | float cur_y = rect.Row(plane_y, ty)[tx]; |
568 | 2.54k | float cur_b = rect.Row(plane_b, ty)[tx]; |
569 | 2.54k | float exposed_b = cur_b - cur_y * 1.2; |
570 | 2.54k | float diff_b = cur_b - cur_y; |
571 | 2.54k | float prev_row = rect.Row(plane_b, ty - 1)[tx]; |
572 | 2.54k | float prev = rect.Row(plane_b, ty)[tx - 1]; |
573 | 2.54k | float diff_prev_row = prev_row - rect.Row(plane_y, ty - 1)[tx]; |
574 | 2.54k | float diff_prev = prev - rect.Row(plane_y, ty)[tx - 1]; |
575 | 2.54k | xmax = std::max(xmax, std::abs(diff_b - diff_prev)); |
576 | 2.54k | ymax = std::max(ymax, std::abs(diff_b - diff_prev_row)); |
577 | 2.54k | if (exposed_b >= 0) { |
578 | 0 | exposed_b *= fabs(cur_b - prev) + fabs(cur_b - prev_row); |
579 | 0 | eb = std::max(eb, exposed_b); |
580 | 0 | } |
581 | 2.54k | } |
582 | 364 | } |
583 | 52 | exposed_blue = eb; |
584 | 52 | db = std::max(xmax, ymax); |
585 | 52 | } |
586 | 52 | void Calc(const Image3F* JXL_RESTRICT opsin, const Rect& rect) { |
587 | 52 | dx = CalcPlane(&opsin->Plane(0), rect); |
588 | 52 | CalcExposedBlue(&opsin->Plane(1), &opsin->Plane(2), rect); |
589 | 52 | } |
590 | 52 | int HowMuchIsXChannelPixelized() const { |
591 | 52 | if (dx >= 0.026) { |
592 | 0 | return 3; |
593 | 0 | } |
594 | 52 | if (dx >= 0.022) { |
595 | 0 | return 2; |
596 | 0 | } |
597 | 52 | if (dx >= 0.015) { |
598 | 0 | return 1; |
599 | 0 | } |
600 | 52 | return 0; |
601 | 52 | } |
602 | 52 | int HowMuchIsBChannelPixelized() const { |
603 | 52 | int add = exposed_blue >= 0.13 ? 1 : 0; |
604 | 52 | if (db > 0.38) { |
605 | 0 | return 2 + add; |
606 | 0 | } |
607 | 52 | if (db > 0.33) { |
608 | 0 | return 1 + add; |
609 | 0 | } |
610 | 52 | if (db > 0.28) { |
611 | 0 | return add; |
612 | 0 | } |
613 | 52 | return 0; |
614 | 52 | } |
615 | | }; |
616 | | |
617 | | void ComputeChromacityAdjustments(const CompressParams& cparams, |
618 | | const Image3F& opsin, const Rect& rect, |
619 | 52 | FrameHeader* frame_header) { |
620 | 52 | if (frame_header->encoding != FrameEncoding::kVarDCT || |
621 | 52 | cparams.max_error_mode) { |
622 | 0 | return; |
623 | 0 | } |
624 | | // 1) Distance based approach for chromacity adjustment: |
625 | 52 | float x_qm_scale_steps[3] = {2.5f, 5.5f, 9.5f}; |
626 | 52 | frame_header->x_qm_scale = 3; |
627 | 156 | for (float x_qm_scale_step : x_qm_scale_steps) { |
628 | 156 | if (cparams.original_butteraugli_distance > x_qm_scale_step) { |
629 | 0 | frame_header->x_qm_scale++; |
630 | 0 | } |
631 | 156 | } |
632 | | // 2) Pixel-based approach for chromacity adjustment: |
633 | | // look at the individual pixels and make a guess how difficult |
634 | | // the image would be based on the worst case pixel. |
635 | 52 | PixelStatsForChromacityAdjustment pixel_stats; |
636 | 52 | if (cparams.speed_tier <= SpeedTier::kSquirrel) { |
637 | 52 | pixel_stats.Calc(&opsin, rect); |
638 | 52 | } |
639 | | // For X take the most severe adjustment. |
640 | 52 | frame_header->x_qm_scale = std::max<int>( |
641 | 52 | frame_header->x_qm_scale, 2 + pixel_stats.HowMuchIsXChannelPixelized()); |
642 | | // B only adjusted by pixel-based approach. |
643 | 52 | frame_header->b_qm_scale = 2 + pixel_stats.HowMuchIsBChannelPixelized(); |
644 | 52 | } |
645 | | |
646 | | void ComputeNoiseParams(const CompressParams& cparams, bool streaming_mode, |
647 | | bool color_is_jpeg, const Image3F& opsin, |
648 | | const FrameDimensions& frame_dim, |
649 | 52 | FrameHeader* frame_header, NoiseParams* noise_params) { |
650 | 52 | if (cparams.photon_noise_iso > 0) { |
651 | 0 | *noise_params = SimulatePhotonNoise(frame_dim.xsize, frame_dim.ysize, |
652 | 0 | cparams.photon_noise_iso); |
653 | 52 | } else if (cparams.manual_noise.size() == NoiseParams::kNumNoisePoints) { |
654 | 0 | for (size_t i = 0; i < NoiseParams::kNumNoisePoints; i++) { |
655 | 0 | noise_params->lut[i] = cparams.manual_noise[i]; |
656 | 0 | } |
657 | 52 | } else if (frame_header->encoding == FrameEncoding::kVarDCT && |
658 | 52 | frame_header->flags & FrameHeader::kNoise && !color_is_jpeg && |
659 | 52 | !streaming_mode) { |
660 | | // Don't start at zero amplitude since adding noise is expensive -- it |
661 | | // significantly slows down decoding, and this is unlikely to |
662 | | // completely go away even with advanced optimizations. After the |
663 | | // kNoiseModelingRampUpDistanceRange we have reached the full level, |
664 | | // i.e. noise is no longer represented by the compressed image, so we |
665 | | // can add full noise by the noise modeling itself. |
666 | 0 | static const float kNoiseModelingRampUpDistanceRange = 0.6; |
667 | 0 | static const float kNoiseLevelAtStartOfRampUp = 0.25; |
668 | 0 | static const float kNoiseRampupStart = 1.0; |
669 | | // TODO(user) test and properly select quality_coef with smooth |
670 | | // filter |
671 | 0 | float quality_coef = 1.0f; |
672 | 0 | const float rampup = (cparams.butteraugli_distance - kNoiseRampupStart) / |
673 | 0 | kNoiseModelingRampUpDistanceRange; |
674 | 0 | if (rampup < 1.0f) { |
675 | 0 | quality_coef = kNoiseLevelAtStartOfRampUp + |
676 | 0 | (1.0f - kNoiseLevelAtStartOfRampUp) * rampup; |
677 | 0 | } |
678 | 0 | if (rampup < 0.0f) { |
679 | 0 | quality_coef = kNoiseRampupStart; |
680 | 0 | } |
681 | 0 | if (!GetNoiseParameter(opsin, noise_params, quality_coef)) { |
682 | 0 | frame_header->flags &= ~FrameHeader::kNoise; |
683 | 0 | } |
684 | 0 | } |
685 | 52 | } |
686 | | |
687 | | Status DownsampleColorChannels(const CompressParams& cparams, |
688 | | const FrameHeader& frame_header, |
689 | 52 | bool color_is_jpeg, Image3F* opsin) { |
690 | 52 | if (color_is_jpeg || frame_header.upsampling == 1 || |
691 | 52 | cparams.already_downsampled) { |
692 | 52 | return true; |
693 | 52 | } |
694 | 0 | if (frame_header.encoding == FrameEncoding::kVarDCT && |
695 | 0 | frame_header.upsampling == 2) { |
696 | | // TODO(lode): use the regular DownsampleImage, or adapt to the custom |
697 | | // coefficients, if there is are custom upscaling coefficients in |
698 | | // CustomTransformData |
699 | 0 | if (cparams.speed_tier <= SpeedTier::kSquirrel) { |
700 | | // TODO(lode): DownsampleImage2_Iterative is currently too slow to |
701 | | // be used for squirrel, make it faster, and / or enable it only for |
702 | | // kitten. |
703 | 0 | JXL_RETURN_IF_ERROR(DownsampleImage2_Iterative(opsin)); |
704 | 0 | } else { |
705 | 0 | JXL_RETURN_IF_ERROR(DownsampleImage2_Sharper(opsin)); |
706 | 0 | } |
707 | 0 | } else { |
708 | 0 | JXL_ASSIGN_OR_RETURN(*opsin, |
709 | 0 | DownsampleImage(*opsin, frame_header.upsampling)); |
710 | 0 | } |
711 | 0 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
712 | 0 | PadImageToBlockMultipleInPlace(opsin); |
713 | 0 | } |
714 | 0 | return true; |
715 | 0 | } |
716 | | |
717 | | template <typename V, typename R> |
718 | 0 | void FindIndexOfSumMaximum(const V* array, const size_t len, R* idx, V* sum) { |
719 | 0 | JXL_ASSERT(len > 0); |
720 | 0 | V maxval = 0; |
721 | 0 | V val = 0; |
722 | 0 | R maxidx = 0; |
723 | 0 | for (size_t i = 0; i < len; ++i) { |
724 | 0 | val += array[i]; |
725 | 0 | if (val > maxval) { |
726 | 0 | maxval = val; |
727 | 0 | maxidx = i; |
728 | 0 | } |
729 | 0 | } |
730 | 0 | *idx = maxidx; |
731 | 0 | *sum = maxval; |
732 | 0 | } |
733 | | |
734 | | Status ComputeJPEGTranscodingData(const jpeg::JPEGData& jpeg_data, |
735 | | const FrameHeader& frame_header, |
736 | | ThreadPool* pool, |
737 | | ModularFrameEncoder* enc_modular, |
738 | 0 | PassesEncoderState* enc_state) { |
739 | 0 | PassesSharedState& shared = enc_state->shared; |
740 | 0 | JxlMemoryManager* memory_manager = enc_state->memory_manager(); |
741 | 0 | const FrameDimensions& frame_dim = shared.frame_dim; |
742 | |
|
743 | 0 | const size_t xsize = frame_dim.xsize_padded; |
744 | 0 | const size_t ysize = frame_dim.ysize_padded; |
745 | 0 | const size_t xsize_blocks = frame_dim.xsize_blocks; |
746 | 0 | const size_t ysize_blocks = frame_dim.ysize_blocks; |
747 | | |
748 | | // no-op chroma from luma |
749 | 0 | JXL_ASSIGN_OR_RETURN(shared.cmap, ColorCorrelationMap::Create( |
750 | 0 | memory_manager, xsize, ysize, false)); |
751 | 0 | shared.ac_strategy.FillDCT8(); |
752 | 0 | FillImage(static_cast<uint8_t>(0), &shared.epf_sharpness); |
753 | |
|
754 | 0 | enc_state->coeffs.clear(); |
755 | 0 | while (enc_state->coeffs.size() < enc_state->passes.size()) { |
756 | 0 | JXL_ASSIGN_OR_RETURN( |
757 | 0 | std::unique_ptr<ACImageT<int32_t>> coeffs, |
758 | 0 | ACImageT<int32_t>::Make(memory_manager, kGroupDim * kGroupDim, |
759 | 0 | frame_dim.num_groups)); |
760 | 0 | enc_state->coeffs.emplace_back(std::move(coeffs)); |
761 | 0 | } |
762 | | |
763 | | // convert JPEG quantization table to a Quantizer object |
764 | 0 | float dcquantization[3]; |
765 | 0 | std::vector<QuantEncoding> qe(DequantMatrices::kNum, |
766 | 0 | QuantEncoding::Library(0)); |
767 | |
|
768 | 0 | auto jpeg_c_map = |
769 | 0 | JpegOrder(frame_header.color_transform, jpeg_data.components.size() == 1); |
770 | |
|
771 | 0 | std::vector<int> qt(192); |
772 | 0 | for (size_t c = 0; c < 3; c++) { |
773 | 0 | size_t jpeg_c = jpeg_c_map[c]; |
774 | 0 | const int32_t* quant = |
775 | 0 | jpeg_data.quant[jpeg_data.components[jpeg_c].quant_idx].values.data(); |
776 | |
|
777 | 0 | dcquantization[c] = 255 * 8.0f / quant[0]; |
778 | 0 | for (size_t y = 0; y < 8; y++) { |
779 | 0 | for (size_t x = 0; x < 8; x++) { |
780 | | // JPEG XL transposes the DCT, JPEG doesn't. |
781 | 0 | qt[c * 64 + 8 * x + y] = quant[8 * y + x]; |
782 | 0 | } |
783 | 0 | } |
784 | 0 | } |
785 | 0 | DequantMatricesSetCustomDC(memory_manager, &shared.matrices, dcquantization); |
786 | 0 | float dcquantization_r[3] = {1.0f / dcquantization[0], |
787 | 0 | 1.0f / dcquantization[1], |
788 | 0 | 1.0f / dcquantization[2]}; |
789 | |
|
790 | 0 | qe[AcStrategy::Type::DCT] = QuantEncoding::RAW(qt); |
791 | 0 | JXL_RETURN_IF_ERROR( |
792 | 0 | DequantMatricesSetCustom(&shared.matrices, qe, enc_modular)); |
793 | | |
794 | | // Ensure that InvGlobalScale() is 1. |
795 | 0 | shared.quantizer = Quantizer(&shared.matrices, 1, kGlobalScaleDenom); |
796 | | // Recompute MulDC() and InvMulDC(). |
797 | 0 | shared.quantizer.RecomputeFromGlobalScale(); |
798 | | |
799 | | // Per-block dequant scaling should be 1. |
800 | 0 | FillImage(static_cast<int32_t>(shared.quantizer.InvGlobalScale()), |
801 | 0 | &shared.raw_quant_field); |
802 | |
|
803 | 0 | std::vector<int32_t> scaled_qtable(192); |
804 | 0 | for (size_t c = 0; c < 3; c++) { |
805 | 0 | for (size_t i = 0; i < 64; i++) { |
806 | 0 | scaled_qtable[64 * c + i] = |
