]> git.sesse.net Git - nageru/blob - futatabi/video_stream.cpp
Fix a 1280x720 hardcoding, replacing it by another one.
[nageru] / futatabi / video_stream.cpp
1 #include "video_stream.h"
2
3 extern "C" {
4 #include <libavformat/avformat.h>
5 #include <libavformat/avio.h>
6 }
7
8 #include "chroma_subsampler.h"
9 #include "shared/context.h"
10 #include "flags.h"
11 #include "flow.h"
12 #include "shared/httpd.h"
13 #include "jpeg_frame_view.h"
14 #include "movit/util.h"
15 #include "shared/mux.h"
16 #include "player.h"
17 #include "util.h"
18 #include "ycbcr_converter.h"
19
20 #include <epoxy/glx.h>
21 #include <jpeglib.h>
22 #include <unistd.h>
23
24 using namespace std;
25 using namespace std::chrono;
26
27 extern HTTPD *global_httpd;
28
29 struct VectorDestinationManager {
30         jpeg_destination_mgr pub;
31         std::vector<uint8_t> dest;
32
33         VectorDestinationManager()
34         {
35                 pub.init_destination = init_destination_thunk;
36                 pub.empty_output_buffer = empty_output_buffer_thunk;
37                 pub.term_destination = term_destination_thunk;
38         }
39
40         static void init_destination_thunk(j_compress_ptr ptr)
41         {
42                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
43         }
44
45         inline void init_destination()
46         {
47                 make_room(0);
48         }
49
50         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
51         {
52                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
53         }
54
55         inline bool empty_output_buffer()
56         {
57                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
58                 return true;
59         }
60
61         inline void make_room(size_t bytes_used)
62         {
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;
67         }
68
69         static void term_destination_thunk(j_compress_ptr ptr)
70         {
71                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
72         }
73
74         inline void term_destination()
75         {
76                 dest.resize(dest.size() - pub.free_in_buffer);
77         }
78 };
79 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
80
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)
82 {
83         VectorDestinationManager dest;
84
85         jpeg_compress_struct cinfo;
86         jpeg_error_mgr jerr;
87         cinfo.err = jpeg_std_error(&jerr);
88         jpeg_create_compress(&cinfo);
89
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);
96
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);
109
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"));
113
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]);
121                 }
122
123                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
124         }
125
126         jpeg_finish_compress(&cinfo);
127         jpeg_destroy_compress(&cinfo);
128
129         return move(dest.dest);
130 }
131
132 VideoStream::VideoStream()
133 {
134         ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr));
135         ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr));
136
137         GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots];
138         GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots];
139         GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots];
140
141         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex);
142         glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex);
143         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex);
144         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex);
145         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex);
146         glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex);
147         check_error();
148
149         constexpr size_t width = 1280, height = 720;  // FIXME: adjustable width, height
150         int levels = find_num_levels(width, height);
151         for (size_t i = 0; i < num_interpolate_slots; ++i) {
152                 glTextureStorage3D(input_tex[i], levels, GL_RGBA8, width, height, 2);
153                 check_error();
154                 glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
155                 check_error();
156                 glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height);
157                 check_error();
158                 glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height);
159                 check_error();
160                 glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height);
161                 check_error();
162                 glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height);
163                 check_error();
164
165                 unique_ptr<InterpolatedFrameResources> resource(new InterpolatedFrameResources);
166                 resource->owner = this;
167                 resource->input_tex = input_tex[i];
168                 resource->gray_tex = gray_tex[i];
169                 resource->fade_y_output_tex = fade_y_output_tex[i];
170                 resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i];
171                 resource->cb_tex = cb_tex[i];
172                 resource->cr_tex = cr_tex[i];
173                 glCreateFramebuffers(2, resource->input_fbos);
174                 check_error();
175                 glCreateFramebuffers(1, &resource->fade_fbo);
176                 check_error();
177
178                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0);
179                 check_error();
180                 glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0);
181                 check_error();
182                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1);
183                 check_error();
184                 glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1);
185                 check_error();
186                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0);
187                 check_error();
188                 glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0);
189                 check_error();
190
191                 GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
192                 glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs);
193                 check_error();
194                 glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs);
195                 check_error();
196                 glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs);
197                 check_error();
198
199                 glCreateBuffers(1, &resource->pbo);
200                 check_error();
201                 glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
202                 check_error();
203                 resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
204                 interpolate_resources.push_back(move(resource));
205         }
206
207         check_error();
208
209         OperatingPoint op;
210         if (global_flags.interpolation_quality == 1) {
211                 op = operating_point1;
212         } else if (global_flags.interpolation_quality == 2) {
213                 op = operating_point2;
214         } else if (global_flags.interpolation_quality == 3) {
215                 op = operating_point3;
216         } else if (global_flags.interpolation_quality == 4) {
217                 op = operating_point4;
218         } else {
219                 assert(false);
220         }
221
222         compute_flow.reset(new DISComputeFlow(width, height, op));
223         interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true));
224         interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false));
225         chroma_subsampler.reset(new ChromaSubsampler);
226         check_error();
227
228         // The “last frame” is initially black.
