]> git.sesse.net Git - nageru/blob - nageru/decklink_output.cpp
e37002918cb4891f645827dcef74dc7115429031
[nageru] / nageru / decklink_output.cpp
1 #include <movit/effect_util.h>
2 #include <movit/util.h>
3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
4 #include <pthread.h>
5 #include <unistd.h>
6
7 #include <mutex>
8
9 #include <epoxy/egl.h>
10
11 #include "chroma_subsampler.h"
12 #include "decklink_output.h"
13 #include "decklink_util.h"
14 #include "flags.h"
15 #include "shared/metrics.h"
16 #include "print_latency.h"
17 #include "shared/timebase.h"
18 #include "v210_converter.h"
19
20 using namespace movit;
21 using namespace std;
22 using namespace std::chrono;
23
24 namespace {
25
26 // This class can be deleted during regular use, so make all the metrics static.
27 once_flag decklink_metrics_inited;
28 LatencyHistogram latency_histogram;
29 atomic<int64_t> metric_decklink_output_width_pixels{-1};
30 atomic<int64_t> metric_decklink_output_height_pixels{-1};
31 atomic<int64_t> metric_decklink_output_frame_rate_den{-1};
32 atomic<int64_t> metric_decklink_output_frame_rate_nom{-1};
33 atomic<int64_t> metric_decklink_output_inflight_frames{0};
34 atomic<int64_t> metric_decklink_output_color_mismatch_frames{0};
35
36 atomic<int64_t> metric_decklink_output_scheduled_frames_dropped{0};
37 atomic<int64_t> metric_decklink_output_scheduled_frames_late{0};
38 atomic<int64_t> metric_decklink_output_scheduled_frames_normal{0};
39 atomic<int64_t> metric_decklink_output_scheduled_frames_preroll{0};
40
41 atomic<int64_t> metric_decklink_output_completed_frames_completed{0};
42 atomic<int64_t> metric_decklink_output_completed_frames_dropped{0};
43 atomic<int64_t> metric_decklink_output_completed_frames_flushed{0};
44 atomic<int64_t> metric_decklink_output_completed_frames_late{0};
45 atomic<int64_t> metric_decklink_output_completed_frames_unknown{0};
46
47 atomic<int64_t> metric_decklink_output_scheduled_samples{0};
48
49 Summary metric_decklink_output_margin_seconds;
50
51 }  // namespace
52
53 DeckLinkOutput::DeckLinkOutput(ResourcePool *resource_pool, QSurface *surface, unsigned width, unsigned height, unsigned card_index)
54         : resource_pool(resource_pool), surface(surface), width(width), height(height), card_index(card_index)
55 {
56         chroma_subsampler.reset(new ChromaSubsampler(resource_pool));
57
58         call_once(decklink_metrics_inited, [](){
59                 latency_histogram.init("decklink_output");
60                 global_metrics.add("decklink_output_width_pixels", &metric_decklink_output_width_pixels, Metrics::TYPE_GAUGE);
61                 global_metrics.add("decklink_output_height_pixels", &metric_decklink_output_height_pixels, Metrics::TYPE_GAUGE);
62                 global_metrics.add("decklink_output_frame_rate_den", &metric_decklink_output_frame_rate_den, Metrics::TYPE_GAUGE);
63                 global_metrics.add("decklink_output_frame_rate_nom", &metric_decklink_output_frame_rate_nom, Metrics::TYPE_GAUGE);
64                 global_metrics.add("decklink_output_inflight_frames", &metric_decklink_output_inflight_frames, Metrics::TYPE_GAUGE);
65                 global_metrics.add("decklink_output_color_mismatch_frames", &metric_decklink_output_color_mismatch_frames);
66
67                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "dropped" }}, &metric_decklink_output_scheduled_frames_dropped);
68                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "late" }}, &metric_decklink_output_scheduled_frames_late);
69                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "normal" }}, &metric_decklink_output_scheduled_frames_normal);
70                 global_metrics.add("decklink_output_scheduled_frames", {{ "status", "preroll" }}, &metric_decklink_output_scheduled_frames_preroll);
71
72                 global_metrics.add("decklink_output_completed_frames", {{ "status", "completed" }}, &metric_decklink_output_completed_frames_completed);
73                 global_metrics.add("decklink_output_completed_frames", {{ "status", "dropped" }}, &metric_decklink_output_completed_frames_dropped);
74                 global_metrics.add("decklink_output_completed_frames", {{ "status", "flushed" }}, &metric_decklink_output_completed_frames_flushed);