807 | 0 | (1 << kCFLFixedPointPrecision) * qt[64 + i] / qt[64 * c + i]; |
808 | 0 | } |
809 | 0 | } |
810 | |
|
811 | 0 | auto jpeg_row = [&](size_t c, size_t y) { |
812 | 0 | return jpeg_data.components[jpeg_c_map[c]].coeffs.data() + |
813 | 0 | jpeg_data.components[jpeg_c_map[c]].width_in_blocks * kDCTBlockSize * |
814 | 0 | y; |
815 | 0 | }; |
816 | |
|
817 | 0 | bool DCzero = (frame_header.color_transform == ColorTransform::kYCbCr); |
818 | | // Compute chroma-from-luma for AC (doesn't seem to be useful for DC) |
819 | 0 | if (frame_header.chroma_subsampling.Is444() && |
820 | 0 | enc_state->cparams.force_cfl_jpeg_recompression && |
821 | 0 | jpeg_data.components.size() == 3) { |
822 | 0 | for (size_t c : {0, 2}) { |
823 | 0 | ImageSB* map = (c == 0 ? &shared.cmap.ytox_map : &shared.cmap.ytob_map); |
824 | 0 | const float kScale = kDefaultColorFactor; |
825 | 0 | const int kOffset = 127; |
826 | 0 | const float kBase = c == 0 ? shared.cmap.base().YtoXRatio(0) |
827 | 0 | : shared.cmap.base().YtoBRatio(0); |
828 | 0 | const float kZeroThresh = |
829 | 0 | kScale * kZeroBiasDefault[c] * |
830 | 0 | 0.9999f; // just epsilon less for better rounding |
831 | |
|
832 | 0 | auto process_row = [&](const uint32_t task, const size_t thread) { |
833 | 0 | size_t ty = task; |
834 | 0 | int8_t* JXL_RESTRICT row_out = map->Row(ty); |
835 | 0 | for (size_t tx = 0; tx < map->xsize(); ++tx) { |
836 | 0 | const size_t y0 = ty * kColorTileDimInBlocks; |
837 | 0 | const size_t x0 = tx * kColorTileDimInBlocks; |
838 | 0 | const size_t y1 = std::min(frame_dim.ysize_blocks, |
839 | 0 | (ty + 1) * kColorTileDimInBlocks); |
840 | 0 | const size_t x1 = std::min(frame_dim.xsize_blocks, |
841 | 0 | (tx + 1) * kColorTileDimInBlocks); |
842 | 0 | int32_t d_num_zeros[257] = {0}; |
843 | | // TODO(veluca): this needs SIMD + fixed point adaptation, and/or |
844 | | // conversion to the new CfL algorithm. |
845 | 0 | for (size_t y = y0; y < y1; ++y) { |
846 | 0 | const int16_t* JXL_RESTRICT row_m = jpeg_row(1, y); |
847 | 0 | const int16_t* JXL_RESTRICT row_s = jpeg_row(c, y); |
848 | 0 | for (size_t x = x0; x < x1; ++x) { |
849 | 0 | for (size_t coeffpos = 1; coeffpos < kDCTBlockSize; coeffpos++) { |
850 | 0 | const float scaled_m = row_m[x * kDCTBlockSize + coeffpos] * |
851 | 0 | scaled_qtable[64 * c + coeffpos] * |
852 | 0 | (1.0f / (1 << kCFLFixedPointPrecision)); |
853 | 0 | const float scaled_s = |
854 | 0 | kScale * row_s[x * kDCTBlockSize + coeffpos] + |
855 | 0 | (kOffset - kBase * kScale) * scaled_m; |
856 | 0 | if (std::abs(scaled_m) > 1e-8f) { |
857 | 0 | float from; |
858 | 0 | float to; |
859 | 0 | if (scaled_m > 0) { |
860 | 0 | from = (scaled_s - kZeroThresh) / scaled_m; |
861 | 0 | to = (scaled_s + kZeroThresh) / scaled_m; |
862 | 0 | } else { |
863 | 0 | from = (scaled_s + kZeroThresh) / scaled_m; |
864 | 0 | to = (scaled_s - kZeroThresh) / scaled_m; |
865 | 0 | } |
866 | 0 | if (from < 0.0f) { |
867 | 0 | from = 0.0f; |
868 | 0 | } |
869 | 0 | if (to > 255.0f) { |
870 | 0 | to = 255.0f; |
871 | 0 | } |
872 | | // Instead of clamping the both values |
873 | | // we just check that range is sane. |
874 | 0 | if (from <= to) { |
875 | 0 | d_num_zeros[static_cast<int>(std::ceil(from))]++; |
876 | 0 | d_num_zeros[static_cast<int>(std::floor(to + 1))]--; |
877 | 0 | } |
878 | 0 | } |
879 | 0 | } |
880 | 0 | } |
881 | 0 | } |
882 | 0 | int best = 0; |
883 | 0 | int32_t best_sum = 0; |
884 | 0 | FindIndexOfSumMaximum(d_num_zeros, 256, &best, &best_sum); |
885 | 0 | int32_t offset_sum = 0; |
886 | 0 | for (int i = 0; i < 256; ++i) { |
887 | 0 | if (i <= kOffset) { |
888 | 0 | offset_sum += d_num_zeros[i]; |
889 | 0 | } |
890 | 0 | } |
891 | 0 | row_out[tx] = 0; |
892 | 0 | if (best_sum > offset_sum + 1) { |
893 | 0 | row_out[tx] = best - kOffset; |
894 | 0 | } |
895 | 0 | } |
896 | 0 | }; |
897 | |
|
898 | 0 | JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, map->ysize(), ThreadPool::NoInit, |
899 | 0 | process_row, "FindCorrelation")); |
900 | 0 | } |
901 | 0 | } |
902 | | |
903 | 0 | JXL_ASSIGN_OR_RETURN( |
904 | 0 | Image3F dc, Image3F::Create(memory_manager, xsize_blocks, ysize_blocks)); |
905 | 0 | if (!frame_header.chroma_subsampling.Is444()) { |
906 | 0 | ZeroFillImage(&dc); |
907 | 0 | for (auto& coeff : enc_state->coeffs) { |
908 | 0 | coeff->ZeroFill(); |
909 | 0 | } |
910 | 0 | } |
911 | | // JPEG DC is from -1024 to 1023. |
912 | 0 | std::vector<size_t> dc_counts[3] = {}; |
913 | 0 | dc_counts[0].resize(2048); |
914 | 0 | dc_counts[1].resize(2048); |
915 | 0 | dc_counts[2].resize(2048); |
916 | 0 | size_t total_dc[3] = {}; |
917 | 0 | for (size_t c : {1, 0, 2}) { |
918 | 0 | if (jpeg_data.components.size() == 1 && c != 1) { |
919 | 0 | for (auto& coeff : enc_state->coeffs) { |
920 | 0 | coeff->ZeroFillPlane(c); |
921 | 0 | } |
922 | 0 | ZeroFillImage(&dc.Plane(c)); |
923 | | // Ensure no division by 0. |
924 | 0 | dc_counts[c][1024] = 1; |
925 | 0 | total_dc[c] = 1; |
926 | 0 | continue; |
927 | 0 | } |
928 | 0 | size_t hshift = frame_header.chroma_subsampling.HShift(c); |
929 | 0 | size_t vshift = frame_header.chroma_subsampling.VShift(c); |
930 | 0 | ImageSB& map = (c == 0 ? shared.cmap.ytox_map : shared.cmap.ytob_map); |
931 | 0 | for (size_t group_index = 0; group_index < frame_dim.num_groups; |
932 | 0 | group_index++) { |
933 | 0 | const size_t gx = group_index % frame_dim.xsize_groups; |
934 | 0 | const size_t gy = group_index / frame_dim.xsize_groups; |
935 | 0 | int32_t* coeffs[kMaxNumPasses]; |
936 | 0 | for (size_t i = 0; i < enc_state->coeffs.size(); i++) { |
937 | 0 | coeffs[i] = enc_state->coeffs[i]->PlaneRow(c, group_index, 0).ptr32; |
938 | 0 | } |
939 | 0 | int32_t block[64]; |
940 | 0 | for (size_t by = gy * kGroupDimInBlocks; |
941 | 0 | by < ysize_blocks && by < (gy + 1) * kGroupDimInBlocks; ++by) { |
942 | 0 | if ((by >> vshift) << vshift != by) continue; |
943 | 0 | const int16_t* JXL_RESTRICT inputjpeg = jpeg_row(c, by >> vshift); |
944 | 0 | const int16_t* JXL_RESTRICT inputjpegY = jpeg_row(1, by); |
945 | 0 | float* JXL_RESTRICT fdc = dc.PlaneRow(c, by >> vshift); |
946 | 0 | const int8_t* JXL_RESTRICT cm = |
947 | 0 | map.ConstRow(by / kColorTileDimInBlocks); |
948 | 0 | for (size_t bx = gx * kGroupDimInBlocks; |
949 | 0 | bx < xsize_blocks && bx < (gx + 1) * kGroupDimInBlocks; ++bx) { |
950 | 0 | if ((bx >> hshift) << hshift != bx) continue; |
951 | 0 | size_t base = (bx >> hshift) * kDCTBlockSize; |
952 | 0 | int idc; |
953 | 0 | if (DCzero) { |
954 | 0 | idc = inputjpeg[base]; |
955 | 0 | } else { |
956 | 0 | idc = inputjpeg[base] + 1024 / qt[c * 64]; |
957 | 0 | } |
958 | 0 | dc_counts[c][std::min(static_cast<uint32_t>(idc + 1024), |
959 | 0 | static_cast<uint32_t>(2047))]++; |
960 | 0 | total_dc[c]++; |
961 | 0 | fdc[bx >> hshift] = idc * dcquantization_r[c]; |
962 | 0 | if (c == 1 || !enc_state->cparams.force_cfl_jpeg_recompression || |
963 | 0 | !frame_header.chroma_subsampling.Is444()) { |
964 | 0 | for (size_t y = 0; y < 8; y++) { |
965 | 0 | for (size_t x = 0; x < 8; x++) { |
966 | 0 | block[y * 8 + x] = inputjpeg[base + x * 8 + y]; |
967 | 0 | } |
968 | 0 | } |
969 | 0 | } else { |
970 | 0 | const int32_t scale = |
971 | 0 | ColorCorrelation::RatioJPEG(cm[bx / kColorTileDimInBlocks]); |
972 | |
|
973 | 0 | for (size_t y = 0; y < 8; y++) { |
974 | 0 | for (size_t x = 0; x < 8; x++) { |
975 | 0 | int Y = inputjpegY[kDCTBlockSize * bx + x * 8 + y]; |
976 | 0 | int QChroma = inputjpeg[kDCTBlockSize * bx + x * 8 + y]; |
977 | | // Fixed-point multiply of CfL scale with quant table ratio |
978 | | // first, and Y value second. |
979 | 0 | int coeff_scale = (scale * scaled_qtable[64 * c + y * 8 + x] + |
980 | 0 | (1 << (kCFLFixedPointPrecision - 1))) >> |
981 | 0 | kCFLFixedPointPrecision; |
982 | 0 | int cfl_factor = |
983 | 0 | (Y * coeff_scale + (1 << (kCFLFixedPointPrecision - 1))) >> |
984 | 0 | kCFLFixedPointPrecision; |
985 | 0 | int QCR = QChroma - cfl_factor; |
986 | 0 | block[y * 8 + x] = QCR; |
987 | 0 | } |
988 | 0 | } |
989 | 0 | } |
990 | 0 | enc_state->progressive_splitter.SplitACCoefficients( |
991 | 0 | block, AcStrategy::FromRawStrategy(AcStrategy::Type::DCT), bx, by, |
992 | 0 | coeffs); |
993 | 0 | for (size_t i = 0; i < enc_state->coeffs.size(); i++) { |
994 | 0 | coeffs[i] += kDCTBlockSize; |
995 | 0 | } |
996 | 0 | } |
997 | 0 | } |
998 | 0 | } |
999 | 0 | } |
1000 | |
|
1001 | 0 | auto& dct = enc_state->shared.block_ctx_map.dc_thresholds; |
1002 | 0 | auto& num_dc_ctxs = enc_state->shared.block_ctx_map.num_dc_ctxs; |
1003 | 0 | num_dc_ctxs = 1; |
1004 | 0 | for (size_t i = 0; i < 3; i++) { |
1005 | 0 | dct[i].clear(); |
1006 | 0 | int num_thresholds = (CeilLog2Nonzero(total_dc[i]) - 12) / 2; |
1007 | | // up to 3 buckets per channel: |
1008 | | // dark/medium/bright, yellow/unsat/blue, green/unsat/red |
1009 | 0 | num_thresholds = std::min(std::max(num_thresholds, 0), 2); |
1010 | 0 | size_t cumsum = 0; |
1011 | 0 | size_t cut = total_dc[i] / (num_thresholds + 1); |
1012 | 0 | for (int j = 0; j < 2048; j++) { |
1013 | 0 | cumsum += dc_counts[i][j]; |
1014 | 0 | if (cumsum > cut) { |
1015 | 0 | dct[i].push_back(j - 1025); |
1016 | 0 | cut = total_dc[i] * (dct[i].size() + 1) / (num_thresholds + 1); |
1017 | 0 | } |
1018 | 0 | } |
1019 | 0 | num_dc_ctxs *= dct[i].size() + 1; |
1020 | 0 | } |
1021 | |
|
1022 | 0 | auto& ctx_map = enc_state->shared.block_ctx_map.ctx_map; |
1023 | 0 | ctx_map.clear(); |
1024 | 0 | ctx_map.resize(3 * kNumOrders * num_dc_ctxs, 0); |
1025 | |
|
1026 | 0 | int lbuckets = (dct[1].size() + 1); |
1027 | 0 | for (size_t i = 0; i < num_dc_ctxs; i++) { |
1028 | | // up to 9 contexts for luma |
1029 | 0 | ctx_map[i] = i / lbuckets; |
1030 | | // up to 3 contexts for chroma |
1031 | 0 | ctx_map[kNumOrders * num_dc_ctxs + i] = |
1032 | 0 | ctx_map[2 * kNumOrders * num_dc_ctxs + i] = |
1033 | 0 | num_dc_ctxs / lbuckets + (i % lbuckets); |
1034 | 0 | } |
1035 | 0 | enc_state->shared.block_ctx_map.num_ctxs = |
1036 | 0 | *std::max_element(ctx_map.begin(), ctx_map.end()) + 1; |
1037 | | |
1038 | | // disable DC frame for now |
1039 | 0 | std::atomic<bool> has_error{false}; |
1040 | 0 | auto compute_dc_coeffs = [&](const uint32_t group_index, |
1041 | 0 | size_t /* thread */) { |
1042 | 0 | if (has_error) return; |
1043 | 0 | const Rect r = enc_state->shared.frame_dim.DCGroupRect(group_index); |
1044 | 0 | if (!enc_modular->AddVarDCTDC(frame_header, dc, r, group_index, |
1045 | 0 | /*nl_dc=*/false, enc_state, |
1046 | 0 | /*jpeg_transcode=*/true)) { |
1047 | 0 | has_error = true; |
