1 #include "video_stream.h"
4 #include <libavformat/avformat.h>
5 #include <libavformat/avio.h>
8 #include "chroma_subsampler.h"
11 #include "jpeg_frame_view.h"
12 #include "movit/util.h"
14 #include "shared/context.h"
15 #include "shared/httpd.h"
16 #include "shared/mux.h"
18 #include "ycbcr_converter.h"
20 #include <epoxy/glx.h>
25 using namespace std::chrono;
27 extern HTTPD *global_httpd;
29 struct VectorDestinationManager {
30 jpeg_destination_mgr pub;
31 std::vector<uint8_t> dest;
33 VectorDestinationManager()
35 pub.init_destination = init_destination_thunk;
36 pub.empty_output_buffer = empty_output_buffer_thunk;
37 pub.term_destination = term_destination_thunk;
40 static void init_destination_thunk(j_compress_ptr ptr)
42 ((VectorDestinationManager *)(ptr->dest))->init_destination();
45 inline void init_destination()
50 static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
52 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
55 inline bool empty_output_buffer()
57 make_room(dest.size()); // Should ignore pub.free_in_buffer!
61 inline void make_room(size_t bytes_used)
63 dest.resize(bytes_used + 4096);
64 dest.resize(dest.capacity());
65 pub.next_output_byte = dest.data() + bytes_used;
66 pub.free_in_buffer = dest.size() - bytes_used;
69 static void term_destination_thunk(j_compress_ptr ptr)
71 ((VectorDestinationManager *)(ptr->dest))->term_destination();
74 inline void term_destination()
76 dest.resize(dest.size() - pub.free_in_buffer);
79 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
81 vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
83 VectorDestinationManager dest;
85 jpeg_compress_struct cinfo;
87 cinfo.err = jpeg_std_error(&jerr);
88 jpeg_create_compress(&cinfo);
90 cinfo.dest = (jpeg_destination_mgr *)&dest;
91 cinfo.input_components = 3;
92 cinfo.in_color_space = JCS_RGB;
93 jpeg_set_defaults(&cinfo);
94 constexpr int quality = 90;
95 jpeg_set_quality(&cinfo, quality, /*force_baseline=*/false);
97 cinfo.image_width = width;
98 cinfo.image_height = height;
99 cinfo.raw_data_in = true;
100 jpeg_set_colorspace(&cinfo, JCS_YCbCr);
101 cinfo.comp_info[0].h_samp_factor = 2;
102 cinfo.comp_info[0].v_samp_factor = 1;
103 cinfo.comp_info[1].h_samp_factor = 1;
104 cinfo.comp_info[1].v_samp_factor = 1;
105 cinfo.comp_info[2].h_samp_factor = 1;
106 cinfo.comp_info[2].v_samp_factor = 1;
107 cinfo.CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though.
108 jpeg_start_compress(&cinfo, true);
110 // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
111 // (and nothing else).
112 jpeg_write_marker(&cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
114 JSAMPROW yptr[8], cbptr[8], crptr[8];
115 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
116 for (unsigned y = 0; y < height; y += 8) {
117 for (unsigned yy = 0; yy < 8; ++yy) {
118 yptr[yy] = const_cast<JSAMPROW>(&y_data[(y + yy) * width]);
119 cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width / 2]);
120 crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width / 2]);
123 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
126 jpeg_finish_compress(&cinfo);
127 jpeg_destroy_compress(&cinfo);
129 return move(dest.dest);
132 VideoStream::VideoStream(AVFormatContext *file_avctx)
133 : avctx(file_avctx), output_fast_forward(file_avctx != nullptr)
135 ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
136 ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
138 GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
139 GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
140 GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
142 glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
143 glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
144 glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
145 glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
146 glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
147 glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
150 size_t width = global_flags.width, height = global_flags.height;
151 int levels = find_num_levels(width, height);
152 for (size_t i = 0; i < num_interpolate_slots; ++i) {
153 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
155 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
157 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
159 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
161 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
163 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
166 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
167 resource->owner = this;
168 resource->input_tex = input_tex[i];
169 resource->gray_tex = gray_tex[i];
170 resource->fade_y_output_tex = fade_y_output_tex[i];
171 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
172 resource->cb_tex = cb_tex[i];
173 resource->cr_tex = cr_tex[i];
174 glCreateFramebuffers(2, resource->input_fbos);
176 glCreateFramebuffers(1, &resource->fade_fbo);
179 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
181 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
183 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
185 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
187 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
189 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
192 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
193 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
195 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
197 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
200 glCreateBuffers(1, &resource->pbo);
202 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
204 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
205 interpolate_resources.push_back(move(resource));
211 if (global_flags.interpolation_quality == 0 ||
212 global_flags.interpolation_quality == 1) {
213 op = operating_point1;
214 } else if (global_flags.interpolation_quality == 2) {
215 op = operating_point2;
216 } else if (global_flags.interpolation_quality == 3) {
217 op = operating_point3;
218 } else if (global_flags.interpolation_quality == 4) {
219 op = operating_point4;
221 // Quality 0 will be changed to 1 in flags.cpp.