229         unique_ptr<uint8_t[]> y(new uint8_t[1280 * 720]);
230         unique_ptr<uint8_t[]> cb_or_cr(new uint8_t[640 * 720]);
231         memset(y.get(), 16, 1280 * 720);
232         memset(cb_or_cr.get(), 128, 640 * 720);
233         last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720);
234 }
235
236 VideoStream::~VideoStream() {}
237
238 void VideoStream::start()
239 {
240         AVFormatContext *avctx = avformat_alloc_context();
241
242         // We use Matroska, because it's pretty much the only mux where FFmpeg
243         // allows writing chroma location to override JFIF's default center placement.
244         // (Note that at the time of writing, however, FFmpeg does not correctly
245         // _read_ this information!)
246         avctx->oformat = av_guess_format("matroska", nullptr, nullptr);
247
248         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
249         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
250         avctx->pb->write_data_type = &VideoStream::write_packet2_thunk;
251         avctx->pb->ignore_boundary_point = 1;
252
253         Mux::Codec video_codec = Mux::CODEC_MJPEG;
254
255         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
256
257         string video_extradata;
258
259         constexpr int width = 1280, height = 720;  // Doesn't matter for MJPEG.
260         stream_mux.reset(new Mux(avctx, width, height, video_codec, video_extradata, /*audio_codec_parameters=*/nullptr,
261                 AVCOL_SPC_BT709, Mux::WITHOUT_AUDIO,
262                 COARSE_TIMEBASE, /*write_callback=*/nullptr, Mux::WRITE_FOREGROUND, {}));
263
264
265         encode_thread = thread(&VideoStream::encode_thread_func, this);
266 }
267
268 void VideoStream::stop()
269 {
270         encode_thread.join();
271 }
272
273 void VideoStream::clear_queue()
274 {
275         deque<QueuedFrame> q;
276
277         {
278                 unique_lock<mutex> lock(queue_lock);
279                 q = move(frame_queue);
280         }
281
282         // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves.
283         // Note that release_texture() is thread-safe.
284         for (const QueuedFrame &qf : q) {
285                 if (qf.type == QueuedFrame::INTERPOLATED ||
286                     qf.type == QueuedFrame::FADED_INTERPOLATED) {
287                         compute_flow->release_texture(qf.flow_tex);
288                 }
289                 if (qf.type == QueuedFrame::INTERPOLATED) {
290                         interpolate->release_texture(qf.output_tex);
291                         interpolate->release_texture(qf.cbcr_tex);
292                 }
293         }
294
295         // Destroy q outside the mutex, as that would be a double-lock.
296 }
297
298 void VideoStream::schedule_original_frame(steady_clock::time_point local_pts,
299                                           int64_t output_pts, function<void()> &&display_func,
300                                           QueueSpotHolder &&queue_spot_holder,
301                                           FrameOnDisk frame)
302 {
303         fprintf(stderr, "output_pts=%ld  original      input_pts=%ld\n", output_pts, frame.pts);
304
305         // Preload the file from disk, so that the encoder thread does not get stalled.
