1 #include "jpeg_frame_view.h"
8 #include <condition_variable>
14 #include <QMouseEvent>
17 #include <movit/resource_pool.h>
18 #include <movit/init.h>
19 #include <movit/util.h>
22 #include "post_to_main_thread.h"
23 #include "vaapi_jpeg_decoder.h"
24 #include "video_stream.h"
25 #include "ycbcr_converter.h"
27 using namespace movit;
30 // Just an arbitrary order for std::map.
31 struct JPEGIDLexicalOrder
33 bool operator() (const JPEGID &a, const JPEGID &b) const
35 if (a.stream_idx != b.stream_idx)
36 return a.stream_idx < b.stream_idx;
39 return a.interpolated < b.interpolated;
44 shared_ptr<Frame> frame;
48 struct PendingDecode {
49 JPEGID primary, secondary;
50 float fade_alpha; // Irrelevant if secondary.stream_idx == -1.
51 JPEGFrameView *destination;
54 thread JPEGFrameView::jpeg_decoder_thread;
56 map<JPEGID, LRUFrame, JPEGIDLexicalOrder> cache; // Under cache_mu.
57 condition_variable any_pending_decodes, cache_updated;
58 deque<PendingDecode> pending_decodes; // Under cache_mu.
59 atomic<size_t> event_counter{0};
60 extern QGLWidget *global_share_widget;
61 extern atomic<bool> should_quit;
63 shared_ptr<Frame> decode_jpeg(const string &filename)
65 shared_ptr<Frame> frame;
66 if (vaapi_jpeg_decoding_usable) {
67 frame = decode_jpeg_vaapi(filename);
68 if (frame != nullptr) {
71 fprintf(stderr, "VA-API hardware decoding failed; falling back to software.\n");
74 frame.reset(new Frame);
76 jpeg_decompress_struct dinfo;
78 dinfo.err = jpeg_std_error(&jerr);
79 jpeg_create_decompress(&dinfo);
81 FILE *fp = fopen(filename.c_str(), "rb");
83 perror(filename.c_str());
86 jpeg_stdio_src(&dinfo, fp);
88 jpeg_read_header(&dinfo, true);
90 if (dinfo.num_components != 3) {
91 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
93 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
94 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
95 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
98 if (dinfo.comp_info[0].h_samp_factor != dinfo.max_h_samp_factor ||
99 dinfo.comp_info[0].v_samp_factor != dinfo.max_v_samp_factor || // Y' must not be subsampled.
100 dinfo.comp_info[1].h_samp_factor != dinfo.comp_info[2].h_samp_factor ||
101 dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[2].v_samp_factor || // Cb and Cr must be identically subsampled.
102 (dinfo.max_h_samp_factor % dinfo.comp_info[1].h_samp_factor) != 0 ||
103 (dinfo.max_v_samp_factor % dinfo.comp_info[1].v_samp_factor) != 0) { // No 2:3 subsampling or other weirdness.
104 fprintf(stderr, "Unsupported subsampling scheme. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
105 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
106 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
107 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
110 dinfo.raw_data_out = true;
112 jpeg_start_decompress(&dinfo);
114 frame->width = dinfo.output_width;
115 frame->height = dinfo.output_height;
116 frame->chroma_subsampling_x = dinfo.max_h_samp_factor / dinfo.comp_info[1].h_samp_factor;
117 frame->chroma_subsampling_y = dinfo.max_v_samp_factor / dinfo.comp_info[1].v_samp_factor;
119 unsigned h_mcu_size = DCTSIZE * dinfo.max_h_samp_factor;
120 unsigned v_mcu_size = DCTSIZE * dinfo.max_v_samp_factor;
121 unsigned mcu_width_blocks = (dinfo.output_width + h_mcu_size - 1) / h_mcu_size;
122 unsigned mcu_height_blocks = (dinfo.output_height + v_mcu_size - 1) / v_mcu_size;
124 unsigned luma_width_blocks = mcu_width_blocks * dinfo.comp_info[0].h_samp_factor;
125 unsigned chroma_width_blocks = mcu_width_blocks * dinfo.comp_info[1].h_samp_factor;
126 unsigned luma_height_blocks = mcu_height_blocks * dinfo.comp_info[0].v_samp_factor;
127 unsigned chroma_height_blocks = mcu_height_blocks * dinfo.comp_info[1].v_samp_factor;
129 // TODO: Decode into a PBO.
130 frame->y.reset(new uint8_t[luma_width_blocks * luma_height_blocks * DCTSIZE2]);
131 frame->cb.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
132 frame->cr.reset(new uint8_t[chroma_width_blocks * chroma_height_blocks * DCTSIZE2]);
133 frame->pitch_y = luma_width_blocks * DCTSIZE;
134 frame->pitch_chroma = chroma_width_blocks * DCTSIZE;
136 JSAMPROW yptr[v_mcu_size], cbptr[v_mcu_size], crptr[v_mcu_size];
137 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
138 for (unsigned y = 0; y < mcu_height_blocks; ++y) {
139 // NOTE: The last elements of cbptr/crptr will be unused for vertically subsampled chroma.