1048 | 0 | return; |
1049 | 0 | } |
1050 | 0 | if (!enc_modular->AddACMetadata(r, group_index, /*jpeg_transcode=*/true, |
1051 | 0 | enc_state)) { |
1052 | 0 | has_error = true; |
1053 | 0 | return; |
1054 | 0 | } |
1055 | 0 | }; |
1056 | 0 | JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, shared.frame_dim.num_dc_groups, |
1057 | 0 | ThreadPool::NoInit, compute_dc_coeffs, |
1058 | 0 | "Compute DC coeffs")); |
1059 | 0 | if (has_error) return JXL_FAILURE("Compute DC coeffs failed"); |
1060 | | |
1061 | 0 | return true; |
1062 | 0 | } |
1063 | | |
1064 | | Status ComputeVarDCTEncodingData(const FrameHeader& frame_header, |
1065 | | const Image3F* linear, |
1066 | | Image3F* JXL_RESTRICT opsin, const Rect& rect, |
1067 | | const JxlCmsInterface& cms, ThreadPool* pool, |
1068 | | ModularFrameEncoder* enc_modular, |
1069 | | PassesEncoderState* enc_state, |
1070 | 52 | AuxOut* aux_out) { |
1071 | 52 | JXL_ASSERT((rect.xsize() % kBlockDim) == 0 && |
1072 | 52 | (rect.ysize() % kBlockDim) == 0); |
1073 | 52 | JxlMemoryManager* memory_manager = enc_state->memory_manager(); |
1074 | | // Save pre-Gaborish opsin for AR control field heuristics computation. |
1075 | 52 | Image3F orig_opsin; |
1076 | 52 | JXL_ASSIGN_OR_RETURN( |
1077 | 52 | orig_opsin, Image3F::Create(memory_manager, rect.xsize(), rect.ysize())); |
1078 | 52 | CopyImageTo(rect, *opsin, Rect(orig_opsin), &orig_opsin); |
1079 | 52 | orig_opsin.ShrinkTo(enc_state->shared.frame_dim.xsize, |
1080 | 52 | enc_state->shared.frame_dim.ysize); |
1081 | | |
1082 | 52 | JXL_RETURN_IF_ERROR(LossyFrameHeuristics(frame_header, enc_state, enc_modular, |
1083 | 52 | linear, opsin, rect, cms, pool, |
1084 | 52 | aux_out)); |
1085 | | |
1086 | 52 | JXL_RETURN_IF_ERROR(InitializePassesEncoder( |
1087 | 52 | frame_header, *opsin, rect, cms, pool, enc_state, enc_modular, aux_out)); |
1088 | | |
1089 | 52 | JXL_RETURN_IF_ERROR( |
1090 | 52 | ComputeARHeuristics(frame_header, enc_state, orig_opsin, rect, pool)); |
1091 | | |
1092 | 52 | JXL_RETURN_IF_ERROR(ComputeACMetadata(pool, enc_state, enc_modular)); |
1093 | | |
1094 | 52 | return true; |
1095 | 52 | } |
1096 | | |
1097 | | void ComputeAllCoeffOrders(PassesEncoderState& enc_state, |
1098 | 52 | const FrameDimensions& frame_dim) { |
1099 | 52 | auto used_orders_info = ComputeUsedOrders( |
1100 | 52 | enc_state.cparams.speed_tier, enc_state.shared.ac_strategy, |
1101 | 52 | Rect(enc_state.shared.raw_quant_field)); |
1102 | 52 | enc_state.used_orders.resize(enc_state.progressive_splitter.GetNumPasses()); |
1103 | 104 | for (size_t i = 0; i < enc_state.progressive_splitter.GetNumPasses(); i++) { |
1104 | 52 | ComputeCoeffOrder( |
1105 | 52 | enc_state.cparams.speed_tier, *enc_state.coeffs[i], |
1106 | 52 | enc_state.shared.ac_strategy, frame_dim, enc_state.used_orders[i], |
1107 | 52 | enc_state.used_acs, used_orders_info.first, used_orders_info.second, |
1108 | 52 | &enc_state.shared.coeff_orders[i * enc_state.shared.coeff_order_size]); |
1109 | 52 | } |
1110 | 52 | enc_state.used_acs |= used_orders_info.first; |
1111 | 52 | } |
1112 | | |
1113 | | // Working area for TokenizeCoefficients (per-group!) |
1114 | | struct EncCache { |
1115 | | // Allocates memory when first called. |
1116 | 52 | Status InitOnce(JxlMemoryManager* memory_manager) { |
1117 | 52 | if (num_nzeroes.xsize() == 0) { |
1118 | 52 | JXL_ASSIGN_OR_RETURN(num_nzeroes, |
1119 | 52 | Image3I::Create(memory_manager, kGroupDimInBlocks, |
1120 | 52 | kGroupDimInBlocks)); |
1121 | 52 | } |
1122 | 52 | return true; |
1123 | 52 | } |
1124 | | // TokenizeCoefficients |
1125 | | Image3I num_nzeroes; |
1126 | | }; |
1127 | | |
1128 | | Status TokenizeAllCoefficients(const FrameHeader& frame_header, |
1129 | | ThreadPool* pool, |
1130 | 52 | PassesEncoderState* enc_state) { |
1131 | 52 | PassesSharedState& shared = enc_state->shared; |
1132 | 52 | std::vector<EncCache> group_caches; |
1133 | 52 | JxlMemoryManager* memory_manager = enc_state->memory_manager(); |
1134 | 52 | const auto tokenize_group_init = [&](const size_t num_threads) { |
1135 | 52 | group_caches.resize(num_threads); |
1136 | 52 | return true; |
1137 | 52 | }; |
1138 | 52 | std::atomic<bool> has_error{false}; |
1139 | 52 | const auto tokenize_group = [&](const uint32_t group_index, |
1140 | 52 | const size_t thread) { |
1141 | 52 | if (has_error) return; |
1142 | | // Tokenize coefficients. |
1143 | 52 | const Rect rect = shared.frame_dim.BlockGroupRect(group_index); |
1144 | 104 | for (size_t idx_pass = 0; idx_pass < enc_state->passes.size(); idx_pass++) { |
1145 | 52 | JXL_ASSERT(enc_state->coeffs[idx_pass]->Type() == ACType::k32); |
1146 | 52 | const int32_t* JXL_RESTRICT ac_rows[3] = { |
1147 | 52 | enc_state->coeffs[idx_pass]->PlaneRow(0, group_index, 0).ptr32, |
1148 | 52 | enc_state->coeffs[idx_pass]->PlaneRow(1, group_index, 0).ptr32, |
1149 | 52 | enc_state->coeffs[idx_pass]->PlaneRow(2, group_index, 0).ptr32, |
1150 | 52 | }; |
1151 | | // Ensure group cache is initialized. |
1152 | 52 | if (!group_caches[thread].InitOnce(memory_manager)) { |
1153 | 0 | has_error = true; |
1154 | 0 | return; |
1155 | 0 | } |
1156 | 52 | TokenizeCoefficients( |
1157 | 52 | &shared.coeff_orders[idx_pass * shared.coeff_order_size], rect, |
1158 | 52 | ac_rows, shared.ac_strategy, frame_header.chroma_subsampling, |
1159 | 52 | &group_caches[thread].num_nzeroes, |
1160 | 52 | &enc_state->passes[idx_pass].ac_tokens[group_index], shared.quant_dc, |
1161 | 52 | shared.raw_quant_field, shared.block_ctx_map); |
1162 | 52 | } |
1163 | 52 | }; |
1164 | 52 | JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, shared.frame_dim.num_groups, |
1165 | 52 | tokenize_group_init, tokenize_group, |
1166 | 52 | "TokenizeGroup")); |
1167 | 52 | if (has_error) return JXL_FAILURE("TokenizeGroup failed"); |
1168 | 52 | return true; |
1169 | 52 | } |
1170 | | |
1171 | | Status EncodeGlobalDCInfo(const PassesSharedState& shared, BitWriter* writer, |
1172 | 52 | AuxOut* aux_out) { |
1173 | | // Encode quantizer DC and global scale. |
1174 | 52 | QuantizerParams params = shared.quantizer.GetParams(); |
1175 | 52 | JXL_RETURN_IF_ERROR( |
1176 | 52 | WriteQuantizerParams(params, writer, kLayerQuant, aux_out)); |
1177 | 52 | EncodeBlockCtxMap(shared.block_ctx_map, writer, aux_out); |
1178 | 52 | ColorCorrelationEncodeDC(shared.cmap.base(), writer, kLayerDC, aux_out); |
1179 | 52 | return true; |
1180 | 52 | } |
1181 | | |
1182 | | // In streaming mode, this function only performs the histogram clustering and |
1183 | | // saves the histogram bitstreams in enc_state, the actual AC global bitstream |
1184 | | // is written in OutputAcGlobal() function after all the groups are processed. |
1185 | | Status EncodeGlobalACInfo(PassesEncoderState* enc_state, BitWriter* writer, |
1186 | 52 | ModularFrameEncoder* enc_modular, AuxOut* aux_out) { |
1187 | 52 | PassesSharedState& shared = enc_state->shared; |
1188 | 52 | JxlMemoryManager* memory_manager = enc_state->memory_manager(); |
1189 | 52 | JXL_RETURN_IF_ERROR(DequantMatricesEncode(memory_manager, shared.matrices, |
1190 | 52 | writer, kLayerQuant, aux_out, |
1191 | 52 | enc_modular)); |
1192 | 52 | size_t num_histo_bits = CeilLog2Nonzero(shared.frame_dim.num_groups); |
1193 | 52 | if (!enc_state->streaming_mode && num_histo_bits != 0) { |
1194 | 0 | BitWriter::Allotment allotment(writer, num_histo_bits); |
1195 | 0 | writer->Write(num_histo_bits, shared.num_histograms - 1); |
1196 | 0 | allotment.ReclaimAndCharge(writer, kLayerAC, aux_out); |
1197 | 0 | } |
1198 | | |
1199 | 104 | for (size_t i = 0; i < enc_state->progressive_splitter.GetNumPasses(); i++) { |
1200 | | // Encode coefficient orders. |
1201 | 52 | if (!enc_state->streaming_mode) { |
1202 | 52 | size_t order_bits = 0; |
1203 | 52 | JXL_RETURN_IF_ERROR(U32Coder::CanEncode( |
1204 | 52 | kOrderEnc, enc_state->used_orders[i], &order_bits)); |
1205 | 52 | BitWriter::Allotment allotment(writer, order_bits); |
1206 | 52 | JXL_CHECK(U32Coder::Write(kOrderEnc, enc_state->used_orders[i], writer)); |
1207 | 52 | allotment.ReclaimAndCharge(writer, kLayerOrder, aux_out); |
1208 | 52 | EncodeCoeffOrders(enc_state->used_orders[i], |
1209 | 52 | &shared.coeff_orders[i * shared.coeff_order_size], |
1210 | 52 | writer, kLayerOrder, aux_out); |
1211 | 52 | } |
1212 | | |
1213 | | // Encode histograms. |
1214 | 52 | HistogramParams hist_params(enc_state->cparams.speed_tier, |
1215 | 52 | shared.block_ctx_map.NumACContexts()); |
1216 | 52 | if (enc_state->cparams.speed_tier > SpeedTier::kTortoise) { |
1217 | 52 | hist_params.lz77_method = HistogramParams::LZ77Method::kNone; |
1218 | 52 | } |
1219 | 52 | if (enc_state->cparams.decoding_speed_tier >= 1) { |
1220 | 0 | hist_params.max_histograms = 6; |
1221 | 0 | } |
1222 | 52 | size_t num_histogram_groups = shared.num_histograms; |
1223 | 52 | if (enc_state->streaming_mode) { |
1224 | 0 | size_t prev_num_histograms = |
1225 | 0 | enc_state->passes[i].codes.encoding_info.size(); |
1226 | 0 | if (enc_state->initialize_global_state) { |
1227 | 0 | prev_num_histograms += kNumFixedHistograms; |
1228 | 0 | hist_params.add_fixed_histograms = true; |
1229 | 0 | } |
1230 | 0 | size_t remaining_histograms = kClustersLimit - prev_num_histograms; |
1231 | | // Heuristic to assign budget of new histograms to DC groups. |
1232 | | // TODO(szabadka) Tune this together with the DC group ordering. |
1233 | 0 | size_t max_histograms = remaining_histograms < 20 |
1234 | 0 | ? std::min<size_t>(remaining_histograms, 4) |
1235 | 0 | : remaining_histograms / 4; |
1236 | 0 | hist_params.max_histograms = |
1237 | 0 | std::min(max_histograms, hist_params.max_histograms); |
1238 | 0 | num_histogram_groups = 1; |
1239 | 0 | } |
1240 | 52 | hist_params.streaming_mode = enc_state->streaming_mode; |
1241 | 52 | hist_params.initialize_global_state = enc_state->initialize_global_state; |
1242 | 52 | BuildAndEncodeHistograms( |
1243 | 52 | memory_manager, hist_params, |
1244 | 52 | num_histogram_groups * shared.block_ctx_map.NumACContexts(), |
1245 | 52 | enc_state->passes[i].ac_tokens, &enc_state->passes[i].codes, |
1246 | 52 | &enc_state->passes[i].context_map, writer, kLayerAC, aux_out); |
1247 | 52 | } |
1248 | | |
1249 | 52 | return true; |
1250 | 52 | } |
1251 | | |
1252 | | Status EncodeGroups(const FrameHeader& frame_header, |
1253 | | PassesEncoderState* enc_state, |
1254 | | ModularFrameEncoder* enc_modular, ThreadPool* pool, |
1255 | | std::vector<std::unique_ptr<BitWriter>>* group_codes, |
1256 | 52 | AuxOut* aux_out) { |
1257 | 52 | const PassesSharedState& shared = enc_state->shared; |