225 compute_flow.reset(new DISComputeFlow(width, height, op));
226 interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
227 interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
228 chroma_subsampler.reset(new ChromaSubsampler);
231 // The “last frame” is initially black.
232 unique_ptr<uint8_t[]> y(new uint8_t[global_flags.width * global_flags.height]);
233 unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[(global_flags.width / 2) * global_flags.height]);
234 memset(y.get(), 16, global_flags.width * global_flags.height);
235 memset(cb_or_cr.get(), 128, (global_flags.width / 2) * global_flags.height);
236 last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), global_flags.width, global_flags.height);
239 VideoStream::~VideoStream()
241 if (last_flow_tex != 0) {
242 compute_flow->release_texture(last_flow_tex);
245 for (const unique_ptr<InterpolatedFrameResources> &resource : interpolate_resources) {
246 glUnmapNamedBuffer(resource->pbo);
248 glDeleteBuffers(1, &resource->pbo);
250 glDeleteFramebuffers(2, resource->input_fbos);
252 glDeleteFramebuffers(1, &resource->fade_fbo);
254 glDeleteTextures(1, &resource->input_tex);
256 glDeleteTextures(1, &resource->gray_tex);
258 glDeleteTextures(1, &resource->fade_y_output_tex);
260 glDeleteTextures(1, &resource->fade_cbcr_output_tex);
262 glDeleteTextures(1, &resource->cb_tex);
264 glDeleteTextures(1, &resource->cr_tex);
267 assert(interpolate_resources.size() == num_interpolate_slots);
270 void VideoStream::start()
272 if (avctx == nullptr) {
273 avctx = avformat_alloc_context();
275 // We use Matroska, because it's pretty much the only mux where FFmpeg
276 // allows writing chroma location to override JFIF's default center placement.
277 // (Note that at the time of writing, however, FFmpeg does not correctly
278 // _read_ this information!)
279 avctx->oformat = av_guess_format("matroska", nullptr, nullptr);
281 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
282 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
283 avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
284 avctx->pb->ignore_boundary_point = 1;
286 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
289 size_t width = global_flags.width, height = global_flags.height; // Doesn't matter for MJPEG.
290 mux.reset(new Mux(avctx, width, height, Mux::CODEC_MJPEG, /*video_extradata=*/"", /*audio_codec_parameters=*/nullptr,
291 AVCOL_SPC_BT709, COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}, Mux::WITH_SUBTITLES));
293 encode_thread = thread(&VideoStream::encode_thread_func, this);
296 void VideoStream::stop()
299 queue_changed.notify_all();
301 encode_thread.join();
304 void VideoStream::clear_queue()
306 deque<QueuedFrame> q;
309 lock_guard<mutex> lock(queue_lock);
310 q = move(frame_queue);
313 // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
314 // Note that release_texture() is thread-safe.
315 for (const QueuedFrame &qf : q) {
316 if (qf.type == QueuedFrame::INTERPOLATED ||
317 qf.type == QueuedFrame::FADED_INTERPOLATED) {
318 if (qf.flow_tex != 0) {
319 compute_flow->release_texture(qf.flow_tex);
322 if (qf.type == QueuedFrame::INTERPOLATED) {
323 interpolate->release_texture(qf.output_tex);
324 interpolate->release_texture(qf.cbcr_tex);
328 // Destroy q outside the mutex, as that would be a double-lock.