75                 global_metrics.add("decklink_output_completed_frames", {{ "status", "late" }}, &metric_decklink_output_completed_frames_late);
76                 global_metrics.add("decklink_output_completed_frames", {{ "status", "unknown" }}, &metric_decklink_output_completed_frames_unknown);
77
78                 global_metrics.add("decklink_output_scheduled_samples", &metric_decklink_output_scheduled_samples);
79                 vector<double> quantiles{0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99};
80                 metric_decklink_output_margin_seconds.init(quantiles, 60.0);
81                 global_metrics.add("decklink_output_margin_seconds", &metric_decklink_output_margin_seconds);
82         });
83 }
84
85 bool DeckLinkOutput::set_device(IDeckLink *decklink)
86 {
87         if (decklink->QueryInterface(IID_IDeckLinkOutput, (void**)&output) != S_OK) {
88                 fprintf(stderr, "Warning: Card %u has no outputs\n", card_index);
89                 return false;
90         }
91
92         IDeckLinkDisplayModeIterator *mode_it;
93         if (output->GetDisplayModeIterator(&mode_it) != S_OK) {
94                 fprintf(stderr, "Warning: Failed to enumerate output display modes for card %u\n", card_index);
95                 return false;
96         }
97
98         video_modes.clear();
99
100         for (const auto &it : summarize_video_modes(mode_it, card_index)) {
101                 if (it.second.width != width || it.second.height != height) {
102                         continue;
103                 }
104
105                 // We could support interlaced modes, but let's stay out of it for now,
106                 // since we don't have interlaced stream output.
107                 if (it.second.interlaced) {
108                         continue;
109                 }
110
111                 video_modes.insert(it);
112         }
113
114         mode_it->Release();
115
116         // HDMI or SDI generally mean “both HDMI and SDI at the same time” on DeckLink cards
117         // that support both; pick_default_video_connection() will generally pick one of those
118         // if they exist. (--prefer-hdmi-input would also affect the selection despite the name
119         // of the option, but since either generally means both, it's inconsequential.)
120         // We're not very likely to need analog outputs, so we don't need a way to change
121         // beyond that.
122         video_connection = pick_default_video_connection(decklink, BMDDeckLinkVideoOutputConnections, card_index);
123         return true;
124 }
125
126 void DeckLinkOutput::start_output(uint32_t mode, int64_t base_pts)
127 {
128         assert(output);
129         assert(!playback_initiated);
130
131         if (video_modes.empty()) {
132                 fprintf(stderr, "ERROR: No matching output modes for %dx%d found\n", width, height);
133                 exit(1);
134         }
135
136         should_quit.unquit();
137         playback_initiated = true;
138         playback_started = false;
139         this->base_pts = base_pts;
140
141         IDeckLinkConfiguration *config = nullptr;
142         if (output->QueryInterface(IID_IDeckLinkConfiguration, (void**)&config) != S_OK) {
143                 fprintf(stderr, "Failed to get configuration interface for output card\n");
144                 exit(1);
145         }
146         if (config->SetFlag(bmdDeckLinkConfigLowLatencyVideoOutput, true) != S_OK) {
147                 fprintf(stderr, "Failed to set low latency output\n");
148                 exit(1);
149         }
150         if (config->SetInt(bmdDeckLinkConfigVideoOutputConnection, video_connection) != S_OK) {
151                 fprintf(stderr, "Failed to set video output connection for card %u\n", card_index);
152                 exit(1);
153         }
154         if (config->SetFlag(bmdDeckLinkConfigUse1080pNotPsF, true) != S_OK) {
155                 fprintf(stderr, "Failed to set PsF flag for card\n");
156                 exit(1);
157         }
158         if (config->SetFlag(bmdDeckLinkConfigSMPTELevelAOutput, true) != S_OK) {
159                 // This affects at least some no-name SDI->HDMI converters.
160                 // Warn, but don't die.
