X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=7a3f437d6ec745048c57be3283005cbf94597937;hb=703e00da89118df9be0354dda621bed023e6030e;hp=dd8cc6013d72559bb08b89c6d5605953ac6b4269;hpb=a76872873dda7a4fc9f41972486c234699f43b23;p=nageru diff --git a/mixer.cpp b/mixer.cpp index dd8cc60..7a3f437 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -4,61 +4,73 @@ #include #include -#include #include #include #include #include +#include #include -#include +#include #include #include #include -#include -#include #include -#include +#include #include #include +#include #include #include +#include #include #include #include #include +#include "DeckLinkAPI.h" +#include "LinuxCOM.h" +#include "alsa_output.h" +#include "basic_stats.h" #include "bmusb/bmusb.h" +#include "bmusb/fake_capture.h" +#ifdef HAVE_CEF +#include "cef_capture.h" +#endif +#include "chroma_subsampler.h" #include "context.h" +#include "decklink_capture.h" +#include "decklink_output.h" #include "defs.h" -#include "h264encode.h" +#include "disk_space_estimator.h" +#include "ffmpeg_capture.h" +#include "flags.h" +#include "input_mapping.h" +#include "metrics.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" +#include "resampling_queue.h" #include "timebase.h" +#include "timecode_renderer.h" +#include "v210_converter.h" +#include "video_encoder.h" + +#undef Status +#include +#include "json.pb.h" +class IDeckLink; class QOpenGLContext; using namespace movit; using namespace std; +using namespace std::chrono; using namespace std::placeholders; +using namespace bmusb; Mixer *global_mixer = nullptr; namespace { -void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples) -{ - for (size_t i = 0; i < num_samples; ++i) { - for (size_t j = 0; j < out_channels; ++j) { - uint32_t s1 = *src++; - uint32_t s2 = *src++; - uint32_t s3 = *src++; - uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24); - dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f); - } - src += 3 * (in_channels - out_channels); - } -} - void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state) { if (interlaced) { @@ -74,194 +86,660 @@ void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced } } -string generate_local_dump_filename(int frame) +void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width) { - time_t now = time(NULL); - tm now_tm; - localtime_r(&now, &now_tm); + bool first; + switch (userdata->pixel_format) { + case PixelFormat_10BitYCbCr: + first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0; + break; + case PixelFormat_8BitYCbCr: + first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0; + break; + case PixelFormat_8BitBGRA: + first = userdata->tex_rgba[field] == 0; + break; + case PixelFormat_8BitYCbCrPlanar: + first = userdata->tex_y[field] == 0 || userdata->tex_cb[field] == 0 || userdata->tex_cr[field] == 0; + break; + default: + assert(false); + } - char timestamp[256]; - strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm); + if (first || + width != userdata->last_width[field] || + height != userdata->last_height[field] || + cbcr_width != userdata->last_cbcr_width[field] || + cbcr_height != userdata->last_cbcr_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. + switch (userdata->pixel_format) { + case PixelFormat_10BitYCbCr: + glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr); + check_error(); + break; + case PixelFormat_8BitYCbCr: { + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, 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, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + break; + } + case PixelFormat_8BitYCbCrPlanar: { + glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_cb[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_cr[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, cbcr_width, cbcr_height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + break; + } + case PixelFormat_8BitBGRA: + glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]); + check_error(); + if (global_flags.can_disable_srgb_decoder) { // See the comments in tweaked_inputs.h. + glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr); + } else { + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr); + } + check_error(); + break; + default: + assert(false); + } + userdata->last_width[field] = width; + userdata->last_height[field] = height; + userdata->last_cbcr_width[field] = cbcr_width; + userdata->last_cbcr_height[field] = cbcr_height; + } + if (global_flags.ten_bit_input && + (first || v210_width != userdata->last_v210_width[field])) { + // Same as above; we need to recreate the texture. + glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr); + check_error(); + userdata->last_v210_width[field] = v210_width; + } +} - // 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; +void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset) +{ + if (interlaced_stride) { + stride *= 2; + } + if (global_flags.flush_pbos) { + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height); + check_error(); + } + + glBindTexture(GL_TEXTURE_2D, tex); + check_error(); + if (interlaced_stride) { + glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2); + check_error(); + } else { + glPixelStorei(GL_UNPACK_ROW_LENGTH, 0); + check_error(); + } + + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset)); + check_error(); + glBindTexture(GL_TEXTURE_2D, 0); + check_error(); + glPixelStorei(GL_UNPACK_ROW_LENGTH, 0); + check_error(); } } // namespace +void JitterHistory::register_metrics(const vector> &labels) +{ + global_metrics.add("input_underestimated_jitter_frames", labels, &metric_input_underestimated_jitter_frames); + global_metrics.add("input_estimated_max_jitter_seconds", labels, &metric_input_estimated_max_jitter_seconds, Metrics::TYPE_GAUGE); +} + +void JitterHistory::unregister_metrics(const vector> &labels) +{ + global_metrics.remove("input_underestimated_jitter_frames", labels); + global_metrics.remove("input_estimated_max_jitter_seconds", labels); +} + +void JitterHistory::frame_arrived(steady_clock::time_point now, int64_t frame_duration, size_t dropped_frames) +{ + if (expected_timestamp > steady_clock::time_point::min()) { + expected_timestamp += dropped_frames * nanoseconds(frame_duration * 1000000000 / TIMEBASE); + double jitter_seconds = fabs(duration(expected_timestamp - now).count()); + history.push_back(orders.insert(jitter_seconds)); + if (jitter_seconds > estimate_max_jitter()) { + ++metric_input_underestimated_jitter_frames; + } + + metric_input_estimated_max_jitter_seconds = estimate_max_jitter(); + + if (history.size() > history_length) { + orders.erase(history.front()); + history.pop_front(); + } + assert(history.size() <= history_length); + } + expected_timestamp = now + nanoseconds(frame_duration * 1000000000 / TIMEBASE); +} + +double JitterHistory::estimate_max_jitter() const +{ + if (orders.empty()) { + return 0.0; + } + size_t elem_idx = lrint((orders.size() - 1) * percentile); + if (percentile <= 0.5) { + return *next(orders.begin(), elem_idx) * multiplier; + } else { + return *prev(orders.end(), orders.size() - elem_idx) * multiplier; + } +} + +void QueueLengthPolicy::register_metrics(const vector> &labels) +{ + global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE); +} + +void QueueLengthPolicy::unregister_metrics(const vector> &labels) +{ + global_metrics.remove("input_queue_safe_length_frames", labels); +} + +void QueueLengthPolicy::update_policy(steady_clock::time_point now, + steady_clock::time_point expected_next_frame, + int64_t input_frame_duration, + int64_t master_frame_duration, + double max_input_card_jitter_seconds, + double max_master_card_jitter_seconds) +{ + double input_frame_duration_seconds = input_frame_duration / double(TIMEBASE); + double master_frame_duration_seconds = master_frame_duration / double(TIMEBASE); + + // Figure out when we can expect the next frame for this card, assuming + // worst-case jitter (ie., the frame is maximally late). + double seconds_until_next_frame = max(duration(expected_next_frame - now).count() + max_input_card_jitter_seconds, 0.0); + + // How many times are the master card expected to tick in that time? + // We assume the master clock has worst-case jitter but not any rate + // discrepancy, ie., it ticks as early as possible every time, but not + // cumulatively. + double frames_needed = (seconds_until_next_frame + max_master_card_jitter_seconds) / master_frame_duration_seconds; + + // As a special case, if the master card ticks faster than the input card, + // we expect the queue to drain by itself even without dropping. But if + // the difference is small (e.g. 60 Hz master and 59.94 input), it would + // go slowly enough that the effect wouldn't really be appreciable. + // We account for this by looking at the situation five frames ahead, + // assuming everything else is the same. + double frames_allowed; + if (master_frame_duration < input_frame_duration) { + frames_allowed = frames_needed + 5 * (input_frame_duration_seconds - master_frame_duration_seconds) / master_frame_duration_seconds; + } else { + frames_allowed = frames_needed; + } + + safe_queue_length = max(floor(frames_allowed), 0); + metric_input_queue_safe_length_frames = safe_queue_length; +} + Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) - : httpd(WIDTH, HEIGHT), + : httpd(), num_cards(num_cards), mixer_surface(create_surface(format)), h264_encoder_surface(create_surface(format)), - level_compressor(OUTPUT_FREQUENCY), - limiter(OUTPUT_FREQUENCY), - compressor(OUTPUT_FREQUENCY) + decklink_output_surface(create_surface(format)) { - httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str()); - httpd.start(9095); - + memcpy(ycbcr_interpretation, global_flags.ycbcr_interpretation, sizeof(ycbcr_interpretation)); CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF)); check_error(); + // This