140 for (unsigned yy = 0; yy < v_mcu_size; ++yy) {
141 yptr[yy] = frame->y.get() + (y * DCTSIZE * dinfo.max_v_samp_factor + yy) * frame->pitch_y;
142 cbptr[yy] = frame->cb.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
143 crptr[yy] = frame->cr.get() + (y * DCTSIZE * dinfo.comp_info[1].v_samp_factor + yy) * frame->pitch_chroma;
146 jpeg_read_raw_data(&dinfo, data, v_mcu_size);
149 (void) jpeg_finish_decompress(&dinfo);
150 jpeg_destroy_decompress(&dinfo);
158 // Assumes cache_mu is held.
159 vector<size_t> lru_timestamps;
160 for (const auto &key_and_value : cache) {
161 lru_timestamps.push_back(key_and_value.second.last_used);
164 size_t cutoff_point = CACHE_SIZE / 10; // Prune away the 10% oldest ones.
165 nth_element(lru_timestamps.begin(), lru_timestamps.begin() + cutoff_point, lru_timestamps.end());
166 size_t must_be_used_after = lru_timestamps[cutoff_point];
167 for (auto it = cache.begin(); it != cache.end(); ) {
168 if (it->second.last_used < must_be_used_after) {
169 it = cache.erase(it);
176 shared_ptr<Frame> decode_jpeg_with_cache(JPEGID id, CacheMissBehavior cache_miss_behavior, bool *did_decode)
180 unique_lock<mutex> lock(cache_mu);
181 auto it = cache.find(id);
182 if (it != cache.end()) {
183 it->second.last_used = event_counter++;
184 return it->second.frame;
188 if (cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE) {
192 assert(!id.interpolated);
194 shared_ptr<Frame> frame = decode_jpeg(filename_for_frame(id.stream_idx, id.pts));
196 unique_lock<mutex> lock(cache_mu);
197 cache[id] = LRUFrame{ frame, event_counter++ };
199 if (cache.size() > CACHE_SIZE) {
205 void jpeg_decoder_thread_func()
207 size_t num_decoded = 0, num_dropped = 0;
209 pthread_setname_np(pthread_self(), "JPEGDecoder");
210 while (!should_quit.load()) {
211 PendingDecode decode;
212 CacheMissBehavior cache_miss_behavior = DECODE_IF_NOT_IN_CACHE;
214 unique_lock<mutex> lock(cache_mu); // TODO: Perhaps under another lock?
215 any_pending_decodes.wait(lock, [] {
216 return !pending_decodes.empty() || should_quit.load();
218 if (should_quit.load()) break;
219 decode = pending_decodes.front();
220 pending_decodes.pop_front();
222 size_t num_pending = 0;
223 for (const PendingDecode &other_decode : pending_decodes) {
224 if (other_decode.destination == decode.destination) {
228 if (num_pending > 3) {
229 cache_miss_behavior = RETURN_NULLPTR_IF_NOT_IN_CACHE;
233 shared_ptr<Frame> primary_frame, secondary_frame;
235 for (int subframe_idx = 0; subframe_idx < 2; ++subframe_idx) {
236 const JPEGID &id = (subframe_idx == 0 ? decode.primary : decode.secondary);
237 if (id.stream_idx == (unsigned)-1) {
238 // No secondary frame.
243 shared_ptr<Frame> frame;
244 if (id.interpolated) {
245 // Interpolated frames are never decoded by us,
246 // put directly into the cache from VideoStream.
247 unique_lock<mutex> lock(cache_mu);
248 cache_updated.wait(lock, [id] {
249 return cache.count(id) != 0 || should_quit.load();
251 if (should_quit.load()) break;
252 found_in_cache = true; // Don't count it as a decode.
254 auto it = cache.find(id);
255 assert(it != cache.end());
257 it->second.last_used = event_counter++;
258 frame = it->second.frame;
259 if (frame == nullptr) {
260 // We inserted a nullptr as signal that the frame was never
261 // interpolated and that we should stop waiting.
262 // But don't let it linger in the cache anymore.
266 frame = decode_jpeg_with_cache(id, cache_miss_behavior, &found_in_cache);
269 if (frame == nullptr) {
270 assert(id.interpolated || cache_miss_behavior == RETURN_NULLPTR_IF_NOT_IN_CACHE);
275 if (!found_in_cache) {
277 if (num_decoded % 1000 == 0) {
278 fprintf(stderr, "Decoded %zu images, dropped %zu (%.2f%% dropped)\n",
279 num_decoded, num_dropped, (100.0 * num_dropped) / (num_decoded + num_dropped));
282 if (subframe_idx == 0) {
283 primary_frame = move(frame);
285 secondary_frame = move(frame);
293 // TODO: Could we get jitter between non-interpolated and interpolated frames here?