1258 | 52 | JxlMemoryManager* memory_manager = shared.memory_manager; |
1259 | 52 | const FrameDimensions& frame_dim = shared.frame_dim; |
1260 | 52 | const size_t num_groups = frame_dim.num_groups; |
1261 | 52 | const size_t num_passes = enc_state->progressive_splitter.GetNumPasses(); |
1262 | 52 | const size_t global_ac_index = frame_dim.num_dc_groups + 1; |
1263 | 52 | const bool is_small_image = |
1264 | 52 | !enc_state->streaming_mode && num_groups == 1 && num_passes == 1; |
1265 | 52 | const size_t num_toc_entries = |
1266 | 52 | is_small_image ? 1 |
1267 | 52 | : AcGroupIndex(0, 0, num_groups, frame_dim.num_dc_groups) + |
1268 | 0 | num_groups * num_passes; |
1269 | 52 | JXL_ASSERT(group_codes->empty()); |
1270 | 52 | group_codes->reserve(num_toc_entries); |
1271 | 104 | for (size_t i = 0; i < num_toc_entries; ++i) { |
1272 | 52 | group_codes->emplace_back(jxl::make_unique<BitWriter>(memory_manager)); |
1273 | 52 | } |
1274 | | |
1275 | 416 | const auto get_output = [&](const size_t index) -> BitWriter* { |
1276 | 416 | return (*group_codes)[is_small_image ? 0 : index].get(); |
1277 | 416 | }; |
1278 | 104 | auto ac_group_code = [&](size_t pass, size_t group) { |
1279 | 104 | return get_output(AcGroupIndex(pass, group, frame_dim.num_groups, |
1280 | 104 | frame_dim.num_dc_groups)); |
1281 | 104 | }; |
1282 | | |
1283 | 52 | if (enc_state->initialize_global_state) { |
1284 | 52 | if (frame_header.flags & FrameHeader::kPatches) { |
1285 | 0 | PatchDictionaryEncoder::Encode(shared.image_features.patches, |
1286 | 0 | get_output(0), kLayerDictionary, aux_out); |
1287 | 0 | } |
1288 | 52 | if (frame_header.flags & FrameHeader::kSplines) { |
1289 | 0 | EncodeSplines(shared.image_features.splines, get_output(0), kLayerSplines, |
1290 | 0 | HistogramParams(), aux_out); |
1291 | 0 | } |
1292 | 52 | if (frame_header.flags & FrameHeader::kNoise) { |
1293 | 0 | EncodeNoise(shared.image_features.noise_params, get_output(0), |
1294 | 0 | kLayerNoise, aux_out); |
1295 | 0 | } |
1296 | | |
1297 | 52 | JXL_RETURN_IF_ERROR(DequantMatricesEncodeDC(shared.matrices, get_output(0), |
1298 | 52 | kLayerQuant, aux_out)); |
1299 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1300 | 52 | JXL_RETURN_IF_ERROR(EncodeGlobalDCInfo(shared, get_output(0), aux_out)); |
1301 | 52 | } |
1302 | 52 | JXL_RETURN_IF_ERROR(enc_modular->EncodeGlobalInfo(enc_state->streaming_mode, |
1303 | 52 | get_output(0), aux_out)); |
1304 | 52 | JXL_RETURN_IF_ERROR(enc_modular->EncodeStream(get_output(0), aux_out, |
1305 | 52 | kLayerModularGlobal, |
1306 | 52 | ModularStreamId::Global())); |
1307 | 52 | } |
1308 | | |
1309 | 52 | std::vector<std::unique_ptr<AuxOut>> aux_outs; |
1310 | 52 | auto resize_aux_outs = [&aux_outs, |
1311 | 156 | aux_out](const size_t num_threads) -> Status { |
1312 | 156 | if (aux_out == nullptr) { |
1313 | 156 | aux_outs.resize(num_threads); |
1314 | 156 | } else { |
1315 | 0 | while (aux_outs.size() > num_threads) { |
1316 | 0 | aux_out->Assimilate(*aux_outs.back()); |
1317 | 0 | aux_outs.pop_back(); |
1318 | 0 | } |
1319 | 0 | while (num_threads > aux_outs.size()) { |
1320 | 0 | aux_outs.emplace_back(jxl::make_unique<AuxOut>()); |
1321 | 0 | } |
1322 | 0 | } |
1323 | 156 | return true; |
1324 | 156 | }; |
1325 | | |
1326 | 52 | const auto process_dc_group = [&](const uint32_t group_index, |
1327 | 52 | const size_t thread) { |
1328 | 52 | AuxOut* my_aux_out = aux_outs[thread].get(); |
1329 | 52 | BitWriter* output = get_output(group_index + 1); |
1330 | 52 | int modular_group_index = group_index; |
1331 | 52 | if (enc_state->streaming_mode) { |
1332 | 0 | JXL_ASSERT(group_index == 0); |
1333 | 0 | modular_group_index = enc_state->dc_group_index; |
1334 | 0 | } |
1335 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT && |
1336 | 52 | !(frame_header.flags & FrameHeader::kUseDcFrame)) { |
1337 | 52 | BitWriter::Allotment allotment(output, 2); |
1338 | 52 | output->Write(2, enc_modular->extra_dc_precision[modular_group_index]); |
1339 | 52 | allotment.ReclaimAndCharge(output, kLayerDC, my_aux_out); |
1340 | 52 | JXL_CHECK(enc_modular->EncodeStream( |
1341 | 52 | output, my_aux_out, kLayerDC, |
1342 | 52 | ModularStreamId::VarDCTDC(modular_group_index))); |
1343 | 52 | } |
1344 | 52 | JXL_CHECK(enc_modular->EncodeStream( |
1345 | 52 | output, my_aux_out, kLayerModularDcGroup, |
1346 | 52 | ModularStreamId::ModularDC(modular_group_index))); |
1347 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1348 | 52 | const Rect& rect = enc_state->shared.frame_dim.DCGroupRect(group_index); |
1349 | 52 | size_t nb_bits = CeilLog2Nonzero(rect.xsize() * rect.ysize()); |
1350 | 52 | if (nb_bits != 0) { |
1351 | 0 | BitWriter::Allotment allotment(output, nb_bits); |
1352 | 0 | output->Write(nb_bits, |
1353 | 0 | enc_modular->ac_metadata_size[modular_group_index] - 1); |
1354 | 0 | allotment.ReclaimAndCharge(output, kLayerControlFields, my_aux_out); |
1355 | 0 | } |
1356 | 52 | JXL_CHECK(enc_modular->EncodeStream( |
1357 | 52 | output, my_aux_out, kLayerControlFields, |
1358 | 52 | ModularStreamId::ACMetadata(modular_group_index))); |
1359 | 52 | } |
1360 | 52 | }; |
1361 | 52 | JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, frame_dim.num_dc_groups, |
1362 | 52 | resize_aux_outs, process_dc_group, |
1363 | 52 | "EncodeDCGroup")); |
1364 | | |
1365 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1366 | 52 | JXL_RETURN_IF_ERROR(EncodeGlobalACInfo( |
1367 | 52 | enc_state, get_output(global_ac_index), enc_modular, aux_out)); |
1368 | 52 | } |
1369 | | |
1370 | 52 | std::atomic<bool> has_error{false}; |
1371 | 52 | const auto process_group = [&](const uint32_t group_index, |
1372 | 52 | const size_t thread) { |
1373 | 52 | if (has_error) return; |
1374 | 52 | AuxOut* my_aux_out = aux_outs[thread].get(); |
1375 | | |
1376 | 52 | size_t ac_group_id = |
1377 | 52 | enc_state->streaming_mode |
1378 | 52 | ? enc_modular->ComputeStreamingAbsoluteAcGroupId( |
1379 | 0 | enc_state->dc_group_index, group_index, shared.frame_dim) |
1380 | 52 | : group_index; |
1381 | | |
1382 | 104 | for (size_t i = 0; i < num_passes; i++) { |
1383 | 52 | JXL_DEBUG_V(2, "Encoding AC group %u [abs %" PRIuS "] pass %" PRIuS, |
1384 | 52 | group_index, ac_group_id, i); |
1385 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1386 | 52 | if (!EncodeGroupTokenizedCoefficients( |
1387 | 52 | group_index, i, enc_state->histogram_idx[group_index], |
1388 | 52 | *enc_state, ac_group_code(i, group_index), my_aux_out)) { |
1389 | 0 | has_error = true; |
1390 | 0 | return; |
1391 | 0 | } |
1392 | 52 | } |
1393 | | // Write all modular encoded data (color?, alpha, depth, extra channels) |
1394 | 52 | if (!enc_modular->EncodeStream( |
1395 | 52 | ac_group_code(i, group_index), my_aux_out, kLayerModularAcGroup, |
1396 | 52 | ModularStreamId::ModularAC(ac_group_id, i))) { |
1397 | 0 | has_error = true; |
1398 | 0 | return; |
1399 | 0 | } |
1400 | 52 | JXL_DEBUG_V(2, |
1401 | 52 | "AC group %u [abs %" PRIuS "] pass %" PRIuS |
1402 | 52 | " encoded size is %" PRIuS " bits", |
1403 | 52 | group_index, ac_group_id, i, |
1404 | 52 | ac_group_code(i, group_index)->BitsWritten()); |
1405 | 52 | } |
1406 | 52 | }; |
1407 | 52 | JXL_RETURN_IF_ERROR(RunOnPool(pool, 0, num_groups, resize_aux_outs, |
1408 | 52 | process_group, "EncodeGroupCoefficients")); |
1409 | 52 | if (has_error) return JXL_FAILURE("EncodeGroupCoefficients failed"); |
1410 | | // Resizing aux_outs to 0 also Assimilates the array. |
1411 | 52 | static_cast<void>(resize_aux_outs(0)); |
1412 | | |
1413 | 52 | for (std::unique_ptr<BitWriter>& bw : *group_codes) { |
1414 | 52 | BitWriter::Allotment allotment(bw.get(), 8); |
1415 | 52 | bw->ZeroPadToByte(); // end of group. |
1416 | 52 | allotment.ReclaimAndCharge(bw.get(), kLayerAC, aux_out); |
1417 | 52 | } |
1418 | 52 | return true; |
1419 | 52 | } |
1420 | | |
1421 | | Status ComputeEncodingData( |
1422 | | const CompressParams& cparams, const FrameInfo& frame_info, |
1423 | | const CodecMetadata* metadata, JxlEncoderChunkedFrameAdapter& frame_data, |
1424 | | const jpeg::JPEGData* jpeg_data, size_t x0, size_t y0, size_t xsize, |
1425 | | size_t ysize, const JxlCmsInterface& cms, ThreadPool* pool, |
1426 | | FrameHeader& mutable_frame_header, ModularFrameEncoder& enc_modular, |
1427 | | PassesEncoderState& enc_state, |
1428 | 52 | std::vector<std::unique_ptr<BitWriter>>* group_codes, AuxOut* aux_out) { |
1429 | 52 | JXL_ASSERT(x0 + xsize <= frame_data.xsize); |
1430 | 52 | JXL_ASSERT(y0 + ysize <= frame_data.ysize); |
1431 | 52 | JxlMemoryManager* memory_manager = enc_state.memory_manager(); |
1432 | 52 | const FrameHeader& frame_header = mutable_frame_header; |
1433 | 52 | PassesSharedState& shared = enc_state.shared; |
1434 | 52 | shared.metadata = metadata; |
1435 | 52 | if (enc_state.streaming_mode) { |
1436 | 0 | shared.frame_dim.Set( |
1437 | 0 | xsize, ysize, frame_header.group_size_shift, |
1438 | 0 | /*max_hshift=*/0, /*max_vshift=*/0, |
1439 | 0 | mutable_frame_header.encoding == FrameEncoding::kModular, |
1440 | 0 | /*upsampling=*/1); |
1441 | 52 | } else { |
1442 | 52 | shared.frame_dim = frame_header.ToFrameDimensions(); |
1443 | 52 | } |
1444 | | |
1445 | 52 | shared.image_features.patches.SetPassesSharedState(&shared); |
1446 | 52 | const FrameDimensions& frame_dim = shared.frame_dim; |
1447 | 52 | JXL_ASSIGN_OR_RETURN( |
1448 | 52 | shared.ac_strategy, |
1449 | 52 | AcStrategyImage::Create(memory_manager, frame_dim.xsize_blocks, |
1450 | 52 | frame_dim.ysize_blocks)); |
1451 | 52 | JXL_ASSIGN_OR_RETURN(shared.raw_quant_field, |
1452 | 52 | ImageI::Create(memory_manager, frame_dim.xsize_blocks, |
1453 | 52 | frame_dim.ysize_blocks)); |
1454 | 52 | JXL_ASSIGN_OR_RETURN(shared.epf_sharpness, |
1455 | 52 | ImageB::Create(memory_manager, frame_dim.xsize_blocks, |
1456 | 52 | frame_dim.ysize_blocks)); |
1457 | 52 | JXL_ASSIGN_OR_RETURN( |
1458 | 52 | shared.cmap, ColorCorrelationMap::Create(memory_manager, frame_dim.xsize, |
1459 | 52 | frame_dim.ysize)); |
1460 | 52 | shared.coeff_order_size = kCoeffOrderMaxSize; |
1461 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1462 | 52 | shared.coeff_orders.resize(frame_header.passes.num_passes * |
1463 | 52 | kCoeffOrderMaxSize); |
1464 | 52 | } |
1465 | | |
1466 | 52 | JXL_ASSIGN_OR_RETURN(shared.quant_dc, |
1467 | 52 | ImageB::Create(memory_manager, frame_dim.xsize_blocks, |
1468 | 52 | frame_dim.ysize_blocks)); |
1469 | 52 | JXL_ASSIGN_OR_RETURN(shared.dc_storage, |
1470 | 52 | Image3F::Create(memory_manager, frame_dim.xsize_blocks, |