331 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
332 int64_t output_pts, function<void()> &&display_func,
333 QueueSpotHolder &&queue_spot_holder,
334 FrameOnDisk frame, const string &subtitle)
336 fprintf(stderr, "output_pts=%ld original input_pts=%ld\n", output_pts, frame.pts);
338 // Preload the file from disk, so that the encoder thread does not get stalled.
339 // TODO: Consider sending it through the queue instead.
340 (void)frame_reader.read_frame(frame);
343 qf.local_pts = local_pts;
344 qf.type = QueuedFrame::ORIGINAL;
345 qf.output_pts = output_pts;
347 qf.display_func = move(display_func);
348 qf.queue_spot_holder = move(queue_spot_holder);
349 qf.subtitle = subtitle;
351 lock_guard<mutex> lock(queue_lock);
352 frame_queue.push_back(move(qf));
353 queue_changed.notify_all();
356 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
357 function<void()> &&display_func,
358 QueueSpotHolder &&queue_spot_holder,
359 FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
360 float fade_alpha, const string &subtitle)
362 fprintf(stderr, "output_pts=%ld faded input_pts=%ld,%ld fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
364 // Get the temporary OpenGL resources we need for doing the fade.
365 // (We share these with interpolated frames, which is slightly
366 // overkill, but there's no need to waste resources on keeping
367 // separate pools around.)
368 BorrowedInterpolatedFrameResources resources;
370 lock_guard<mutex> lock(queue_lock);
371 if (interpolate_resources.empty()) {
372 fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
375 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
376 interpolate_resources.pop_front();
381 shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
382 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
384 ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
387 qf.local_pts = local_pts;
388 qf.type = QueuedFrame::FADED;
389 qf.output_pts = output_pts;
390 qf.frame1 = frame1_spec;
391 qf.display_func = move(display_func);
392 qf.queue_spot_holder = move(queue_spot_holder);
393 qf.subtitle = subtitle;
395 qf.secondary_frame = frame2_spec;
397 // Subsample and split Cb/Cr.
398 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
400 // Read it down (asynchronously) to the CPU.
401 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
402 glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
404 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
406 glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3, BUFFER_OFFSET(global_flags.width * global_flags.height));
408 glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3 - (global_flags.width / 2) * global_flags.height, BUFFER_OFFSET(global_flags.width * global_flags.height + (global_flags.width / 2) * global_flags.height));
410 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
412 // Set a fence we can wait for to make sure the CPU sees the read.
413 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
415 qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
417 qf.resources = move(resources);
418 qf.local_pts = local_pts;
420 lock_guard<mutex> lock(queue_lock);
421 frame_queue.push_back(move(qf));
422 queue_changed.notify_all();
425 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
426 int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
427 QueueSpotHolder &&queue_spot_holder,
428 FrameOnDisk frame1, FrameOnDisk frame2,
429 float alpha, FrameOnDisk secondary_frame, float fade_alpha, const string &subtitle)
431 if (secondary_frame.pts != -1) {
432 fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f secondary_pts=%ld fade_alpha=%.2f\n", output_pts, frame1.pts, frame2.pts, alpha, secondary_frame.pts, fade_alpha);
434 fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
437 // Get the temporary OpenGL resources we need for doing the interpolation.
438 BorrowedInterpolatedFrameResources resources;
440 lock_guard<mutex> lock(queue_lock);
441 if (interpolate_resources.empty()) {
442 fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
445 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
446 interpolate_resources.pop_front();
450 qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
451 qf.output_pts = output_pts;
452 qf.display_decoded_func = move(display_func);
453 qf.queue_spot_holder = move(queue_spot_holder);
454 qf.local_pts = local_pts;
455 qf.subtitle = subtitle;
459 // Convert frame0 and frame1 to OpenGL textures.