306         // TODO: Consider sending it through the queue instead.
307         (void)frame_reader.read_frame(frame);
308
309         QueuedFrame qf;
310         qf.local_pts = local_pts;
311         qf.type = QueuedFrame::ORIGINAL;
312         qf.output_pts = output_pts;
313         qf.frame1 = frame;
314         qf.display_func = move(display_func);
315         qf.queue_spot_holder = move(queue_spot_holder);
316
317         unique_lock<mutex> lock(queue_lock);
318         frame_queue.push_back(move(qf));
319         queue_changed.notify_all();
320 }
321
322 void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts,
323                                        function<void()> &&display_func,
324                                        QueueSpotHolder &&queue_spot_holder,
325                                        FrameOnDisk frame1_spec, FrameOnDisk frame2_spec,
326                                        float fade_alpha)
327 {
328         fprintf(stderr, "output_pts=%ld  faded         input_pts=%ld,%ld  fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha);
329
330         // Get the temporary OpenGL resources we need for doing the fade.
331         // (We share these with interpolated frames, which is slightly
332         // overkill, but there's no need to waste resources on keeping
333         // separate pools around.)
334         BorrowedInterpolatedFrameResources resources;
335         {
336                 unique_lock<mutex> lock(queue_lock);
337                 if (interpolate_resources.empty()) {
338                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
339                         return;
340                 }
341                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
342                 interpolate_resources.pop_front();
343         }
344
345         bool did_decode;
346
347         shared_ptr<Frame> frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
348         shared_ptr<Frame> frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
349
350         ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
351
352         QueuedFrame qf;
353         qf.local_pts = local_pts;
354         qf.type = QueuedFrame::FADED;
355         qf.output_pts = output_pts;
356         qf.frame1 = frame1_spec;
357         qf.display_func = move(display_func);
358         qf.queue_spot_holder = move(queue_spot_holder);
359
360         qf.secondary_frame = frame2_spec;
361
362         // Subsample and split Cb/Cr.
363         chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
364
365         // Read it down (asynchronously) to the CPU.
366         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
367         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
368         check_error();
369         glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
370         check_error();
371         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
372         check_error();
373         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
374         check_error();
375         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
376
377         // Set a fence we can wait for to make sure the CPU sees the read.
378         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
379         check_error();
380         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
381         check_error();
382         qf.resources = move(resources);
383         qf.local_pts = local_pts;
384
385         unique_lock<mutex> lock(queue_lock);
386         frame_queue.push_back(move(qf));
387         queue_changed.notify_all();
388 }
389
390 void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts,
391                                               int64_t output_pts, function<void(shared_ptr<Frame>)> &&display_func,
392                                               QueueSpotHolder &&queue_spot_holder,
393                                               FrameOnDisk frame1, FrameOnDisk frame2,
394                                               float alpha, FrameOnDisk secondary_frame, float fade_alpha)
395 {
396         if (secondary_frame.pts != -1) {
397                 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);
398         } else {
399                 fprintf(stderr, "output_pts=%ld  interpolated  input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha);
400         }
401
402         // Get the temporary OpenGL resources we need for doing the interpolation.
403         BorrowedInterpolatedFrameResources resources;
404         {
405                 unique_lock<mutex> lock(queue_lock);
406                 if (interpolate_resources.empty()) {
407                         fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n");
408                         return;
409                 }
410                 resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release());
411                 interpolate_resources.pop_front();
412         }
413
414         QueuedFrame qf;
415         qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED;
416         qf.output_pts = output_pts;
417         qf.display_decoded_func = move(display_func);
418         qf.queue_spot_holder = move(queue_spot_holder);
419         qf.local_pts = local_pts;
420
421         check_error();
422
423         // Convert frame0 and frame1 to OpenGL textures.