1471 | 52 | frame_dim.ysize_blocks)); |
1472 | 52 | shared.dc = &shared.dc_storage; |
1473 | | |
1474 | 52 | const size_t num_extra_channels = metadata->m.num_extra_channels; |
1475 | 52 | const ExtraChannelInfo* alpha_eci = metadata->m.Find(ExtraChannel::kAlpha); |
1476 | 52 | const ExtraChannelInfo* black_eci = metadata->m.Find(ExtraChannel::kBlack); |
1477 | 52 | const size_t alpha_idx = alpha_eci - metadata->m.extra_channel_info.data(); |
1478 | 52 | const size_t black_idx = black_eci - metadata->m.extra_channel_info.data(); |
1479 | 52 | const ColorEncoding c_enc = metadata->m.color_encoding; |
1480 | | |
1481 | | // Make the image patch bigger than the currently processed group in streaming |
1482 | | // mode so that we can take into account border pixels around the group when |
1483 | | // computing inverse Gaborish and adaptive quantization map. |
1484 | 52 | int max_border = enc_state.streaming_mode ? kBlockDim : 0; |
1485 | 52 | Rect frame_rect(0, 0, frame_data.xsize, frame_data.ysize); |
1486 | 52 | Rect frame_area_rect = Rect(x0, y0, xsize, ysize); |
1487 | 52 | Rect patch_rect = frame_area_rect.Extend(max_border, frame_rect); |
1488 | 52 | JXL_ASSERT(patch_rect.IsInside(frame_rect)); |
1489 | | |
1490 | | // Allocating a large enough image avoids a copy when padding. |
1491 | 104 | JXL_ASSIGN_OR_RETURN( |
1492 | 104 | Image3F color, |
1493 | 104 | Image3F::Create(memory_manager, RoundUpToBlockDim(patch_rect.xsize()), |
1494 | 104 | RoundUpToBlockDim(patch_rect.ysize()))); |
1495 | 104 | color.ShrinkTo(patch_rect.xsize(), patch_rect.ysize()); |
1496 | 104 | std::vector<ImageF> extra_channels(num_extra_channels); |
1497 | 104 | for (auto& extra_channel : extra_channels) { |
1498 | 0 | JXL_ASSIGN_OR_RETURN( |
1499 | 0 | extra_channel, |
1500 | 0 | ImageF::Create(memory_manager, patch_rect.xsize(), patch_rect.ysize())); |
1501 | 0 | } |
1502 | 52 | ImageF* alpha = alpha_eci ? &extra_channels[alpha_idx] : nullptr; |
1503 | 52 | ImageF* black = black_eci ? &extra_channels[black_idx] : nullptr; |
1504 | 52 | bool has_interleaved_alpha = false; |
1505 | 52 | JxlChunkedFrameInputSource input = frame_data.GetInputSource(); |
1506 | 52 | if (!frame_data.IsJPEG()) { |
1507 | 52 | JXL_RETURN_IF_ERROR(CopyColorChannels(input, patch_rect, frame_info, |
1508 | 52 | metadata->m, pool, &color, alpha, |
1509 | 52 | &has_interleaved_alpha)); |
1510 | 52 | } |
1511 | 52 | JXL_RETURN_IF_ERROR(CopyExtraChannels(input, patch_rect, frame_info, |
1512 | 52 | metadata->m, has_interleaved_alpha, |
1513 | 52 | pool, &extra_channels)); |
1514 | | |
1515 | 52 | shared.image_features.patches.SetPassesSharedState(&shared); |
1516 | 52 | enc_state.cparams = cparams; |
1517 | | |
1518 | 52 | Image3F linear_storage; |
1519 | 52 | Image3F* linear = nullptr; |
1520 | | |
1521 | 52 | if (!jpeg_data) { |
1522 | 52 | if (frame_header.color_transform == ColorTransform::kXYB && |
1523 | 52 | frame_info.ib_needs_color_transform) { |
1524 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT && |
1525 | 52 | cparams.speed_tier <= SpeedTier::kKitten) { |
1526 | 0 | JXL_ASSIGN_OR_RETURN(linear_storage, |
1527 | 0 | Image3F::Create(memory_manager, patch_rect.xsize(), |
1528 | 0 | patch_rect.ysize())); |
1529 | 0 | linear = &linear_storage; |
1530 | 0 | } |
1531 | 52 | ToXYB(c_enc, metadata->m.IntensityTarget(), black, pool, &color, cms, |
1532 | 52 | linear); |
1533 | 52 | } else { |
1534 | | // Nothing to do. |
1535 | | // RGB or YCbCr: forward YCbCr is not implemented, this is only used when |
1536 | | // the input is already in YCbCr |
1537 | | // If encoding a special DC or reference frame: input is already in XYB. |
1538 | 0 | } |
1539 | 52 | bool lossless = cparams.IsLossless(); |
1540 | 52 | if (alpha && !alpha_eci->alpha_associated && |
1541 | 52 | frame_header.frame_type == FrameType::kRegularFrame && |
1542 | 52 | !ApplyOverride(cparams.keep_invisible, cparams.IsLossless()) && |
1543 | 52 | cparams.ec_resampling == cparams.resampling && |
1544 | 52 | !cparams.disable_percepeptual_optimizations) { |
1545 | | // simplify invisible pixels |
1546 | 0 | SimplifyInvisible(&color, *alpha, lossless); |
1547 | 0 | if (linear) { |
1548 | 0 | SimplifyInvisible(linear, *alpha, lossless); |
1549 | 0 | } |
1550 | 0 | } |
1551 | 52 | PadImageToBlockMultipleInPlace(&color); |
1552 | 52 | } |
1553 | | |
1554 | | // Rectangle within color that corresponds to the currently processed group in |
1555 | | // streaming mode. |
1556 | 52 | Rect group_rect(x0 - patch_rect.x0(), y0 - patch_rect.y0(), |
1557 | 52 | RoundUpToBlockDim(xsize), RoundUpToBlockDim(ysize)); |
1558 | | |
1559 | 52 | if (enc_state.initialize_global_state && !jpeg_data) { |
1560 | 52 | ComputeChromacityAdjustments(cparams, color, group_rect, |
1561 | 52 | &mutable_frame_header); |
1562 | 52 | } |
1563 | | |
1564 | 52 | bool has_jpeg_data = (jpeg_data != nullptr); |
1565 | 52 | ComputeNoiseParams(cparams, enc_state.streaming_mode, has_jpeg_data, color, |
1566 | 52 | frame_dim, &mutable_frame_header, |
1567 | 52 | &shared.image_features.noise_params); |
1568 | | |
1569 | 52 | JXL_RETURN_IF_ERROR( |
1570 | 52 | DownsampleColorChannels(cparams, frame_header, has_jpeg_data, &color)); |
1571 | | |
1572 | 52 | if (cparams.ec_resampling != 1 && !cparams.already_downsampled) { |
1573 | 0 | for (ImageF& ec : extra_channels) { |
1574 | 0 | JXL_ASSIGN_OR_RETURN(ec, DownsampleImage(ec, cparams.ec_resampling)); |
1575 | 0 | } |
1576 | 0 | } |
1577 | | |
1578 | 52 | if (!enc_state.streaming_mode) { |
1579 | 52 | group_rect = Rect(color); |
1580 | 52 | } |
1581 | | |
1582 | 52 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
1583 | 52 | enc_state.passes.resize(enc_state.progressive_splitter.GetNumPasses()); |
1584 | 52 | for (PassesEncoderState::PassData& pass : enc_state.passes) { |
1585 | 52 | pass.ac_tokens.resize(shared.frame_dim.num_groups); |
1586 | 52 | } |
1587 | 52 | if (jpeg_data) { |
1588 | 0 | JXL_RETURN_IF_ERROR(ComputeJPEGTranscodingData( |
1589 | 0 | *jpeg_data, frame_header, pool, &enc_modular, &enc_state)); |
1590 | 52 | } else { |
1591 | 52 | JXL_RETURN_IF_ERROR(ComputeVarDCTEncodingData( |
1592 | 52 | frame_header, linear, &color, group_rect, cms, pool, &enc_modular, |
1593 | 52 | &enc_state, aux_out)); |
1594 | 52 | } |
1595 | 52 | ComputeAllCoeffOrders(enc_state, frame_dim); |
1596 | 52 | if (!enc_state.streaming_mode) { |
1597 | 52 | shared.num_histograms = 1; |
1598 | 52 | enc_state.histogram_idx.resize(frame_dim.num_groups); |
1599 | 52 | } |
1600 | 52 | JXL_RETURN_IF_ERROR( |
1601 | 52 | TokenizeAllCoefficients(frame_header, pool, &enc_state)); |
1602 | 52 | } |
1603 | | |
1604 | 52 | if (cparams.modular_mode || !extra_channels.empty()) { |
1605 | 0 | JXL_RETURN_IF_ERROR(enc_modular.ComputeEncodingData( |
1606 | 0 | frame_header, metadata->m, &color, extra_channels, group_rect, |
1607 | 0 | frame_dim, frame_area_rect, &enc_state, cms, pool, aux_out, |
1608 | 0 | /*do_color=*/cparams.modular_mode)); |
1609 | 0 | } |
1610 | | |
1611 | 52 | if (!enc_state.streaming_mode) { |
1612 | 52 | if (cparams.speed_tier < SpeedTier::kTortoise || |
1613 | 52 | !cparams.ModularPartIsLossless() || cparams.responsive || |
1614 | 52 | !cparams.custom_fixed_tree.empty()) { |
1615 | | // Use local trees if doing lossless modular, unless at very slow speeds. |
1616 | 52 | JXL_RETURN_IF_ERROR(enc_modular.ComputeTree(pool)); |
1617 | 52 | JXL_RETURN_IF_ERROR(enc_modular.ComputeTokens(pool)); |
1618 | 52 | } |
1619 | 52 | mutable_frame_header.UpdateFlag(shared.image_features.patches.HasAny(), |
1620 | 52 | FrameHeader::kPatches); |
1621 | 52 | mutable_frame_header.UpdateFlag(shared.image_features.splines.HasAny(), |
1622 | 52 | FrameHeader::kSplines); |
1623 | 52 | } |
1624 | | |
1625 | 52 | JXL_RETURN_IF_ERROR(EncodeGroups(frame_header, &enc_state, &enc_modular, pool, |
1626 | 52 | group_codes, aux_out)); |
1627 | 52 | if (enc_state.streaming_mode) { |
1628 | 0 | const size_t group_index = enc_state.dc_group_index; |
1629 | 0 | enc_modular.ClearStreamData(ModularStreamId::VarDCTDC(group_index)); |
1630 | 0 | enc_modular.ClearStreamData(ModularStreamId::ACMetadata(group_index)); |
1631 | 0 | enc_modular.ClearModularStreamData(); |
1632 | 0 | } |
1633 | 52 | return true; |
1634 | 52 | } |
1635 | | |
1636 | | Status PermuteGroups(const CompressParams& cparams, |
1637 | | const FrameDimensions& frame_dim, size_t num_passes, |
1638 | | std::vector<coeff_order_t>* permutation, |
1639 | 52 | std::vector<std::unique_ptr<BitWriter>>* group_codes) { |
1640 | 52 | const size_t num_groups = frame_dim.num_groups; |
1641 | 52 | if (!cparams.centerfirst || (num_passes == 1 && num_groups == 1)) { |
1642 | 52 | return true; |
1643 | 52 | } |
1644 | | // Don't permute global DC/AC or DC. |
1645 | 0 | permutation->resize(frame_dim.num_dc_groups + 2); |
1646 | 0 | std::iota(permutation->begin(), permutation->end(), 0); |
1647 | 0 | std::vector<coeff_order_t> ac_group_order(num_groups); |
1648 | 0 | std::iota(ac_group_order.begin(), ac_group_order.end(), 0); |
1649 | 0 | size_t group_dim = frame_dim.group_dim; |
1650 | | |
1651 | | // The center of the image is either given by parameters or chosen |
1652 | | // to be the middle of the image by default if center_x, center_y resp. |
1653 | | // are not provided. |
1654 | |
|
1655 | 0 | int64_t imag_cx; |
1656 | 0 | if (cparams.center_x != static_cast<size_t>(-1)) { |
1657 | 0 | JXL_RETURN_IF_ERROR(cparams.center_x < frame_dim.xsize); |
1658 | 0 | imag_cx = cparams.center_x; |
1659 | 0 | } else { |
1660 | 0 | imag_cx = frame_dim.xsize / 2; |
1661 | 0 | } |
1662 | | |
1663 | 0 | int64_t imag_cy; |
1664 | 0 | if (cparams.center_y != static_cast<size_t>(-1)) { |
1665 | 0 | JXL_RETURN_IF_ERROR(cparams.center_y < frame_dim.ysize); |
1666 | 0 | imag_cy = cparams.center_y; |
1667 | 0 | } else { |
1668 | 0 | imag_cy = frame_dim.ysize / 2; |
1669 | 0 | } |
1670 | | |
1671 | | // The center of the group containing the center of the image. |
1672 | 0 | int64_t cx = (imag_cx / group_dim) * group_dim + group_dim / 2; |
1673 | 0 | int64_t cy = (imag_cy / group_dim) * group_dim + group_dim / 2; |
1674 | | // This identifies in what area of the central group the center of the image |
1675 | | // lies in. |
1676 | 0 | double direction = -std::atan2(imag_cy - cy, imag_cx - cx); |
1677 | | // This identifies the side of the central group the center of the image |
1678 | | // lies closest to. This can take values 0, 1, 2, 3 corresponding to left, |
1679 | | // bottom, right, top. |
1680 | 0 | int64_t side = std::fmod((direction + 5 * kPi / 4), 2 * kPi) * 2 / kPi; |