294 decode.destination->setDecodedFrame(primary_frame, secondary_frame, decode.fade_alpha);
298 void JPEGFrameView::shutdown()
300 any_pending_decodes.notify_all();
301 jpeg_decoder_thread.join();
304 JPEGFrameView::JPEGFrameView(QWidget *parent)
305 : QGLWidget(parent, global_share_widget) {
308 void JPEGFrameView::setFrame(unsigned stream_idx, int64_t pts, bool interpolated, int secondary_stream_idx, int64_t secondary_pts, float fade_alpha)
310 current_stream_idx = stream_idx; // TODO: Does this interact with fades?
312 unique_lock<mutex> lock(cache_mu);
313 PendingDecode decode;
314 decode.primary = JPEGID{ stream_idx, pts, interpolated };
315 decode.secondary = JPEGID{ (unsigned)secondary_stream_idx, secondary_pts, /*interpolated=*/false };
316 decode.fade_alpha = fade_alpha;
317 decode.destination = this;
318 pending_decodes.push_back(decode);
319 any_pending_decodes.notify_all();
322 void JPEGFrameView::insert_interpolated_frame(unsigned stream_idx, int64_t pts, shared_ptr<Frame> frame)
324 JPEGID id{ stream_idx, pts, true };
326 // We rely on the frame not being evicted from the cache before
327 // jpeg_decoder_thread() sees it and can display it (otherwise,
328 // that thread would hang). With a default cache of 1000 elements,
329 // that would sound like a reasonable assumption.
330 unique_lock<mutex> lock(cache_mu);
331 cache[id] = LRUFrame{ std::move(frame), event_counter++ };
332 cache_updated.notify_all();
335 ResourcePool *resource_pool = nullptr;
337 void JPEGFrameView::initializeGL()
340 glDisable(GL_DEPTH_TEST);
343 static once_flag once;
345 resource_pool = new ResourcePool;
346 jpeg_decoder_thread = std::thread(jpeg_decoder_thread_func);
349 ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_RGBA, resource_pool));
351 ImageFormat inout_format;
352 inout_format.color_space = COLORSPACE_sRGB;
353 inout_format.gamma_curve = GAMMA_sRGB;
355 overlay_chain.reset(new EffectChain(overlay_base_width, overlay_base_height, resource_pool));
356 overlay_input = (movit::FlatInput *)overlay_chain->add_input(new FlatInput(inout_format, FORMAT_GRAYSCALE, GL_UNSIGNED_BYTE, overlay_base_width, overlay_base_height));
358 overlay_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
359 overlay_chain->finalize();
362 void JPEGFrameView::resizeGL(int width, int height)
365 glViewport(0, 0, width, height);
368 // Save these, as width() and height() will lie with DPI scaling.
373 void JPEGFrameView::paintGL()
375 glViewport(0, 0, gl_width, gl_height);
376 if (current_frame == nullptr) {
377 glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
378 glClear(GL_COLOR_BUFFER_BIT);
383 current_chain->render_to_screen();
385 if (overlay_image != nullptr) {
386 if (overlay_input_needs_refresh) {
387 overlay_input->set_width(overlay_width);
388 overlay_input->set_height(overlay_height);
389 overlay_input->set_pixel_data(overlay_image->bits());
391 glViewport(gl_width - overlay_width, 0, overlay_width, overlay_height);
392 overlay_chain->render_to_screen();
401 void JPEGFrameView::setDecodedFrame(shared_ptr<Frame> frame, shared_ptr<Frame> secondary_frame, float fade_alpha)
403 post_to_main_thread([this, frame, secondary_frame, fade_alpha] {
404 current_frame = frame;
405 current_secondary_frame = secondary_frame;
407 if (secondary_frame != nullptr) {
408 current_chain = ycbcr_converter->prepare_chain_for_fade(frame, secondary_frame, fade_alpha);
410 current_chain = ycbcr_converter->prepare_chain_for_conversion(frame);
416 void JPEGFrameView::mousePressEvent(QMouseEvent *event)
418 if (event->type() == QEvent::MouseButtonPress && event->button() == Qt::LeftButton) {
423 void JPEGFrameView::set_overlay(const string &text)
426 overlay_image.reset();
430 float dpr = QGuiApplication::primaryScreen()->devicePixelRatio();
431 overlay_width = lrint(overlay_base_width * dpr);
432 overlay_height = lrint(overlay_base_height * dpr);
434 overlay_image.reset(new QImage(overlay_width, overlay_height, QImage::Format_Grayscale8));
435 overlay_image->setDevicePixelRatio(dpr);
436 overlay_image->fill(0);
437 QPainter painter(overlay_image.get());
439 painter.setPen(Qt::white);
440 QFont font = painter.font();
441 font.setPointSize(12);
442 painter.setFont(font);
444 painter.drawText(QRectF(0, 0, overlay_base_width, overlay_base_height), Qt::AlignCenter, QString::fromStdString(text));
446 // Don't refresh immediately; we might not have an OpenGL context here.
447 overlay_input_needs_refresh = true;