460 for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
461 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
463 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
464 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], global_flags.width, global_flags.height);
467 glGenerateTextureMipmap(resources->input_tex);
469 glGenerateTextureMipmap(resources->gray_tex);
473 if (last_flow_tex != 0 && frame1 == last_frame1 && frame2 == last_frame2) {
474 // Reuse the flow from previous computation. This frequently happens
475 // if we slow down by more than 2x, so that there are multiple interpolated
476 // frames between each original.
477 flow_tex = last_flow_tex;
480 // Cache miss, so release last_flow_tex.
481 qf.flow_tex = last_flow_tex;
484 flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
487 // Store the flow texture for possible reuse next frame.
488 last_flow_tex = flow_tex;
489 last_frame1 = frame1;
490 last_frame2 = frame2;
493 if (secondary_frame.pts != -1) {
494 // Fade. First kick off the interpolation.
495 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
498 // Now decode the image we are fading against.
500 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
502 // Then fade against it, putting it into the fade Y' and CbCr textures.
503 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, global_flags.width, global_flags.height, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, global_flags.width, global_flags.height);
505 // Subsample and split Cb/Cr.
506 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
508 interpolate_no_split->release_texture(qf.output_tex);
510 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, flow_tex, global_flags.width, global_flags.height, alpha);
513 // Subsample and split Cb/Cr.
514 chroma_subsampler->subsample_chroma(qf.cbcr_tex, global_flags.width, global_flags.height, resources->cb_tex, resources->cr_tex);
517 // We could have released qf.flow_tex here, but to make sure we don't cause a stall
518 // when trying to reuse it for the next frame, we can just as well hold on to it
519 // and release it only when the readback is done.
521 // TODO: This is maybe less relevant now that qf.flow_tex contains the texture we used
522 // _last_ frame, not this one.
524 // Read it down (asynchronously) to the CPU.
525 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
526 glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
528 if (secondary_frame.pts != -1) {
529 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
531 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 4, BUFFER_OFFSET(0));
534 glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3, BUFFER_OFFSET(global_flags.width * global_flags.height));
536 glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, global_flags.width * global_flags.height * 3 - (global_flags.width / 2) * global_flags.height, BUFFER_OFFSET(global_flags.width * global_flags.height + (global_flags.width / 2) * global_flags.height));
538 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
540 // Set a fence we can wait for to make sure the CPU sees the read.
541 glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
543 qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
545 qf.resources = move(resources);
547 lock_guard<mutex> lock(queue_lock);
548 frame_queue.push_back(move(qf));
549 queue_changed.notify_all();
552 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
553 int64_t output_pts, function<void()> &&display_func,
554 QueueSpotHolder &&queue_spot_holder, const string &subtitle)
557 qf.type = QueuedFrame::REFRESH;
558 qf.output_pts = output_pts;
559 qf.display_func = move(display_func);
560 qf.queue_spot_holder = move(queue_spot_holder);
561 qf.subtitle = subtitle;
563 lock_guard<mutex> lock(queue_lock);
564 frame_queue.push_back(move(qf));
565 queue_changed.notify_all();
570 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
572 size_t chroma_width = width / 2;
574 const uint8_t *y = (const uint8_t *)contents;
575 const uint8_t *cb = (const uint8_t *)contents + width * height;
576 const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
578 shared_ptr<Frame> frame(new Frame);
579 frame->y.reset(new uint8_t[width * height]);
580 frame->cb.reset(new uint8_t[chroma_width * height]);
581 frame->cr.reset(new uint8_t[chroma_width * height]);
582 for (unsigned yy = 0; yy < height; ++yy) {
583 memcpy(frame->y.get() + width * yy, y + width * yy, width);
584 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
585 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
587 frame->is_semiplanar = false;
588 frame->width = width;
589 frame->height = height;
590 frame->chroma_subsampling_x = 2;
591 frame->chroma_subsampling_y = 1;
592 frame->pitch_y = width;
593 frame->pitch_chroma = chroma_width;
599 void VideoStream::encode_thread_func()
601 pthread_setname_np(pthread_self(), "VideoStream");
602 QSurface *surface = create_surface();
603 QOpenGLContext *context = create_context(surface);
604 bool ok = make_current(context, surface);
606 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
610 while (!should_quit) {
613 unique_lock<mutex> lock(queue_lock);
615 // Wait until we have a frame to play.