1681 | 0 | auto get_distance_from_center = [&](size_t gid) { |
1682 | 0 | Rect r = frame_dim.GroupRect(gid); |
1683 | 0 | int64_t gcx = r.x0() + group_dim / 2; |
1684 | 0 | int64_t gcy = r.y0() + group_dim / 2; |
1685 | 0 | int64_t dx = gcx - cx; |
1686 | 0 | int64_t dy = gcy - cy; |
1687 | | // The angle is determined by taking atan2 and adding an appropriate |
1688 | | // starting point depending on the side we want to start on. |
1689 | 0 | double angle = std::remainder( |
1690 | 0 | std::atan2(dy, dx) + kPi / 4 + side * (kPi / 2), 2 * kPi); |
1691 | | // Concentric squares in clockwise order. |
1692 | 0 | return std::make_pair(std::max(std::abs(dx), std::abs(dy)), angle); |
1693 | 0 | }; |
1694 | 0 | std::sort(ac_group_order.begin(), ac_group_order.end(), |
1695 | 0 | [&](coeff_order_t a, coeff_order_t b) { |
1696 | 0 | return get_distance_from_center(a) < get_distance_from_center(b); |
1697 | 0 | }); |
1698 | 0 | std::vector<coeff_order_t> inv_ac_group_order(ac_group_order.size(), 0); |
1699 | 0 | for (size_t i = 0; i < ac_group_order.size(); i++) { |
1700 | 0 | inv_ac_group_order[ac_group_order[i]] = i; |
1701 | 0 | } |
1702 | 0 | for (size_t i = 0; i < num_passes; i++) { |
1703 | 0 | size_t pass_start = permutation->size(); |
1704 | 0 | for (coeff_order_t v : inv_ac_group_order) { |
1705 | 0 | permutation->push_back(pass_start + v); |
1706 | 0 | } |
1707 | 0 | } |
1708 | 0 | std::vector<std::unique_ptr<BitWriter>> new_group_codes(group_codes->size()); |
1709 | 0 | for (size_t i = 0; i < permutation->size(); i++) { |
1710 | 0 | new_group_codes[(*permutation)[i]] = std::move((*group_codes)[i]); |
1711 | 0 | } |
1712 | 0 | group_codes->swap(new_group_codes); |
1713 | 0 | return true; |
1714 | 0 | } |
1715 | | |
1716 | | bool CanDoStreamingEncoding(const CompressParams& cparams, |
1717 | | const FrameInfo& frame_info, |
1718 | | const CodecMetadata& metadata, |
1719 | 52 | const JxlEncoderChunkedFrameAdapter& frame_data) { |
1720 | 52 | if (cparams.buffering == 0) { |
1721 | 0 | return false; |
1722 | 0 | } |
1723 | 52 | if (cparams.buffering == -1) { |
1724 | 52 | if (cparams.speed_tier < SpeedTier::kTortoise) return false; |
1725 | 52 | if (cparams.speed_tier < SpeedTier::kSquirrel && |
1726 | 52 | cparams.butteraugli_distance > 0.5f) { |
1727 | 0 | return false; |
1728 | 0 | } |
1729 | 52 | if (cparams.speed_tier == SpeedTier::kSquirrel && |
1730 | 52 | cparams.butteraugli_distance >= 3.f) { |
1731 | 0 | return false; |
1732 | 0 | } |
1733 | 52 | } |
1734 | | |
1735 | | // TODO(veluca): handle different values of `buffering`. |
1736 | 52 | if (frame_data.xsize <= 2048 && frame_data.ysize <= 2048) { |
1737 | 52 | return false; |
1738 | 52 | } |
1739 | 0 | if (frame_data.IsJPEG()) { |
1740 | 0 | return false; |
1741 | 0 | } |
1742 | 0 | if (cparams.noise == Override::kOn || cparams.patches == Override::kOn) { |
1743 | 0 | return false; |
1744 | 0 | } |
1745 | 0 | if (cparams.progressive_dc != 0 || frame_info.dc_level != 0) { |
1746 | 0 | return false; |
1747 | 0 | } |
1748 | 0 | if (cparams.resampling != 1 || cparams.ec_resampling != 1) { |
1749 | 0 | return false; |
1750 | 0 | } |
1751 | 0 | if (cparams.max_error_mode) { |
1752 | 0 | return false; |
1753 | 0 | } |
1754 | 0 | if (!cparams.ModularPartIsLossless() || cparams.responsive > 0) { |
1755 | 0 | if (metadata.m.num_extra_channels > 0 || cparams.modular_mode) { |
1756 | 0 | return false; |
1757 | 0 | } |
1758 | 0 | } |
1759 | 0 | ColorTransform ok_color_transform = |
1760 | 0 | cparams.modular_mode ? ColorTransform::kNone : ColorTransform::kXYB; |
1761 | 0 | if (cparams.color_transform != ok_color_transform) { |
1762 | 0 | return false; |
1763 | 0 | } |
1764 | 0 | return true; |
1765 | 0 | } |
1766 | | |
1767 | | void ComputePermutationForStreaming(size_t xsize, size_t ysize, |
1768 | | size_t group_size, size_t num_passes, |
1769 | | std::vector<coeff_order_t>& permutation, |
1770 | 0 | std::vector<size_t>& dc_group_order) { |
1771 | | // This is only valid in VarDCT mode, otherwise there can be group shift. |
1772 | 0 | const size_t dc_group_size = group_size * kBlockDim; |
1773 | 0 | const size_t group_xsize = DivCeil(xsize, group_size); |
1774 | 0 | const size_t group_ysize = DivCeil(ysize, group_size); |
1775 | 0 | const size_t dc_group_xsize = DivCeil(xsize, dc_group_size); |
1776 | 0 | const size_t dc_group_ysize = DivCeil(ysize, dc_group_size); |
1777 | 0 | const size_t num_groups = group_xsize * group_ysize; |
1778 | 0 | const size_t num_dc_groups = dc_group_xsize * dc_group_ysize; |
1779 | 0 | const size_t num_sections = 2 + num_dc_groups + num_passes * num_groups; |
1780 | 0 | permutation.resize(num_sections); |
1781 | 0 | size_t new_ix = 0; |
1782 | | // DC Global is first |
1783 | 0 | permutation[0] = new_ix++; |
1784 | | // TODO(szabadka) Change the dc group order to center-first. |
1785 | 0 | for (size_t dc_y = 0; dc_y < dc_group_ysize; ++dc_y) { |
1786 | 0 | for (size_t dc_x = 0; dc_x < dc_group_xsize; ++dc_x) { |
1787 | 0 | size_t dc_ix = dc_y * dc_group_xsize + dc_x; |
1788 | 0 | dc_group_order.push_back(dc_ix); |
1789 | 0 | permutation[1 + dc_ix] = new_ix++; |
1790 | 0 | size_t ac_y0 = dc_y * kBlockDim; |
1791 | 0 | size_t ac_x0 = dc_x * kBlockDim; |
1792 | 0 | size_t ac_y1 = std::min<size_t>(group_ysize, ac_y0 + kBlockDim); |
1793 | 0 | size_t ac_x1 = std::min<size_t>(group_xsize, ac_x0 + kBlockDim); |
1794 | 0 | for (size_t pass = 0; pass < num_passes; ++pass) { |
1795 | 0 | for (size_t ac_y = ac_y0; ac_y < ac_y1; ++ac_y) { |
1796 | 0 | for (size_t ac_x = ac_x0; ac_x < ac_x1; ++ac_x) { |
1797 | 0 | size_t group_ix = ac_y * group_xsize + ac_x; |
1798 | 0 | size_t old_ix = |
1799 | 0 | AcGroupIndex(pass, group_ix, num_groups, num_dc_groups); |
1800 | 0 | permutation[old_ix] = new_ix++; |
1801 | 0 | } |
1802 | 0 | } |
1803 | 0 | } |
1804 | 0 | } |
1805 | 0 | } |
1806 | | // AC Global is last |
1807 | 0 | permutation[1 + num_dc_groups] = new_ix++; |
1808 | 0 | JXL_ASSERT(new_ix == num_sections); |
1809 | 0 | } |
1810 | | |
1811 | | constexpr size_t kGroupSizeOffset[4] = { |
1812 | | static_cast<size_t>(0), |
1813 | | static_cast<size_t>(1024), |
1814 | | static_cast<size_t>(17408), |
1815 | | static_cast<size_t>(4211712), |
1816 | | }; |
1817 | | constexpr size_t kTOCBits[4] = {12, 16, 24, 32}; |
1818 | | |
1819 | 0 | size_t TOCBucket(size_t group_size) { |
1820 | 0 | size_t bucket = 0; |
1821 | 0 | while (bucket < 3 && group_size >= kGroupSizeOffset[bucket + 1]) ++bucket; |
1822 | 0 | return bucket; |
1823 | 0 | } |
1824 | | |
1825 | 0 | size_t TOCSize(const std::vector<size_t>& group_sizes) { |
1826 | 0 | size_t toc_bits = 0; |
1827 | 0 | for (size_t group_size : group_sizes) { |
1828 | 0 | toc_bits += kTOCBits[TOCBucket(group_size)]; |
1829 | 0 | } |
1830 | 0 | return (toc_bits + 7) / 8; |
1831 | 0 | } |
1832 | | |
1833 | | PaddedBytes EncodeTOC(JxlMemoryManager* memory_manager, |
1834 | 0 | const std::vector<size_t>& group_sizes, AuxOut* aux_out) { |
1835 | 0 | BitWriter writer{memory_manager}; |
1836 | 0 | BitWriter::Allotment allotment(&writer, 32 * group_sizes.size()); |
1837 | 0 | for (size_t group_size : group_sizes) { |
1838 | 0 | JXL_CHECK(U32Coder::Write(kTocDist, group_size, &writer)); |
1839 | 0 | } |
1840 | 0 | writer.ZeroPadToByte(); // before first group |
1841 | 0 | allotment.ReclaimAndCharge(&writer, kLayerTOC, aux_out); |
1842 | 0 | return std::move(writer).TakeBytes(); |
1843 | 0 | } |
1844 | | |
1845 | | void ComputeGroupDataOffset(size_t frame_header_size, size_t dc_global_size, |
1846 | | size_t num_sections, size_t& min_dc_global_size, |
1847 | 0 | size_t& group_offset) { |
1848 | 0 | size_t max_toc_bits = (num_sections - 1) * 32; |
1849 | 0 | size_t min_toc_bits = (num_sections - 1) * 12; |
1850 | 0 | size_t max_padding = (max_toc_bits - min_toc_bits + 7) / 8; |
1851 | 0 | min_dc_global_size = dc_global_size; |
1852 | 0 | size_t dc_global_bucket = TOCBucket(min_dc_global_size); |
1853 | 0 | while (TOCBucket(min_dc_global_size + max_padding) > dc_global_bucket) { |
1854 | 0 | dc_global_bucket = TOCBucket(min_dc_global_size + max_padding); |
1855 | 0 | min_dc_global_size = kGroupSizeOffset[dc_global_bucket]; |
1856 | 0 | } |
1857 | 0 | JXL_ASSERT(TOCBucket(min_dc_global_size) == dc_global_bucket); |
1858 | 0 | JXL_ASSERT(TOCBucket(min_dc_global_size + max_padding) == dc_global_bucket); |
1859 | 0 | max_toc_bits += kTOCBits[dc_global_bucket]; |
1860 | 0 | size_t max_toc_size = (max_toc_bits + 7) / 8; |
1861 | 0 | group_offset = frame_header_size + max_toc_size + min_dc_global_size; |
1862 | 0 | } |
1863 | | |
1864 | | size_t ComputeDcGlobalPadding(const std::vector<size_t>& group_sizes, |
1865 | | size_t frame_header_size, |
1866 | | size_t group_data_offset, |
1867 | 0 | size_t min_dc_global_size) { |
1868 | 0 | std::vector<size_t> new_group_sizes = group_sizes; |
1869 | 0 | new_group_sizes[0] = min_dc_global_size; |
1870 | 0 | size_t toc_size = TOCSize(new_group_sizes); |
1871 | 0 | size_t actual_offset = frame_header_size + toc_size + group_sizes[0]; |
1872 | 0 | return group_data_offset - actual_offset; |
1873 | 0 | } |
1874 | | |
1875 | | Status OutputGroups(std::vector<std::unique_ptr<BitWriter>>&& group_codes, |
1876 | | std::vector<size_t>* group_sizes, |
1877 | 0 | JxlEncoderOutputProcessorWrapper* output_processor) { |
1878 | 0 | JXL_ASSERT(group_codes.size() >= 4); |
1879 | 0 | { |
1880 | 0 | PaddedBytes dc_group = std::move(*group_codes[1]).TakeBytes(); |
1881 | 0 | group_sizes->push_back(dc_group.size()); |
1882 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, dc_group)); |
1883 | 0 | } |
1884 | 0 | for (size_t i = 3; i < group_codes.size(); ++i) { |
1885 | 0 | PaddedBytes ac_group = std::move(*group_codes[i]).TakeBytes(); |
1886 | 0 | group_sizes->push_back(ac_group.size()); |
1887 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, ac_group)); |
1888 | 0 | } |
1889 | 0 | return true; |
1890 | 0 | } |
1891 | | |
1892 | | void RemoveUnusedHistograms(std::vector<uint8_t>& context_map, |
1893 | 0 | EntropyEncodingData& codes) { |
1894 | 0 | std::vector<int> remap(256, -1); |
1895 | 0 | std::vector<uint8_t> inv_remap; |
1896 | 0 | for (uint8_t& context : context_map) { |
1897 | 0 | const uint8_t histo_ix = context; |
1898 | 0 | if (remap[histo_ix] == -1) { |
1899 | 0 | remap[histo_ix] = inv_remap.size(); |
1900 | 0 | inv_remap.push_back(histo_ix); |
1901 | 0 | } |
1902 | 0 | context = remap[histo_ix]; |
1903 | 0 | } |
1904 | 0 | EntropyEncodingData new_codes; |
1905 | 0 | new_codes.use_prefix_code = codes.use_prefix_code; |