161                 fprintf(stderr, "WARNING: Failed to enable SMTPE Level A; resolutions like 1080p60 might have issues.\n");
162         }
163
164         BMDDisplayModeSupport support;
165         IDeckLinkDisplayMode *display_mode;
166         BMDPixelFormat pixel_format = global_flags.ten_bit_output ? bmdFormat10BitYUV : bmdFormat8BitYUV;
167         if (output->DoesSupportVideoMode(mode, pixel_format, bmdVideoOutputFlagDefault,
168                                          &support, &display_mode) != S_OK) {
169                 fprintf(stderr, "Couldn't ask for format support\n");
170                 exit(1);
171         }
172
173         if (support == bmdDisplayModeNotSupported) {
174                 fprintf(stderr, "Requested display mode not supported\n");
175                 exit(1);
176         }
177
178         current_mode_flags = display_mode->GetFlags();
179
180         BMDTimeValue time_value;
181         BMDTimeScale time_scale;
182         if (display_mode->GetFrameRate(&time_value, &time_scale) != S_OK) {
183                 fprintf(stderr, "Couldn't get frame rate\n");
184                 exit(1);
185         }
186
187         metric_decklink_output_width_pixels = width;
188         metric_decklink_output_height_pixels = height;
189         metric_decklink_output_frame_rate_nom = time_value;
190         metric_decklink_output_frame_rate_den = time_scale;
191
192         frame_duration = time_value * TIMEBASE / time_scale;
193
194         display_mode->Release();
195
196         HRESULT result = output->EnableVideoOutput(mode, bmdVideoOutputFlagDefault);
197         if (result != S_OK) {
198                 fprintf(stderr, "Couldn't enable output with error 0x%x\n", result);
199                 exit(1);
200         }
201         if (output->SetScheduledFrameCompletionCallback(this) != S_OK) {
202                 fprintf(stderr, "Couldn't set callback\n");
203                 exit(1);
204         }
205         assert(OUTPUT_FREQUENCY == 48000);
206         if (output->EnableAudioOutput(bmdAudioSampleRate48kHz, bmdAudioSampleType32bitInteger, 2, bmdAudioOutputStreamTimestamped) != S_OK) {
207                 fprintf(stderr, "Couldn't enable audio output\n");
208                 exit(1);
209         }
210         if (output->BeginAudioPreroll() != S_OK) {
211                 fprintf(stderr, "Couldn't begin audio preroll\n");
212                 exit(1);
213         }
214
215         present_thread = thread([this]{
216                 QOpenGLContext *context = create_context(this->surface);
217                 eglBindAPI(EGL_OPENGL_API);
218                 if (!make_current(context, this->surface)) {
219                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
220                                 eglGetError());
221                         exit(1);
222                 }
223                 present_thread_func();
224                 delete_context(context);
225         });
226 }
227
228 void DeckLinkOutput::end_output()
229 {
230         if (!playback_initiated) {
231                 return;
232         }
233
234         should_quit.quit();
235         frame_queues_changed.notify_all();
236         present_thread.join();
237         playback_initiated = false;
238
239         output->StopScheduledPlayback(0, nullptr, 0);
240         output->DisableVideoOutput();
241         output->DisableAudioOutput();
242
243         // Wait until all frames are accounted for, and free them.