616 queue_changed.wait(lock, [this] {
617 return !frame_queue.empty() || should_quit;
622 steady_clock::time_point frame_start = frame_queue.front().local_pts;
624 // Now sleep until the frame is supposed to start (the usual case),
625 // _or_ clear_queue() happened.
627 if (output_fast_forward) {
628 aborted = frame_queue.empty() || frame_queue.front().local_pts != frame_start;
630 aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start] {
631 return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
635 // clear_queue() happened, so don't play this frame after all.
638 qf = move(frame_queue.front());
639 frame_queue.pop_front();
642 // Hack: We mux the subtitle packet one time unit before the actual frame,
643 // so that Nageru is sure to get it first.
644 if (!qf.subtitle.empty()) {
646 av_init_packet(&pkt);
647 pkt.stream_index = mux->get_subtitle_stream_idx();
648 assert(pkt.stream_index != -1);
649 pkt.data = (uint8_t *)qf.subtitle.data();
650 pkt.size = qf.subtitle.size();
652 pkt.duration = lrint(TIMEBASE / global_flags.output_framerate); // Doesn't really matter for Nageru.
653 mux->add_packet(pkt, qf.output_pts - 1, qf.output_pts - 1);
656 if (qf.type == QueuedFrame::ORIGINAL) {
657 // Send the JPEG frame on, unchanged.
658 string jpeg = frame_reader.read_frame(qf.frame1);
660 av_init_packet(&pkt);
661 pkt.stream_index = 0;
662 pkt.data = (uint8_t *)jpeg.data();
663 pkt.size = jpeg.size();
664 pkt.flags = AV_PKT_FLAG_KEY;
665 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
667 last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
668 } else if (qf.type == QueuedFrame::FADED) {
669 glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
671 shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height);
673 // Now JPEG encode it, and send it on to the stream.
674 vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
677 av_init_packet(&pkt);
678 pkt.stream_index = 0;
679 pkt.data = (uint8_t *)jpeg.data();
680 pkt.size = jpeg.size();
681 pkt.flags = AV_PKT_FLAG_KEY;
682 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
683 last_frame = move(jpeg);
684 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
685 glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
687 // Send it on to display.
688 shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, global_flags.width, global_flags.height);
689 if (qf.display_decoded_func != nullptr) {
690 qf.display_decoded_func(frame);
693 // Now JPEG encode it, and send it on to the stream.
694 vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), global_flags.width, global_flags.height);
695 if (qf.flow_tex != 0) {
696 compute_flow->release_texture(qf.flow_tex);
698 if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
699 interpolate->release_texture(qf.output_tex);
700 interpolate->release_texture(qf.cbcr_tex);
704 av_init_packet(&pkt);
705 pkt.stream_index = 0;
706 pkt.data = (uint8_t *)jpeg.data();
707 pkt.size = jpeg.size();
708 pkt.flags = AV_PKT_FLAG_KEY;
709 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
710 last_frame = move(jpeg);
711 } else if (qf.type == QueuedFrame::REFRESH) {
713 av_init_packet(&pkt);
714 pkt.stream_index = 0;
715 pkt.data = (uint8_t *)last_frame.data();
716 pkt.size = last_frame.size();
717 pkt.flags = AV_PKT_FLAG_KEY;
718 mux->add_packet(pkt, qf.output_pts, qf.output_pts);
722 if (qf.display_func != nullptr) {
728 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
730 VideoStream *video_stream = (VideoStream *)opaque;
731 return video_stream->write_packet2(buf, buf_size, type, time);
734 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
736 if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
737 seen_sync_markers = true;
738 } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
739 // We don't know if this is a keyframe or not (the muxer could
740 // avoid marking it), so we just have to make the best of it.
741 type = AVIO_DATA_MARKER_SYNC_POINT;
744 if (type == AVIO_DATA_MARKER_HEADER) {
745 stream_mux_header.append((char *)buf, buf_size);
746 global_httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header);
748 global_httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });