X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=80b0df1ea0df60d852bc530cb0d5535e99e2f92a;hb=8d575ed1bf8c01daaeb4bdf87de53d089448ba93;hp=3d4cccb4a6b4a372ca6e7e36c8442925334ac324;hpb=10ff4d7e0cba7d0c5d11317aa5e11679382d3cfe;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 3d4cccb..80b0df1 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -33,7 +33,7 @@ #include "decklink_capture.h" #include "defs.h" #include "flags.h" -#include "h264encode.h" +#include "video_encoder.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" #include "timebase.h" @@ -92,23 +92,6 @@ void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced } } -string generate_local_dump_filename(int frame) -{ - time_t now = time(NULL); - tm now_tm; - localtime_r(&now, &now_tm); - - char timestamp[256]; - strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm); - - // Use the frame number to disambiguate between two cuts starting - // on the same second. - char filename[256]; - snprintf(filename, sizeof(filename), "%s%s-f%02d%s", - LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX); - return filename; -} - } // namespace void QueueLengthPolicy::update_policy(int queue_length) @@ -127,9 +110,9 @@ void QueueLengthPolicy::update_policy(int queue_length) if (queue_length >= int(safe_queue_length)) { been_at_safe_point_since_last_starvation = true; } - if (++frames_with_at_least_one >= 50 && safe_queue_length > 0) { + if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) { --safe_queue_length; - fprintf(stderr, "Card %u: Spare frames for more than 50 frames, reducing safe limit to %u frames\n", + fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n", card_index, safe_queue_length); frames_with_at_least_one = 0; } @@ -159,6 +142,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards)); for (unsigned i = 0; i < NUM_OUTPUTS; ++i) { output_channel[i].parent = this; + output_channel[i].channel = i; } ImageFormat inout_format; @@ -174,10 +158,9 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) display_chain->set_dither_bits(0); // Don't bother. display_chain->finalize(); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); - h264_encoder->open_output_file(generate_local_dump_filename(/*frame=*/0).c_str()); + video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); - // Start listening for clients only once H264Encoder has written its header, if any. + // Start listening for clients only once VideoEncoder has written its header, if any. httpd.start(9095); // First try initializing the PCI devices, then USB, until we have the desired number of cards. @@ -274,7 +257,9 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) // and there's a limit to how important the peak meter is. peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0); - alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); + if (global_flags.enable_alsa_output) { + alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); + } } Mixer::~Mixer() @@ -292,7 +277,7 @@ Mixer::~Mixer() cards[card_index].capture->stop_dequeue_thread(); } - h264_encoder.reset(nullptr); + video_encoder.reset(nullptr); } void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture) @@ -305,19 +290,7 @@ void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, Ca card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB. card->capture->set_video_frame_allocator(card->frame_allocator.get()); card->surface = create_surface(format); - card->capture->set_dequeue_thread_callbacks( - [card]{ - eglBindAPI(EGL_OPENGL_API); - card->context = create_context(card->surface); - if (!make_current(card->context, card->surface)) { - printf("failed to create bmusb context\n"); - exit(1); - } - }, - [this]{ - resource_pool->clean_context(); - }); - card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); card->capture->configure_card(); } @@ -386,7 +359,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, } } - int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom; + int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom; size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0; if (num_samples > OUTPUT_FREQUENCY / 10) { @@ -437,7 +410,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, if (dropped_frames > MAX_FPS * 2) { fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n", card_index, card->last_timecode, timecode); - card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); dropped_frames = 0; } else if (dropped_frames > 0) { // Insert silence as needed. @@ -519,59 +492,64 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, size_t y_offset = video_frame.size / 2 + video_offset / 2; for (unsigned field = 0; field < num_fields; ++field) { - unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22); - - if (userdata->tex_y[field] == 0 || - userdata->tex_cbcr[field] == 0 || - video_format.width != userdata->last_width[field] || - video_format.height != userdata->last_height[field]) { - // We changed resolution since last use of this texture, so we need to create - // a new object. Note that this each card has its own PBOFrameAllocator, - // we don't need to worry about these flip-flopping between resolutions. + // Put the actual texture upload in a lambda that is executed in the main thread. + // It is entirely possible to do this in the same thread (and it might even be + // faster, depending on the GPU and driver), but it appears to be trickling + // driver bugs very easily. + // + // Note that this means we must hold on to the actual frame data in + // until the upload command is run, but we hold on to much longer than that + // (in fact, all the way until we no longer use the texture in rendering). + auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() { + unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22); + + if (userdata->tex_y[field] == 0 || + userdata->tex_cbcr[field] == 0 || + video_format.width != userdata->last_width[field] || + video_format.height != userdata->last_height[field]) { + // We changed resolution since last use of this texture, so we need to create + // a new object. Note that this each card has its own PBOFrameAllocator, + // we don't need to worry about these flip-flopping between resolutions. + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + userdata->last_width[field] = video_format.width; + userdata->last_height[field] = video_format.height; + } + + GLuint pbo = userdata->pbo; + check_error(); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo); + check_error(); + + size_t field_y_start = y_offset + video_format.width * field_start_line; + size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t); + + if (global_flags.flush_pbos) { + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height); + check_error(); + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t)); + check_error(); + } + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); check_error(); - glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start)); check_error(); glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); check_error(); - glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start)); check_error(); - userdata->last_width[field] = video_format.width; - userdata->last_height[field] = video_format.height; - } - - GLuint