244         {
245                 unique_lock<mutex> lock(frame_queue_mutex);
246                 while (!(frame_freelist.empty() && num_frames_in_flight == 0)) {
247                         frame_queues_changed.wait(lock, [this]{ return !frame_freelist.empty(); });
248                         frame_freelist.pop();
249                 }
250         }
251 }
252
253 void DeckLinkOutput::send_frame(GLuint y_tex, GLuint cbcr_tex, YCbCrLumaCoefficients output_ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, int64_t pts, int64_t duration)
254 {
255         assert(!should_quit.should_quit());
256
257         if ((current_mode_flags & bmdDisplayModeColorspaceRec601) && output_ycbcr_coefficients == YCBCR_REC_709) {
258                 if (!last_frame_had_mode_mismatch) {
259                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 601 Y'CbCr coefficients.\n");
260                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec601 (or =auto).\n");
261                 }
262                 last_frame_had_mode_mismatch = true;
263                 ++metric_decklink_output_color_mismatch_frames;
264         } else if ((current_mode_flags & bmdDisplayModeColorspaceRec709) && output_ycbcr_coefficients == YCBCR_REC_601) {
265                 if (!last_frame_had_mode_mismatch) {
266                         fprintf(stderr, "WARNING: Chosen output mode expects Rec. 709 Y'CbCr coefficients.\n");
267                         fprintf(stderr, "         Consider --output-ycbcr-coefficients=rec709 (or =auto).\n");
268                 }
269                 last_frame_had_mode_mismatch = true;
270                 ++metric_decklink_output_color_mismatch_frames;
271         } else {
272                 last_frame_had_mode_mismatch = false;
273         }
274
275         unique_ptr<Frame> frame = get_frame();
276         if (global_flags.ten_bit_output) {
277                 chroma_subsampler->create_v210(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
278         } else {
279                 chroma_subsampler->create_uyvy(y_tex, cbcr_tex, width, height, frame->uyvy_tex);
280         }
281
282         // Download the UYVY texture to the PBO.
283         glPixelStorei(GL_PACK_ROW_LENGTH, 0);
284         check_error();
285
286         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
287         check_error();
288
289         if (global_flags.ten_bit_output) {
290                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
291                 check_error();
292                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, BUFFER_OFFSET(0));
293                 check_error();
294         } else {
295                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
296                 check_error();
297                 glGetTexImage(GL_TEXTURE_2D, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, BUFFER_OFFSET(0));
298                 check_error();
299         }
300
301         glBindTexture(GL_TEXTURE_2D, 0);
302         check_error();
303         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
304         check_error();
305
306         glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
307         check_error();
308
309         frame->fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
310         check_error();
311         glFlush();  // Make the DeckLink thread see the fence as soon as possible.
312         check_error();
313
314         frame->input_frames = input_frames;
315         frame->received_ts = find_received_timestamp(input_frames);
316         frame->pts = pts;
317         frame->duration = duration;
318
319         {
320                 lock_guard<mutex> lock(frame_queue_mutex);
321                 pending_video_frames.push(move(frame));
322         }
323         frame_queues_changed.notify_all();
324 }
325
326 void DeckLinkOutput::send_audio(int64_t pts, const std::vector<float> &samples)
327 {
328         unique_ptr<int32_t[]> int_samples(new int32_t[samples.size()]);
329         for (size_t i = 0; i < samples.size(); ++i) {
330                 int_samples[i] = lrintf(samples[i] * 2147483648.0f);
331         }
332
333         uint32_t frames_written;
334         HRESULT result = output->ScheduleAudioSamples(int_samples.get(), samples.size() / 2,
335                 pts, TIMEBASE, &frames_written);
336         if (result != S_OK) {
337                 fprintf(stderr, "ScheduleAudioSamples(pts=%" PRId64 ") failed (result=0x%08x)\n", pts, result);
338         } else {
339                 if (frames_written != samples.size() / 2) {
340                         fprintf(stderr, "ScheduleAudioSamples() returned short write (%u/%zu)\n", frames_written, samples.size() / 2);
341                 }
342         }
343         metric_decklink_output_scheduled_samples += samples.size() / 2;
344 }
345
346 void DeckLinkOutput::wait_for_frame(int64_t pts, int *dropped_frames, int64_t *frame_duration, bool *is_preroll, steady_clock::time_point *frame_timestamp)
347 {
348         assert(!should_quit.should_quit());
349
350         *dropped_frames = 0;
351         *frame_duration = this->frame_duration;
352
353         const BMDTimeValue buffer = lrint(*frame_duration * global_flags.output_buffer_frames);
354         const BMDTimeValue max_overshoot = lrint(*frame_duration * global_flags.output_slop_frames);
355         BMDTimeValue target_time = pts - buffer;
356
357         // While prerolling, we send out frames as quickly as we can.