424         for (size_t frame_no = 0; frame_no < 2; ++frame_no) {
425                 FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1;
426                 bool did_decode;
427                 shared_ptr<Frame> frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
428                 ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], 1280, 720);
429         }
430
431         glGenerateTextureMipmap(resources->input_tex);
432         check_error();
433         glGenerateTextureMipmap(resources->gray_tex);
434         check_error();
435
436         // Compute the interpolated frame.
437         qf.flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
438         check_error();
439
440         if (secondary_frame.pts != -1) {
441                 // Fade. First kick off the interpolation.
442                 tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
443                 check_error();
444
445                 // Now decode the image we are fading against.
446                 bool did_decode;
447                 shared_ptr<Frame> frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &frame_reader, &did_decode);
448
449                 // Then fade against it, putting it into the fade Y' and CbCr textures.
450                 ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, 1280, 720, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720);
451
452                 // Subsample and split Cb/Cr.
453                 chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
454
455                 interpolate_no_split->release_texture(qf.output_tex);
456         } else {
457                 tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha);
458                 check_error();
459
460                 // Subsample and split Cb/Cr.
461                 chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources->cb_tex, resources->cr_tex);
462         }
463
464         // We could have released qf.flow_tex here, but to make sure we don't cause a stall
465         // when trying to reuse it for the next frame, we can just as well hold on to it
466         // and release it only when the readback is done.
467
468         // Read it down (asynchronously) to the CPU.
469         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
470         glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo);
471         check_error();
472         if (secondary_frame.pts != -1) {
473                 glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
474         } else {
475                 glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
476         }
477         check_error();
478         glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
479         check_error();
480         glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720));
481         check_error();
482         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
483
484         // Set a fence we can wait for to make sure the CPU sees the read.
485         glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
486         check_error();
487         qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
488         check_error();
489         qf.resources = move(resources);
490
491         unique_lock<mutex> lock(queue_lock);
492         frame_queue.push_back(move(qf));
493         queue_changed.notify_all();
494 }
495
496 void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts,
497                                          int64_t output_pts, function<void()> &&display_func,
498                                          QueueSpotHolder &&queue_spot_holder)
499 {
500         QueuedFrame qf;
501         qf.type = QueuedFrame::REFRESH;
502         qf.output_pts = output_pts;
503         qf.display_func = move(display_func);
504         qf.queue_spot_holder = move(queue_spot_holder);
505
506         unique_lock<mutex> lock(queue_lock);
507         frame_queue.push_back(move(qf));
508         queue_changed.notify_all();
509 }
510
511 namespace {
512
513 shared_ptr<Frame> frame_from_pbo(void *contents, size_t width, size_t height)
514 {
515         size_t chroma_width = width / 2;
516
517         const uint8_t *y = (const uint8_t *)contents;
518         const uint8_t *cb = (const uint8_t *)contents + width * height;
519         const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height;
520
521         shared_ptr<Frame> frame(new Frame);
522         frame->y.reset(new uint8_t[width * height]);
523         frame->cb.reset(new uint8_t[chroma_width * height]);
524         frame->cr.reset(new uint8_t[chroma_width * height]);
525         for (unsigned yy = 0; yy < height; ++yy) {
526                 memcpy(frame->y.get() + width * yy, y + width * yy, width);
527                 memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width);
528                 memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width);
529         }
530         frame->is_semiplanar = false;
531         frame->width = width;
532         frame->height = height;
533         frame->chroma_subsampling_x = 2;
534         frame->chroma_subsampling_y = 1;
535         frame->pitch_y = width;
536         frame->pitch_chroma = chroma_width;
537         return frame;
538 }
539
540 }  // namespace
541
542 void VideoStream::encode_thread_func()
543 {
544         pthread_setname_np(pthread_self(), "VideoStream");
545         QSurface *surface = create_surface();
546         QOpenGLContext *context = create_context(surface);
547         bool ok = make_current(context, surface);
548         if (!ok) {
549                 fprintf(stderr, "Video stream couldn't get an OpenGL context\n");
550                 exit(1);
551         }
552
553         for ( ;; ) {
554                 QueuedFrame qf;
555                 {
556                         unique_lock<mutex> lock(queue_lock);
557
558                         // Wait until we have a frame to play.
559                         queue_changed.wait(lock, [this]{
560                                 return !frame_queue.empty();
561                         });
562                         steady_clock::time_point frame_start = frame_queue.front().local_pts;
563
564                         // Now sleep until the frame is supposed to start (the usual case),
565                         // _or_ clear_queue() happened.
566                         bool aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{
567                                 return frame_queue.empty() || frame_queue.front().local_pts != frame_start;
568                         });
569                         if (aborted) {
570                                 // clear_queue() happened, so don't play this frame after all.
571                                 continue;
572                         }
573                         qf = move(frame_queue.front());
574                         frame_queue.pop_front();
575                 }
576
577                 if (qf.type == QueuedFrame::ORIGINAL) {
578                         // Send the JPEG frame on, unchanged.
579                         string jpeg = frame_reader.read_frame(qf.frame1);
580                         AVPacket pkt;
581                         av_init_packet(&pkt);
582                         pkt.stream_index = 0;
583                         pkt.data = (uint8_t *)jpeg.data();
584                         pkt.size = jpeg.size();
585                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
586
587                         last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size());
588                 } else if (qf.type == QueuedFrame::FADED) {
589                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
590
591                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
592
593                         // Now JPEG encode it, and send it on to the stream.
594                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
595
596                         AVPacket pkt;
597                         av_init_packet(&pkt);
598                         pkt.stream_index = 0;
599                         pkt.data = (uint8_t *)jpeg.data();
600                         pkt.size = jpeg.size();
601                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
602                         last_frame = move(jpeg);
603                 } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) {
604                         glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
605
606                         // Send it on to display.
607                         shared_ptr<Frame> frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720);
608                         if (qf.display_decoded_func != nullptr) {
609                                 qf.display_decoded_func(frame);
610                         }
611
612                         // Now JPEG encode it, and send it on to the stream.
613                         vector<uint8_t> jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720);
614                         compute_flow->release_texture(qf.flow_tex);
615                         if (qf.type != QueuedFrame::FADED_INTERPOLATED) {
616                                 interpolate->release_texture(qf.output_tex);
617                                 interpolate->release_texture(qf.cbcr_tex);
618                         }
619
620                         AVPacket pkt;
621                         av_init_packet(&pkt);
622                         pkt.stream_index = 0;
623                         pkt.data = (uint8_t *)jpeg.data();
624                         pkt.size = jpeg.size();
625                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
626                         last_frame = move(jpeg);
627                 } else if (qf.type == QueuedFrame::REFRESH) {
628                         AVPacket pkt;
629                         av_init_packet(&pkt);
630                         pkt.stream_index = 0;
631                         pkt.data = (uint8_t *)last_frame.data();
632                         pkt.size = last_frame.size();
633                         stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts);
634                 } else {
635                         assert(false);
636                 }
637                 if (qf.display_func != nullptr) {
638                         qf.display_func();
639                 }
640         }
641 }
642
643 int VideoStream::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
644 {
645         VideoStream *video_stream = (VideoStream *)opaque;
646         return video_stream->write_packet2(buf, buf_size, type, time);
647 }
648
649 int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
650 {
651         if (type == AVIO_DATA_MARKER_SYNC_POINT || type == AVIO_DATA_MARKER_BOUNDARY_POINT) {
652                 seen_sync_markers = true;
653         } else if (type == AVIO_DATA_MARKER_UNKNOWN && !seen_sync_markers) {
654                 // We don't know if this is a keyframe or not (the muxer could
655                 // avoid marking it), so we just have to make the best of it.
656                 type = AVIO_DATA_MARKER_SYNC_POINT;
657         }
658
659         if (type == AVIO_DATA_MARKER_HEADER) {
660                 stream_mux_header.append((char *)buf, buf_size);
661                 global_httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header);
662         } else {
663                 global_httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
664         }
665         return buf_size;
666 }