pbo = userdata->pbo; - check_error(); - glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo); - check_error(); - - size_t field_y_start = y_offset + video_format.width * field_start_line; - size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t); - - if (global_flags.flush_pbos) { - glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height); + glBindTexture(GL_TEXTURE_2D, 0); check_error(); - glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t)); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); check_error(); - } - - glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start)); - check_error(); - glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start)); - check_error(); - glBindTexture(GL_TEXTURE_2D, 0); - check_error(); - glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); - check_error(); - RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); - check_error(); - assert(fence.get() != nullptr); - glFlush(); // Make sure the main thread doesn't have to wait until we push out enough frames to make a new command buffer. - check_error(); + }; if (field == 1) { // Don't upload the second field as fast as we can; wait until @@ -599,7 +577,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, new_frame.length = frame_length; new_frame.field = field; new_frame.interlaced = video_format.interlaced; - new_frame.ready_fence = fence; + new_frame.upload_func = upload_func; new_frame.dropped_frames = dropped_frames; card->new_frames.push(move(new_frame)); card->new_frames_changed.notify_all(); @@ -666,19 +644,17 @@ void Mixer::thread_func() insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state); check_error(); - // The new texture might still be uploaded, - // tell the GPU to wait until it's there. - if (new_frame->ready_fence) { - glWaitSync(new_frame->ready_fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); - check_error(); - new_frame->ready_fence.reset(); - check_error(); + // The new texture might need uploading before use. + if (new_frame->upload_func) { + new_frame->upload_func(); + new_frame->upload_func = nullptr; } } - render_one_frame(); + int64_t duration = new_frames[master_card_index].length; + render_one_frame(duration); ++frame; - pts_int += new_frames[master_card_index].length; + pts_int += duration; clock_gettime(CLOCK_MONOTONIC, &now); double elapsed = now.tv_sec - start.tv_sec + @@ -691,12 +667,7 @@ void Mixer::thread_func() } if (should_cut.exchange(false)) { // Test and clear. - string filename = generate_local_dump_filename(frame); - printf("Starting new recording: %s\n", filename.c_str()); - h264_encoder->close_output_file(); - h264_encoder->shutdown(); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); - h264_encoder->open_output_file(filename.c_str()); + video_encoder->do_cut(frame); } #if 0 @@ -774,7 +745,7 @@ void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_sam } } -void Mixer::render_one_frame() +void Mixer::render_one_frame(int64_t duration) { // Get the main chain from the theme, and set its state immediately. Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state); @@ -783,7 +754,7 @@ void Mixer::render_one_frame() //theme_main_chain.chain->enable_phase_timing(true); GLuint y_tex, cbcr_tex; - bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex); + bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex); assert(got_frame); // Render main chain. @@ -805,7 +776,7 @@ void Mixer::render_one_frame() glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded. - RefCountedGLsync fence = h264_encoder->end_frame(pts_int + av_delay, theme_main_chain.input_frames); + RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames); // The live frame just shows the RGBA texture we just rendered. // It owns rgba_tex now. @@ -857,7 +828,10 @@ void Mixer::audio_thread_func() { unique_lock lock(audio_mutex); - audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); }); + audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); }); + if (should_quit) { + return; + } task = audio_task_queue.front(); audio_task_queue.pop(); } @@ -879,9 +853,7 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) samples_card.resize(num_samples * 2); { unique_lock lock(cards[card_index].audio_mutex); - if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) { - printf("Card %d reported previous underrun.\n", card_index); - } + cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples); } if (card_index == selected_audio_card) { samples_out = move(samples_card); @@ -1026,7 +998,7 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) } // And finally add them to the output. - h264_encoder->add_audio(frame_pts_int, move(samples_out)); + video_encoder->add_audio(frame_pts_int, move(samples_out)); } void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) @@ -1107,6 +1079,7 @@ void Mixer::start() void Mixer::quit() { should_quit = true; + audio_task_queue_changed.notify_one(); mixer_thread.join(); audio_thread.join(); } @@ -1170,9 +1143,47 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) has_ready_frame = true; } - if (has_new_frame_ready_callback) { + if (new_frame_ready_callback) { new_frame_ready_callback(); } + + // Reduce the number of callbacks by filtering duplicates. The reason + // why we bother doing this is that Qt seemingly can get into a state + // where its builds up an essentially unbounded queue of signals, + // consuming more and more memory, and there's no good way of collapsing + // user-defined signals or limiting the length of the queue. + if (transition_names_updated_callback) { + vector transition_names = global_mixer->get_transition_names(); + bool changed = false; + if (transition_names.size() != last_transition_names.size()) { + changed = true; + } else { + for (unsigned i = 0; i < transition_names.size(); ++i) { + if (transition_names[i] != last_transition_names[i]) { + changed = true; + break; + } + } + } + if (changed) { + transition_names_updated_callback(transition_names); + last_transition_names = transition_names; + } + } + if (name_updated_callback) { + string name = global_mixer->get_channel_name(channel); + if (name != last_name) { + name_updated_callback(name); + last_name = name; + } + } + if (color_updated_callback) { + string color = global_mixer->get_channel_color(channel); + if (color != last_color) { + color_updated_callback(color); + last_color = color; + } + } } bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) @@ -1203,7 +1214,21 @@ bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback) { new_frame_ready_callback = callback; - has_new_frame_ready_callback = true; +} + +void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback) +{ + transition_names_updated_callback = callback; +} + +void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback) +{ + name_updated_callback = callback; +} + +void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback) +{ + color_updated_callback = callback; } mutex RefCountedGLsync::fence_lock;