358         if (target_time < base_pts) {
359                 *is_preroll = true;
360                 ++metric_decklink_output_scheduled_frames_preroll;
361                 return;
362         }
363
364         *is_preroll = !playback_started;
365
366         if (!playback_started) {
367                 if (output->EndAudioPreroll() != S_OK) {
368                         fprintf(stderr, "Could not end audio preroll\n");
369                         exit(1);  // TODO
370                 }
371                 if (output->StartScheduledPlayback(base_pts, TIMEBASE, 1.0) != S_OK) {
372                         fprintf(stderr, "Could not start playback\n");
373                         exit(1);  // TODO
374                 }
375                 playback_started = true;
376         }
377
378         BMDTimeValue stream_frame_time;
379         double playback_speed;
380         output->GetScheduledStreamTime(TIMEBASE, &stream_frame_time, &playback_speed);
381
382         *frame_timestamp = steady_clock::now() +
383                 nanoseconds((target_time - stream_frame_time) * 1000000000 / TIMEBASE);
384
385         metric_decklink_output_margin_seconds.count_event(
386                 (target_time - stream_frame_time) / double(TIMEBASE));
387
388         // If we're ahead of time, wait for the frame to (approximately) start.
389         if (stream_frame_time < target_time) {
390                 should_quit.sleep_until(*frame_timestamp);
391                 ++metric_decklink_output_scheduled_frames_normal;
392                 return;
393         }
394
395         // If we overshot the previous frame by just a little,
396         // fire off one immediately.
397         if (stream_frame_time < target_time + max_overshoot) {
398                 fprintf(stderr, "Warning: Frame was %ld ms late (but not skipping it due to --output-slop-frames).\n",
399                         lrint((stream_frame_time - target_time) * 1000.0 / TIMEBASE));
400                 ++metric_decklink_output_scheduled_frames_late;
401                 return;
402         }
403
404         // Oops, we missed by more than one frame. Return immediately,
405         // but drop so that we catch up.
406         *dropped_frames = (stream_frame_time - target_time + *frame_duration - 1) / *frame_duration;
407         const int64_t ns_per_frame = this->frame_duration * 1000000000 / TIMEBASE;
408         *frame_timestamp += nanoseconds(*dropped_frames * ns_per_frame);
409         fprintf(stderr, "Dropped %d output frames; skipping.\n", *dropped_frames);
410         metric_decklink_output_scheduled_frames_dropped += *dropped_frames;
411         ++metric_decklink_output_scheduled_frames_normal;
412 }
413
414 uint32_t DeckLinkOutput::pick_video_mode(uint32_t mode) const
415 {
416         if (video_modes.count(mode)) {
417                 return mode;
418         }
419
420         // Prioritize 59.94 > 60 > 29.97. If none of those are found, then pick the highest one.
421         for (const pair<int, int> &desired : vector<pair<int, int>>{ { 60000, 1001 }, { 60, 0 }, { 30000, 1001 } }) {
422                 for (const auto &it : video_modes) {
423                         if (it.second.frame_rate_num * desired.second == desired.first * it.second.frame_rate_den) {
424                                 return it.first;
425                         }
426                 }
427         }
428
429         uint32_t best_mode = 0;
430         double best_fps = 0.0;
431         for (const auto &it : video_modes) {
432                 double fps = double(it.second.frame_rate_num) / it.second.frame_rate_den;
433                 if (fps > best_fps) {
434                         best_mode = it.first;
435                         best_fps = fps;
436                 }
437         }
438         return best_mode;
439 }
440
441 YCbCrLumaCoefficients DeckLinkOutput::preferred_ycbcr_coefficients() const
442 {
443         if (current_mode_flags & bmdDisplayModeColorspaceRec601) {
444                 return YCBCR_REC_601;
445         } else {
446                 // Don't bother checking bmdDisplayModeColorspaceRec709;
447                 // if none is set, 709 is a good default anyway.
448                 return YCBCR_REC_709;
449         }
450 }
451
452 HRESULT DeckLinkOutput::ScheduledFrameCompleted(/* in */ IDeckLinkVideoFrame *completedFrame, /* in */ BMDOutputFrameCompletionResult result)
453 {
454         Frame *frame = static_cast<Frame *>(completedFrame);
455         switch (result) {
456         case bmdOutputFrameCompleted:
457                 ++metric_decklink_output_completed_frames_completed;
458                 break;
459         case bmdOutputFrameDisplayedLate:
460                 fprintf(stderr, "Output frame displayed late (pts=%" PRId64 ")\n", frame->pts);
461                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
462                 ++metric_decklink_output_completed_frames_late;
463                 break;
464         case bmdOutputFrameDropped:
465                 fprintf(stderr, "Output frame was dropped (pts=%" PRId64 "ld)\n", frame->pts);
466                 fprintf(stderr, "Consider increasing --output-buffer-frames if this persists.\n");
467                 ++metric_decklink_output_completed_frames_dropped;
468                 break;
469         case bmdOutputFrameFlushed:
470                 fprintf(stderr, "Output frame was flushed (pts=%" PRId64 "ld)\n", frame->pts);
471                 ++metric_decklink_output_completed_frames_flushed;
472                 break;
473         default:
474                 fprintf(stderr, "Output frame completed with unknown status %d\n", result);
475                 ++metric_decklink_output_completed_frames_unknown;
476                 break;
477         }
478
479         static int frameno = 0;
480         print_latency("DeckLink output latency (frame received → output on HDMI):", frame->received_ts, false, &frameno, &latency_histogram);
481
482         {
483                 lock_guard<mutex> lock(frame_queue_mutex);
484                 frame_freelist.push(unique_ptr<Frame>(frame));
485                 --num_frames_in_flight;
486                 --metric_decklink_output_inflight_frames;
487         }
488
489         return S_OK;
490 }
491
492 HRESULT DeckLinkOutput::ScheduledPlaybackHasStopped()
493 {
494         printf("playback stopped!\n");
495         return S_OK;
496 }
497
498 unique_ptr<DeckLinkOutput::Frame> DeckLinkOutput::get_frame()
499 {
500         lock_guard<mutex> lock(frame_queue_mutex);
501
502         if (!frame_freelist.empty()) {
503                 unique_ptr<Frame> frame = move(frame_freelist.front());
504                 frame_freelist.pop();
505                 return frame;
506         }
507
508         unique_ptr<Frame> frame(new Frame);
509
510         size_t stride;
511         if (global_flags.ten_bit_output) {
512                 stride = v210Converter::get_v210_stride(width);
513                 GLint v210_width = stride / sizeof(uint32_t);
514                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGB10_A2, v210_width, height);
515
516                 // We need valid texture state, or NVIDIA won't allow us to write to the texture.
517                 glBindTexture(GL_TEXTURE_2D, frame->uyvy_tex);
518                 check_error();
519                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
520                 check_error();
521         } else {
522                 stride = width * 2;
523                 frame->uyvy_tex = resource_pool->create_2d_texture(GL_RGBA8, width / 2, height);
524         }
525
526         glGenBuffers(1, &frame->pbo);
527         check_error();
528         glBindBuffer(GL_PIXEL_PACK_BUFFER, frame->pbo);
529         check_error();
530         glBufferStorage(GL_PIXEL_PACK_BUFFER, stride * height, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
531         check_error();
532         frame->uyvy_ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, stride * height, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
533         check_error();
534         frame->uyvy_ptr_local.reset(new uint8_t[stride * height]);
535         frame->resource_pool = resource_pool;
536
537         return frame;
538 }
539
540 void DeckLinkOutput::present_thread_func()
541 {
542         pthread_setname_np(pthread_self(), "DeckLinkOutput");
543         for ( ;; ) {
544                 unique_ptr<Frame> frame;
545                 {
546                         unique_lock<mutex> lock(frame_queue_mutex);
547                         frame_queues_changed.wait(lock, [this]{
548                                 return should_quit.should_quit() || !pending_video_frames.empty();
549                         });
550                         if (should_quit.should_quit()) {
551                                 return;
552                         }
553                         frame = move(pending_video_frames.front());
554                         pending_video_frames.pop();
555                         ++num_frames_in_flight;
556                         ++metric_decklink_output_inflight_frames;
557                 }
558
559                 for ( ;; ) {
560                         int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
561                         if (err == GL_TIMEOUT_EXPIRED) {
562                                 // NVIDIA likes to busy-wait; yield instead.