1906 | 0 | new_codes.lz77 = codes.lz77; |
1907 | 0 | for (uint8_t histo_idx : inv_remap) { |
1908 | 0 | new_codes.encoding_info.emplace_back( |
1909 | 0 | std::move(codes.encoding_info[histo_idx])); |
1910 | 0 | new_codes.uint_config.emplace_back(codes.uint_config[histo_idx]); |
1911 | 0 | new_codes.encoded_histograms.emplace_back( |
1912 | 0 | std::move(codes.encoded_histograms[histo_idx])); |
1913 | 0 | } |
1914 | 0 | codes = std::move(new_codes); |
1915 | 0 | } |
1916 | | |
1917 | | Status OutputAcGlobal(PassesEncoderState& enc_state, |
1918 | | const FrameDimensions& frame_dim, |
1919 | | std::vector<size_t>* group_sizes, |
1920 | | JxlEncoderOutputProcessorWrapper* output_processor, |
1921 | 0 | AuxOut* aux_out) { |
1922 | 0 | JXL_ASSERT(frame_dim.num_groups > 1); |
1923 | 0 | JxlMemoryManager* memory_manager = enc_state.memory_manager(); |
1924 | 0 | BitWriter writer{memory_manager}; |
1925 | 0 | { |
1926 | 0 | size_t num_histo_bits = CeilLog2Nonzero(frame_dim.num_groups); |
1927 | 0 | BitWriter::Allotment allotment(&writer, num_histo_bits + 1); |
1928 | 0 | writer.Write(1, 1); // default dequant matrices |
1929 | 0 | writer.Write(num_histo_bits, frame_dim.num_dc_groups - 1); |
1930 | 0 | allotment.ReclaimAndCharge(&writer, kLayerAC, aux_out); |
1931 | 0 | } |
1932 | 0 | const PassesSharedState& shared = enc_state.shared; |
1933 | 0 | for (size_t i = 0; i < enc_state.progressive_splitter.GetNumPasses(); i++) { |
1934 | | // Encode coefficient orders. |
1935 | 0 | size_t order_bits = 0; |
1936 | 0 | JXL_RETURN_IF_ERROR( |
1937 | 0 | U32Coder::CanEncode(kOrderEnc, enc_state.used_orders[i], &order_bits)); |
1938 | 0 | BitWriter::Allotment allotment(&writer, order_bits); |
1939 | 0 | JXL_CHECK(U32Coder::Write(kOrderEnc, enc_state.used_orders[i], &writer)); |
1940 | 0 | allotment.ReclaimAndCharge(&writer, kLayerOrder, aux_out); |
1941 | 0 | EncodeCoeffOrders(enc_state.used_orders[i], |
1942 | 0 | &shared.coeff_orders[i * shared.coeff_order_size], |
1943 | 0 | &writer, kLayerOrder, aux_out); |
1944 | | // Fix up context map and entropy codes to remove any fix histograms that |
1945 | | // were not selected by clustering. |
1946 | 0 | RemoveUnusedHistograms(enc_state.passes[i].context_map, |
1947 | 0 | enc_state.passes[i].codes); |
1948 | 0 | EncodeHistograms(enc_state.passes[i].context_map, enc_state.passes[i].codes, |
1949 | 0 | &writer, kLayerAC, aux_out); |
1950 | 0 | } |
1951 | 0 | { |
1952 | 0 | BitWriter::Allotment allotment(&writer, 8); |
1953 | 0 | writer.ZeroPadToByte(); // end of group. |
1954 | 0 | allotment.ReclaimAndCharge(&writer, kLayerAC, aux_out); |
1955 | 0 | } |
1956 | 0 | PaddedBytes ac_global = std::move(writer).TakeBytes(); |
1957 | 0 | group_sizes->push_back(ac_global.size()); |
1958 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, ac_global)); |
1959 | 0 | return true; |
1960 | 0 | } |
1961 | | |
1962 | | Status EncodeFrameStreaming(JxlMemoryManager* memory_manager, |
1963 | | const CompressParams& cparams, |
1964 | | const FrameInfo& frame_info, |
1965 | | const CodecMetadata* metadata, |
1966 | | JxlEncoderChunkedFrameAdapter& frame_data, |
1967 | | const JxlCmsInterface& cms, ThreadPool* pool, |
1968 | | JxlEncoderOutputProcessorWrapper* output_processor, |
1969 | 0 | AuxOut* aux_out) { |
1970 | 0 | PassesEncoderState enc_state{memory_manager}; |
1971 | 0 | SetProgressiveMode(cparams, &enc_state.progressive_splitter); |
1972 | 0 | FrameHeader frame_header(metadata); |
1973 | 0 | std::unique_ptr<jpeg::JPEGData> jpeg_data; |
1974 | 0 | if (frame_data.IsJPEG()) { |
1975 | 0 | jpeg_data = make_unique<jpeg::JPEGData>(frame_data.TakeJPEGData()); |
1976 | 0 | } |
1977 | 0 | JXL_RETURN_IF_ERROR(MakeFrameHeader(frame_data.xsize, frame_data.ysize, |
1978 | 0 | cparams, enc_state.progressive_splitter, |
1979 | 0 | frame_info, jpeg_data.get(), true, |
1980 | 0 | &frame_header)); |
1981 | 0 | const size_t num_passes = enc_state.progressive_splitter.GetNumPasses(); |
1982 | 0 | ModularFrameEncoder enc_modular(memory_manager, frame_header, cparams, true); |
1983 | 0 | std::vector<coeff_order_t> permutation; |
1984 | 0 | std::vector<size_t> dc_group_order; |
1985 | 0 | size_t group_size = frame_header.ToFrameDimensions().group_dim; |
1986 | 0 | ComputePermutationForStreaming(frame_data.xsize, frame_data.ysize, group_size, |
1987 | 0 | num_passes, permutation, dc_group_order); |
1988 | 0 | enc_state.shared.num_histograms = dc_group_order.size(); |
1989 | 0 | size_t dc_group_size = group_size * kBlockDim; |
1990 | 0 | size_t dc_group_xsize = DivCeil(frame_data.xsize, dc_group_size); |
1991 | 0 | size_t min_dc_global_size = 0; |
1992 | 0 | size_t group_data_offset = 0; |
1993 | 0 | PaddedBytes frame_header_bytes{memory_manager}; |
1994 | 0 | PaddedBytes dc_global_bytes{memory_manager}; |
1995 | 0 | std::vector<size_t> group_sizes; |
1996 | 0 | size_t start_pos = output_processor->CurrentPosition(); |
1997 | 0 | for (size_t i = 0; i < dc_group_order.size(); ++i) { |
1998 | 0 | size_t dc_ix = dc_group_order[i]; |
1999 | 0 | size_t dc_y = dc_ix / dc_group_xsize; |
2000 | 0 | size_t dc_x = dc_ix % dc_group_xsize; |
2001 | 0 | size_t y0 = dc_y * dc_group_size; |
2002 | 0 | size_t x0 = dc_x * dc_group_size; |
2003 | 0 | size_t ysize = std::min<size_t>(dc_group_size, frame_data.ysize - y0); |
2004 | 0 | size_t xsize = std::min<size_t>(dc_group_size, frame_data.xsize - x0); |
2005 | 0 | size_t group_xsize = DivCeil(xsize, group_size); |
2006 | 0 | size_t group_ysize = DivCeil(ysize, group_size); |
2007 | 0 | JXL_DEBUG_V(2, |
2008 | 0 | "Encoding DC group #%" PRIuS " dc_y = %" PRIuS " dc_x = %" PRIuS |
2009 | 0 | " (x0, y0) = (%" PRIuS ", %" PRIuS ") (xsize, ysize) = (%" PRIuS |
2010 | 0 | ", %" PRIuS ")", |
2011 | 0 | dc_ix, dc_y, dc_x, x0, y0, xsize, ysize); |
2012 | 0 | enc_state.streaming_mode = true; |
2013 | 0 | enc_state.initialize_global_state = (i == 0); |
2014 | 0 | enc_state.dc_group_index = dc_ix; |
2015 | 0 | enc_state.histogram_idx = std::vector<size_t>(group_xsize * group_ysize, i); |
2016 | 0 | std::vector<std::unique_ptr<BitWriter>> group_codes; |
2017 | 0 | JXL_RETURN_IF_ERROR(ComputeEncodingData( |
2018 | 0 | cparams, frame_info, metadata, frame_data, jpeg_data.get(), x0, y0, |
2019 | 0 | xsize, ysize, cms, pool, frame_header, enc_modular, enc_state, |
2020 | 0 | &group_codes, aux_out)); |
2021 | 0 | JXL_ASSERT(enc_state.special_frames.empty()); |
2022 | 0 | if (i == 0) { |
2023 | 0 | BitWriter writer{memory_manager}; |
2024 | 0 | JXL_RETURN_IF_ERROR(WriteFrameHeader(frame_header, &writer, aux_out)); |
2025 | 0 | BitWriter::Allotment allotment(&writer, 8); |
2026 | 0 | writer.Write(1, 1); // write permutation |
2027 | 0 | EncodePermutation(permutation.data(), /*skip=*/0, permutation.size(), |
2028 | 0 | &writer, kLayerHeader, aux_out); |
2029 | 0 | writer.ZeroPadToByte(); |
2030 | 0 | allotment.ReclaimAndCharge(&writer, kLayerHeader, aux_out); |
2031 | 0 | frame_header_bytes = std::move(writer).TakeBytes(); |
2032 | 0 | dc_global_bytes = std::move(*group_codes[0]).TakeBytes(); |
2033 | 0 | ComputeGroupDataOffset(frame_header_bytes.size(), dc_global_bytes.size(), |
2034 | 0 | permutation.size(), min_dc_global_size, |
2035 | 0 | group_data_offset); |
2036 | 0 | JXL_DEBUG_V(2, "Frame header size: %" PRIuS, frame_header_bytes.size()); |
2037 | 0 | JXL_DEBUG_V(2, "DC global size: %" PRIuS ", min size for TOC: %" PRIuS, |
2038 | 0 | dc_global_bytes.size(), min_dc_global_size); |
2039 | 0 | JXL_DEBUG_V(2, "Num groups: %" PRIuS " group data offset: %" PRIuS, |
2040 | 0 | permutation.size(), group_data_offset); |
2041 | 0 | group_sizes.push_back(dc_global_bytes.size()); |
2042 | 0 | output_processor->Seek(start_pos + group_data_offset); |
2043 | 0 | } |
2044 | 0 | JXL_RETURN_IF_ERROR( |
2045 | 0 | OutputGroups(std::move(group_codes), &group_sizes, output_processor)); |
2046 | 0 | } |
2047 | 0 | if (frame_header.encoding == FrameEncoding::kVarDCT) { |
2048 | 0 | JXL_RETURN_IF_ERROR( |
2049 | 0 | OutputAcGlobal(enc_state, frame_header.ToFrameDimensions(), |
2050 | 0 | &group_sizes, output_processor, aux_out)); |
2051 | 0 | } else { |
2052 | 0 | group_sizes.push_back(0); |
2053 | 0 | } |
2054 | 0 | JXL_ASSERT(group_sizes.size() == permutation.size()); |
2055 | 0 | size_t end_pos = output_processor->CurrentPosition(); |
2056 | 0 | output_processor->Seek(start_pos); |
2057 | 0 | size_t padding_size = |
2058 | 0 | ComputeDcGlobalPadding(group_sizes, frame_header_bytes.size(), |
2059 | 0 | group_data_offset, min_dc_global_size); |
2060 | 0 | group_sizes[0] += padding_size; |
2061 | 0 | PaddedBytes toc_bytes = EncodeTOC(memory_manager, group_sizes, aux_out); |
2062 | 0 | std::vector<uint8_t> padding_bytes(padding_size); |
2063 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, frame_header_bytes)); |
2064 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, toc_bytes)); |
2065 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, dc_global_bytes)); |
2066 | 0 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, padding_bytes)); |
2067 | 0 | JXL_DEBUG_V(2, "TOC size: %" PRIuS " padding bytes after DC global: %" PRIuS, |
2068 | 0 | toc_bytes.size(), padding_size); |
2069 | 0 | JXL_ASSERT(output_processor->CurrentPosition() == |
2070 | 0 | start_pos + group_data_offset); |
2071 | 0 | output_processor->Seek(end_pos); |
2072 | 0 | return true; |
2073 | 0 | } |
2074 | | |
2075 | | Status EncodeFrameOneShot(JxlMemoryManager* memory_manager, |
2076 | | const CompressParams& cparams, |
2077 | | const FrameInfo& frame_info, |
2078 | | const CodecMetadata* metadata, |
2079 | | JxlEncoderChunkedFrameAdapter& frame_data, |
2080 | | const JxlCmsInterface& cms, ThreadPool* pool, |
2081 | | JxlEncoderOutputProcessorWrapper* output_processor, |
2082 | 52 | AuxOut* aux_out) { |
2083 | 52 | PassesEncoderState enc_state{memory_manager}; |
2084 | 52 | SetProgressiveMode(cparams, &enc_state.progressive_splitter); |
2085 | 52 | FrameHeader frame_header(metadata); |
2086 | 52 | std::unique_ptr<jpeg::JPEGData> jpeg_data; |
2087 | 52 | if (frame_data.IsJPEG()) { |
2088 | 0 | jpeg_data = make_unique<jpeg::JPEGData>(frame_data.TakeJPEGData()); |
2089 | 0 | } |
2090 | 52 | JXL_RETURN_IF_ERROR(MakeFrameHeader(frame_data.xsize, frame_data.ysize, |
2091 | 52 | cparams, enc_state.progressive_splitter, |
2092 | 52 | frame_info, jpeg_data.get(), false, |
2093 | 52 | &frame_header)); |
2094 | 52 | const size_t num_passes = enc_state.progressive_splitter.GetNumPasses(); |
2095 | 52 | ModularFrameEncoder enc_modular(memory_manager, frame_header, cparams, false); |
2096 | 52 | std::vector<std::unique_ptr<BitWriter>> group_codes; |
2097 | 52 | JXL_RETURN_IF_ERROR(ComputeEncodingData( |
2098 | 52 | cparams, frame_info, metadata, frame_data, jpeg_data.get(), 0, 0, |
2099 | 52 | frame_data.xsize, frame_data.ysize, cms, pool, frame_header, enc_modular, |
2100 | 52 | enc_state, &group_codes, aux_out)); |
2101 | | |
2102 | 52 | BitWriter writer{memory_manager}; |
2103 | 52 | writer.AppendByteAligned(enc_state.special_frames); |
2104 | 52 | JXL_RETURN_IF_ERROR(WriteFrameHeader(frame_header, &writer, aux_out)); |
2105 | | |
2106 | 52 | std::vector<coeff_order_t> permutation; |
2107 | 52 | JXL_RETURN_IF_ERROR(PermuteGroups(cparams, enc_state.shared.frame_dim, |
2108 | 52 | num_passes, &permutation, &group_codes)); |
2109 | | |
2110 | 52 | JXL_RETURN_IF_ERROR( |
2111 | 52 | WriteGroupOffsets(group_codes, permutation, &writer, aux_out)); |
2112 | | |
2113 | 52 | writer.AppendByteAligned(group_codes); |
2114 | 52 | PaddedBytes frame_bytes = std::move(writer).TakeBytes(); |
2115 | 52 | JXL_RETURN_IF_ERROR(AppendData(*output_processor, frame_bytes)); |
2116 | | |
2117 | 52 | return true; |
2118 | 52 | } |
2119 | | |
2120 | | } // namespace |
2121 | | |
2122 | | Status EncodeFrame(JxlMemoryManager* memory_manager, |
2123 | | const CompressParams& cparams_orig, |
2124 | | const FrameInfo& frame_info, const CodecMetadata* metadata, |
2125 | | JxlEncoderChunkedFrameAdapter& frame_data, |
2126 | | const JxlCmsInterface& cms, ThreadPool* pool, |
2127 | | JxlEncoderOutputProcessorWrapper* output_processor, |
2128 | 52 | AuxOut* aux_out) { |
2129 | 52 | CompressParams cparams = cparams_orig; |
2130 | 52 | if (cparams.speed_tier == SpeedTier::kTectonicPlate && |
2131 | 52 | !cparams.IsLossless()) { |
2132 | 0 | cparams.speed_tier = SpeedTier::kGlacier; |
2133 | 0 | } |
2134 | | // Lightning mode is handled externally, so switch to Thunder mode to handle |
2135 | | // potentially weird cases. |
2136 | 52 | if (cparams.speed_tier == SpeedTier::kLightning) { |
2137 | 0 | cparams.speed_tier = SpeedTier::kThunder; |
2138 | 0 | } |
2139 | 52 | if (cparams.speed_tier == SpeedTier::kTectonicPlate) { |
2140 | 0 | std::vector<CompressParams> all_params; |
2141 | 0 | std::vector<size_t> size; |
2142 | |
|
2143 | 0 | CompressParams cparams_attempt = cparams_orig; |
2144 | 0 | cparams_attempt.speed_tier = SpeedTier::kGlacier; |
2145 | 0 | cparams_attempt.options.max_properties = 4; |
2146 | |
|
2147 | 0 | for (float x : {0.0f, 80.f}) { |
2148 | 0 | cparams_attempt.channel_colors_percent = x; |
2149 | 0 | for (float y : {0.0f, 95.0f}) { |
2150 | 0 | cparams_attempt.channel_colors_pre_transform_percent = y; |
2151 | | // 70000 ensures that the number of palette colors is representable in |
2152 | | // modular headers. |
2153 | 0 | for (int K : {0, 1 << 10, 70000}) { |
2154 | 0 | cparams_attempt.palette_colors = K; |
2155 | 0 | for (int tree_mode : |
2156 | 0 | {-1, static_cast<int>(ModularOptions::TreeMode::kNoWP), |
2157 | 0 | static_cast<int>(ModularOptions::TreeMode::kDefault)}) { |
2158 | 0 | if (tree_mode == -1) { |
2159 | | // LZ77 only |
2160 | 0 | cparams_attempt.options.nb_repeats = 0; |
2161 | 0 | } else { |
2162 | 0 | cparams_attempt.options.nb_repeats = 1; |
2163 | 0 | cparams_attempt.options.wp_tree_mode = |
2164 | 0 | static_cast<ModularOptions::TreeMode>(tree_mode); |
2165 | 0 | } |
2166 | 0 | for (Predictor pred : {Predictor::Zero, Predictor::Variable}) { |
2167 | 0 | cparams_attempt.options.predictor = pred; |
2168 | 0 | for (int g : {0, -1, 3}) { |
2169 | 0 | cparams_attempt.modular_group_size_shift = g; |
2170 | 0 | for (Override patches : {Override::kDefault, Override::kOff}) { |
2171 | 0 | cparams_attempt.patches = patches; |
2172 | 0 | all_params.push_back(cparams_attempt); |
2173 | 0 | } |
2174 | 0 | } |
2175 | 0 | } |
2176 | 0 | } |
2177 | 0 | } |
2178 | 0 | } |
2179 | 0 | } |
2180 | |
|
2181 | 0 | size.resize(all_params.size()); |
2182 | |
|
2183 | 0 | std::atomic<bool> has_error{false}; |
2184 | |
|
2185 | 0 | JXL_RETURN_IF_ERROR(RunOnPool( |
2186 | 0 | pool, 0, all_params.size(), ThreadPool::NoInit, |
2187 | 0 | [&](size_t task, size_t) { |
2188 | 0 | if (has_error) return; |
2189 | 0 | std::vector<uint8_t> output(64); |
2190 | 0 | uint8_t* next_out = output.data(); |
2191 | 0 | size_t avail_out = output.size(); |
2192 | 0 | JxlEncoderOutputProcessorWrapper local_output(memory_manager); |
2193 | 0 | local_output.SetAvailOut(&next_out, &avail_out); |
2194 | 0 | if (!EncodeFrame(memory_manager, all_params[task], frame_info, |
2195 | 0 | metadata, frame_data, cms, nullptr, &local_output, |
2196 | 0 | aux_out)) { |
2197 | 0 | has_error = true; |
2198 | 0 | return; |
2199 | 0 | } |
2200 | 0 | size[task] = local_output.CurrentPosition(); |
2201 | 0 | }, |
2202 | 0 | "Compress kTectonicPlate")); |
2203 | 0 | if (has_error) return JXL_FAILURE("Compress kTectonicPlate failed"); |
2204 | | |
2205 | 0 | size_t best_idx = 0; |
2206 | 0 | for (size_t i = 1; i < all_params.size(); i++) { |
2207 | 0 | if (size[best_idx] > size[i]) { |
2208 | 0 | best_idx = i; |
2209 | 0 | } |
2210 | 0 | } |
2211 | 0 | cparams = all_params[best_idx]; |
2212 | 0 | } |
2213 | | |
2214 | 52 | JXL_RETURN_IF_ERROR(ParamsPostInit(&cparams)); |
2215 | | |
2216 | 52 | if (cparams.butteraugli_distance < 0) { |
2217 | 0 | return JXL_FAILURE("Expected non-negative distance"); |
2218 | 0 | } |
2219 | | |
2220 | 52 | if (cparams.progressive_dc < 0) { |
2221 | 52 | if (cparams.progressive_dc != -1) { |
2222 | 0 | return JXL_FAILURE("Invalid progressive DC setting value (%d)", |
2223 | 0 | cparams.progressive_dc); |
2224 | 0 | } |
2225 | 52 | cparams.progressive_dc = 0; |
2226 | 52 | } |
2227 | 52 | if (cparams.ec_resampling < cparams.resampling) { |
2228 | 0 | cparams.ec_resampling = cparams.resampling; |
2229 | 0 | } |
2230 | 52 | if (cparams.resampling > 1 || frame_info.is_preview) { |
2231 | 0 | cparams.progressive_dc = 0; |
2232 | 0 | } |
2233 | | |
2234 | 52 | if (frame_info.dc_level + cparams.progressive_dc > 4) { |
2235 | 0 | return JXL_FAILURE("Too many levels of progressive DC"); |
2236 | 0 | } |
2237 | | |
2238 | 52 | if (cparams.butteraugli_distance != 0 && |
2239 | 52 | cparams.butteraugli_distance < kMinButteraugliDistance) { |
2240 | 0 | return JXL_FAILURE("Butteraugli distance is too low (%f)", |
2241 | 0 | cparams.butteraugli_distance); |
2242 | 0 | } |
2243 | | |
2244 | 52 | if (frame_data.IsJPEG()) { |
2245 | 0 | cparams.gaborish = Override::kOff; |
2246 | 0 | cparams.epf = 0; |
2247 | 0 | cparams.modular_mode = false; |
2248 | 0 | } |
2249 | | |
2250 | 52 | if (frame_data.xsize == 0 || frame_data.ysize == 0) { |
2251 | 0 | return JXL_FAILURE("Empty image"); |
2252 | 0 | } |
2253 | | |
2254 | | // Assert that this metadata is correctly set up for the compression params, |
2255 | | // this should have been done by enc_file.cc |
2256 | 52 | JXL_ASSERT(metadata->m.xyb_encoded == |
2257 | 52 | (cparams.color_transform == ColorTransform::kXYB)); |
2258 | | |
2259 | 52 | if (frame_data.IsJPEG() && cparams.color_transform == ColorTransform::kXYB) { |
2260 | 0 | return JXL_FAILURE("Can't add JPEG frame to XYB codestream"); |
2261 | 0 | } |
2262 | | |
2263 | 52 | if (CanDoStreamingEncoding(cparams, frame_info, *metadata, frame_data)) { |
2264 | 0 | return EncodeFrameStreaming(memory_manager, cparams, frame_info, metadata, |
2265 | 0 | frame_data, cms, pool, output_processor, |
2266 | 0 | aux_out); |
2267 | 52 | } else { |
2268 | 52 | return EncodeFrameOneShot(memory_manager, cparams, frame_info, metadata, |
2269 | 52 | frame_data, cms, pool, output_processor, aux_out); |
2270 | 52 | } |
2271 | 52 | } |
2272 | | |
2273 | | Status EncodeFrame(JxlMemoryManager* memory_manager, |
2274 | | const CompressParams& cparams_orig, |
2275 | | const FrameInfo& frame_info, const CodecMetadata* metadata, |
2276 | | const ImageBundle& ib, const JxlCmsInterface& cms, |
2277 | 0 | ThreadPool* pool, BitWriter* writer, AuxOut* aux_out) { |
2278 | 0 | JxlEncoderChunkedFrameAdapter frame_data(ib.xsize(), ib.ysize(), |
2279 | 0 | ib.extra_channels().size()); |
2280 | 0 | std::vector<uint8_t> color; |
2281 | 0 | if (ib.IsJPEG()) { |
2282 | 0 | frame_data.SetJPEGData(*ib.jpeg_data); |
2283 | 0 | } else { |
2284 | 0 | uint32_t num_channels = |
2285 | 0 | ib.IsGray() && frame_info.ib_needs_color_transform ? 1 : 3; |
2286 | 0 | size_t stride = ib.xsize() * num_channels * 4; |
2287 | 0 | color.resize(ib.ysize() * stride); |
2288 | 0 | JXL_RETURN_IF_ERROR(ConvertToExternal( |
2289 | 0 | ib, /*bits_per_sample=*/32, /*float_out=*/true, num_channels, |
2290 | 0 | JXL_NATIVE_ENDIAN, stride, pool, color.data(), color.size(), |
2291 | 0 | /*out_callback=*/{}, Orientation::kIdentity)); |
2292 | 0 | JxlPixelFormat format{num_channels, JXL_TYPE_FLOAT, JXL_NATIVE_ENDIAN, 0}; |
2293 | 0 | frame_data.SetFromBuffer(0, color.data(), color.size(), format); |
2294 | 0 | } |
2295 | 0 | for (size_t ec = 0; ec < ib.extra_channels().size(); ++ec) { |
2296 | 0 | JxlPixelFormat ec_format{1, JXL_TYPE_FLOAT, JXL_NATIVE_ENDIAN, 0}; |
2297 | 0 | size_t ec_stride = ib.xsize() * 4; |
2298 | 0 | std::vector<uint8_t> ec_data(ib.ysize() * ec_stride); |
2299 | 0 | const ImageF* channel = &ib.extra_channels()[ec]; |
2300 | 0 | JXL_RETURN_IF_ERROR(ConvertChannelsToExternal( |
2301 | 0 | &channel, 1, |
2302 | 0 | /*bits_per_sample=*/32, |
2303 | 0 | /*float_out=*/true, JXL_NATIVE_ENDIAN, ec_stride, pool, ec_data.data(), |
2304 | 0 | ec_data.size(), /*out_callback=*/{}, Orientation::kIdentity)); |
2305 | 0 | frame_data.SetFromBuffer(1 + ec, ec_data.data(), ec_data.size(), ec_format); |
2306 | 0 | } |
2307 | 0 | FrameInfo fi = frame_info; |
2308 | 0 | fi.origin = ib.origin; |
2309 | 0 | fi.blend = ib.blend; |
2310 | 0 | fi.blendmode = ib.blendmode; |
2311 | 0 | fi.duration = ib.duration; |
2312 | 0 | fi.timecode = ib.timecode; |
2313 | 0 | fi.name = ib.name; |
2314 | 0 | std::vector<uint8_t> output(64); |
2315 | 0 | uint8_t* next_out = output.data(); |
2316 | 0 | size_t avail_out = output.size(); |
2317 | 0 | JxlEncoderOutputProcessorWrapper output_processor(memory_manager); |
2318 | 0 | output_processor.SetAvailOut(&next_out, &avail_out); |
2319 | 0 | JXL_RETURN_IF_ERROR(EncodeFrame(memory_manager, cparams_orig, fi, metadata, |
2320 | 0 | frame_data, cms, pool, &output_processor, |
2321 | 0 | aux_out)); |
2322 | 0 | output_processor.SetFinalizedPosition(); |
2323 | 0 | output_processor.CopyOutput(output, next_out, avail_out); |
2324 | 0 | writer->AppendByteAligned(Bytes(output)); |
2325 | 0 | return true; |
2326 | 0 | } |
2327 | | |
2328 | | } // namespace jxl |