563                                 this_thread::sleep_for(milliseconds(1));
564                         } else {
565                                 break;
566                         }
567                 }
568                 check_error();
569                 frame->fence.reset();
570
571                 if (global_flags.ten_bit_output) {
572                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, v210Converter::get_v210_stride(width) * height);
573                 } else {
574                         memcpy(frame->uyvy_ptr_local.get(), frame->uyvy_ptr, width * height * 2);
575                 }
576
577                 // Release any input frames we needed to render this frame.
578                 frame->input_frames.clear();
579
580                 BMDTimeValue pts = frame->pts;
581                 BMDTimeValue duration = frame->duration;
582                 HRESULT res = output->ScheduleVideoFrame(frame.get(), pts, duration, TIMEBASE);
583                 if (res == S_OK) {
584                         frame.release();  // Owned by the driver now.
585                 } else {
586                         fprintf(stderr, "Could not schedule video frame! (error=0x%08x)\n", res);
587
588                         lock_guard<mutex> lock(frame_queue_mutex);
589                         frame_freelist.push(move(frame));
590                         --num_frames_in_flight;
591                         --metric_decklink_output_inflight_frames;
592                 }
593         }
594 }
595
596 HRESULT STDMETHODCALLTYPE DeckLinkOutput::QueryInterface(REFIID, LPVOID *)
597 {
598         return E_NOINTERFACE;
599 }
600
601 ULONG STDMETHODCALLTYPE DeckLinkOutput::AddRef()
602 {
603         return refcount.fetch_add(1) + 1;
604 }
605
606 ULONG STDMETHODCALLTYPE DeckLinkOutput::Release()
607 {
608         int new_ref = refcount.fetch_sub(1) - 1;
609         if (new_ref == 0)
610                 delete this;
611         return new_ref;
612 }
613
614 DeckLinkOutput::Frame::~Frame()
615 {
616         glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo);
617         check_error();
618         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
619         check_error();
620         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
621         check_error();
622         glDeleteBuffers(1, &pbo);
623         check_error();
624         resource_pool->release_2d_texture(uyvy_tex);
625         check_error();
626 }
627
628 HRESULT STDMETHODCALLTYPE DeckLinkOutput::Frame::QueryInterface(REFIID, LPVOID *)
629 {
630         return E_NOINTERFACE;
631 }
632
633 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::AddRef()
634 {
635         return refcount.fetch_add(1) + 1;
636 }
637
638 ULONG STDMETHODCALLTYPE DeckLinkOutput::Frame::Release()
639 {
640         int new_ref = refcount.fetch_sub(1) - 1;
641         if (new_ref == 0)
642                 delete this;
643         return new_ref;
644 }
645
646 long DeckLinkOutput::Frame::GetWidth()
647 {
648         return global_flags.width;
649 }
650
651 long DeckLinkOutput::Frame::GetHeight()
652 {
653         return global_flags.height;
654 }
655
656 long DeckLinkOutput::Frame::GetRowBytes()
657 {
658         if (global_flags.ten_bit_output) {
659                 return v210Converter::get_v210_stride(global_flags.width);
660         } else {
661                 return global_flags.width * 2;
662         }
663 }
664
665 BMDPixelFormat DeckLinkOutput::Frame::GetPixelFormat()
666 {
667         if (global_flags.ten_bit_output) {
668                 return bmdFormat10BitYUV;
669         } else {
670                 return bmdFormat8BitYUV;
671         }
672 }
673
674 BMDFrameFlags DeckLinkOutput::Frame::GetFlags()
675 {
676         return bmdFrameFlagDefault;
677 }
678
679 HRESULT DeckLinkOutput::Frame::GetBytes(/* out */ void **buffer)
680 {
681         *buffer = uyvy_ptr_local.get();
682         return S_OK;
683 }
684
685 HRESULT DeckLinkOutput::Frame::GetTimecode(/* in */ BMDTimecodeFormat format, /* out */ IDeckLinkTimecode **timecode)
686 {
687         fprintf(stderr, "STUB: GetTimecode()\n");
688         return E_NOTIMPL;
689 }
690
691 HRESULT DeckLinkOutput::Frame::GetAncillaryData(/* out */ IDeckLinkVideoFrameAncillary **ancillary)
692 {
693         fprintf(stderr, "STUB: GetAncillaryData()\n");
694         return E_NOTIMPL;
695 }