nearly always should be true. + global_flags.can_disable_srgb_decoder = + epoxy_has_gl_extension("GL_EXT_texture_sRGB_decode") && + epoxy_has_gl_extension("GL_ARB_sampler_objects"); + // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision // will be halved when sampling them, and we need to compensate here. movit_texel_subpixel_precision /= 2.0; resource_pool.reset(new ResourcePool); - theme.reset(new Theme("theme.lua", 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; inout_format.color_space = COLORSPACE_sRGB; inout_format.gamma_curve = GAMMA_sRGB; - // Display chain; shows the live output produced by the main chain (its RGBA version). - display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get())); + // Matches the 4:2:0 format created by the main chain. + YCbCrFormat ycbcr_format; + ycbcr_format.chroma_subsampling_x = 2; + ycbcr_format.chroma_subsampling_y = 2; + if (global_flags.ycbcr_rec709_coefficients) { + ycbcr_format.luma_coefficients = YCBCR_REC_709; + } else { + ycbcr_format.luma_coefficients = YCBCR_REC_601; + } + ycbcr_format.full_range = false; + ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth; + ycbcr_format.cb_x_position = 0.0f; + ycbcr_format.cr_x_position = 0.0f; + ycbcr_format.cb_y_position = 0.5f; + ycbcr_format.cr_y_position = 0.5f; + + // Display chain; shows the live output produced by the main chain (or rather, a copy of it). + display_chain.reset(new EffectChain(global_flags.width, global_flags.height, resource_pool.get())); check_error(); - display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT); // FIXME: GL_UNSIGNED_BYTE is really wrong. + GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE; + display_input = new YCbCrInput(inout_format, ycbcr_format, global_flags.width, global_flags.height, YCBCR_INPUT_SPLIT_Y_AND_CBCR, type); display_chain->add_input(display_input); display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED); display_chain->set_dither_bits(0); // Don't bother. display_chain->finalize(); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd)); + video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator)); - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - printf("Configuring card %d...\n", card_index); - CaptureCard *card = &cards[card_index]; - card->usb = new BMUSBCapture(card_index); - card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7)); - card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB. - card->usb->set_video_frame_allocator(card->frame_allocator.get()); - card->surface = create_surface(format); - card->usb->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); + // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy. + theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards)); + + // Must be instantiated after the theme, as the theme decides the number of FFmpeg inputs. + std::vector video_inputs = theme->get_video_inputs(); + audio_mixer.reset(new AudioMixer(num_cards, video_inputs.size())); + + httpd.add_endpoint("/channels", bind(&Mixer::get_channels_json, this), HTTPD::ALLOW_ALL_ORIGINS); + for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) { + char url[256]; + snprintf(url, sizeof(url), "/channels/%d/color", channel_idx); + httpd.add_endpoint(url, bind(&Mixer::get_channel_color_http, this, unsigned(channel_idx)), HTTPD::ALLOW_ALL_ORIGINS); + } + + // Start listening for clients only once VideoEncoder has written its header, if any. + httpd.start(global_flags.http_port); + + // First try initializing the then PCI devices, then USB, then + // fill up with fake cards until we have the desired number of cards. + unsigned num_pci_devices = 0; + unsigned card_index = 0; + + { + IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance(); + if (decklink_iterator != nullptr) { + for ( ; card_index < num_cards; ++card_index) { + IDeckLink *decklink; + if (decklink_iterator->Next(&decklink) != S_OK) { + break; } - }, - [this]{ - resource_pool->clean_context(); - }); - card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); - card->usb->configure_card(); + + DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index); + DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index); + if (!output->set_device(decklink)) { + delete output; + output = nullptr; + } + configure_card(card_index, capture, CardType::LIVE_CARD, output); + ++num_pci_devices; + } + decklink_iterator->Release(); + fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices); + } else { + fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n"); + } } - BMUSBCapture::start_bm_thread(); + unsigned num_usb_devices = BMUSBCapture::num_cards(); + for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) { + BMUSBCapture *capture = new BMUSBCapture(usb_card_index); + capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index)); + configure_card(card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr); + } + fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices); - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - cards[card_index].usb->start_bm_capture(); + unsigned num_fake_cards = 0; + for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) { + FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio); + configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr); } - // Set up stuff for NV12 conversion. + if (num_fake_cards > 0) { + fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards); + } - // Cb/Cr shader. - string cbcr_vert_shader = read_file("vs-cbcr.130.vert"); - string cbcr_frag_shader = - "#version 130 \n" - "in vec2 tc0; \n" - "uniform sampler2D cbcr_tex; \n" - "void main() { \n" - " gl_FragColor = texture2D(cbcr_tex, tc0); \n" - "} \n"; - vector frag_shader_outputs; - cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs); + // Initialize all video inputs the theme asked for. Note that these are + // all put _after_ the regular cards, which stop at - 1. + for (unsigned video_card_index = 0; video_card_index < video_inputs.size(); ++card_index, ++video_card_index) { + if (card_index >= MAX_VIDEO_CARDS) { + fprintf(stderr, "ERROR: Not enough card slots available for the videos the theme requested.\n"); + exit(1); + } + configure_card(card_index, video_inputs[video_card_index], CardType::FFMPEG_INPUT, /*output=*/nullptr); + video_inputs[video_card_index]->set_card_index(card_index); + } + num_video_inputs = video_inputs.size(); + +#ifdef HAVE_CEF + // Same, for HTML inputs. + std::vector html_inputs = theme->get_html_inputs(); + for (unsigned html_card_index = 0; html_card_index < html_inputs.size(); ++card_index, ++html_card_index) { + if (card_index >= MAX_VIDEO_CARDS) { + fprintf(stderr, "ERROR: Not enough card slots available for the HTML inputs the theme requested.\n"); + exit(1); + } + configure_card(card_index, html_inputs[html_card_index], CardType::CEF_INPUT, /*output=*/nullptr); + html_inputs[html_card_index]->set_card_index(card_index); + } + num_html_inputs = html_inputs.size(); +#endif - r128.init(2, OUTPUT_FREQUENCY); - r128.integr_start(); + BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1)); + BMUSBCapture::start_bm_thread(); + + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + cards[card_index].queue_length_policy.reset(card_index); + } + + chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get())); + + if (global_flags.ten_bit_input) { + if (!v210Converter::has_hardware_support()) { + fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n"); + fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n"); + exit(1); + } + v210_converter.reset(new v210Converter()); + + // These are all the widths listed in the Blackmagic SDK documentation + // (section 2.7.3, “Display Modes”). + v210_converter->precompile_shader(720); + v210_converter->precompile_shader(1280); + v210_converter->precompile_shader(1920); + v210_converter->precompile_shader(2048); + v210_converter->precompile_shader(3840); + v210_converter->precompile_shader(4096); + } + if (global_flags.ten_bit_output) { + if (!v210Converter::has_hardware_support()) { + fprintf(stderr, "ERROR: --ten-bit-output requires support for OpenGL compute shaders\n"); + fprintf(stderr, " (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n"); + exit(1); + } + } - locut.init(FILTER_HPF, 2); + timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height)); + display_timecode_in_stream = global_flags.display_timecode_in_stream; + display_timecode_on_stdout = global_flags.display_timecode_on_stdout; - // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise, - // 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); + if (global_flags.enable_alsa_output) { + alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); + } + if (global_flags.output_card != -1) { + desired_output_card_index = global_flags.output_card; + set_output_card_internal(global_flags.output_card); + } - alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); + output_jitter_history.register_metrics({{ "card", "output" }}); } Mixer::~Mixer() { - resource_pool->release_glsl_program(cbcr_program_num); + httpd.stop(); BMUSBCapture::stop_bm_thread(); - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { { - unique_lock lock(bmusb_mutex); + unique_lock lock(card_mutex); cards[card_index].should_quit = true; // Unblock thread. - cards[card_index].new_data_ready_changed.notify_all(); + cards[card_index].new_frames_changed.notify_all(); + } + cards[card_index].capture->stop_dequeue_thread(); + if (cards[card_index].output) { + cards[card_index].output->end_output(); + cards[card_index].output.reset(); } - cards[card_index].usb->stop_dequeue_thread(); } - h264_encoder.reset(nullptr); + video_encoder.reset(nullptr); } -namespace { - -int unwrap_timecode(uint16_t current_wrapped, int last) +void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output) { - uint16_t last_wrapped = last & 0xffff; - if (current_wrapped > last_wrapped) { - return (last & ~0xffff) | current_wrapped; + printf("Configuring card %d...\n", card_index); + + CaptureCard *card = &cards[card_index]; + if (card->capture != nullptr) { + card->capture->stop_dequeue_thread(); + } + card->capture.reset(capture); + card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE); + card->is_cef_capture = (card_type == CardType::CEF_INPUT); + card->may_have_dropped_last_frame = false; + card->type = card_type; + if (card->output.get() != output) { + card->output.reset(output); + } + + PixelFormat pixel_format; + if (card_type == CardType::FFMPEG_INPUT) { + pixel_format = capture->get_current_pixel_format(); + } else if (card_type == CardType::CEF_INPUT) { + pixel_format = PixelFormat_8BitBGRA; + } else if (global_flags.ten_bit_input) { + pixel_format = PixelFormat_10BitYCbCr; } else { - return 0x10000 + ((last & ~0xffff) | current_wrapped); + pixel_format = PixelFormat_8BitYCbCr; + } + + card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7)); + if (card->frame_allocator == nullptr) { + card->frame_allocator.reset(new PBOFrameAllocator(pixel_format, 8 << 20, global_flags.width, global_flags.height)); // 8 MB. + } + card->capture->set_video_frame_allocator(card->frame_allocator.get()); + if (card->surface == nullptr) { + card->surface = create_surface_with_same_format(mixer_surface); + } + while (!card->new_frames.empty()) card->new_frames.pop_front(); + card->last_timecode = -1; + card->capture->set_pixel_format(pixel_format); + card->capture->configure_card(); + + // NOTE: start_bm_capture() happens in thread_func(). + + DeviceSpec device; + if (card_type == CardType::FFMPEG_INPUT) { + device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards}; + } else { + device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index}; } + audio_mixer->reset_resampler(device); + audio_mixer->set_display_name(device, card->capture->get_description()); + audio_mixer->trigger_state_changed_callback(); + + // Unregister old metrics, if any. + if (!card->labels.empty()) { + const vector> &labels = card->labels; + card->jitter_history.unregister_metrics(labels); + card->queue_length_policy.unregister_metrics(labels); + global_metrics.remove("input_received_frames", labels); + global_metrics.remove("input_dropped_frames_jitter", labels); + global_metrics.remove("input_dropped_frames_error", labels); + global_metrics.remove("input_dropped_frames_resets", labels); + global_metrics.remove("input_queue_length_frames", labels); + global_metrics.remove("input_queue_duped_frames", labels); + + global_metrics.remove("input_has_signal_bool", labels); + global_metrics.remove("input_is_connected_bool", labels); + global_metrics.remove("input_interlaced_bool", labels); + global_metrics.remove("input_width_pixels", labels); + global_metrics.remove("input_height_pixels", labels); + global_metrics.remove("input_frame_rate_nom", labels); + global_metrics.remove("input_frame_rate_den", labels); + global_metrics.remove("input_sample_rate_hz", labels); + } + + // Register metrics. + vector> labels; + char card_name[64]; + snprintf(card_name, sizeof(card_name), "%d", card_index); + labels.emplace_back("card", card_name); + + switch (card_type) { + case CardType::LIVE_CARD: + labels.emplace_back("cardtype", "live"); + break; + case CardType::FAKE_CAPTURE: + labels.emplace_back("cardtype", "fake"); + break; + case CardType::FFMPEG_INPUT: + labels.emplace_back("cardtype", "ffmpeg"); + break; + case CardType::CEF_INPUT: + labels.emplace_back("cardtype", "cef"); + break; + default: + assert(false); + } + card->jitter_history.register_metrics(labels); + card->queue_length_policy.register_metrics(labels); + global_metrics.add("input_received_frames", labels, &card->metric_input_received_frames); + global_metrics.add("input_dropped_frames_jitter", labels, &card->metric_input_dropped_frames_jitter); + global_metrics.add("input_dropped_frames_error", labels, &card->metric_input_dropped_frames_error); + global_metrics.add("input_dropped_frames_resets", labels, &card->metric_input_resets); + global_metrics.add("input_queue_length_frames", labels, &card->metric_input_queue_length_frames, Metrics::TYPE_GAUGE); + global_metrics.add("input_queue_duped_frames", labels, &card->metric_input_duped_frames); + + global_metrics.add("input_has_signal_bool", labels, &card->metric_input_has_signal_bool, Metrics::TYPE_GAUGE); + global_metrics.add("input_is_connected_bool", labels, &card->metric_input_is_connected_bool, Metrics::TYPE_GAUGE); + global_metrics.add("input_interlaced_bool", labels, &card->metric_input_interlaced_bool, Metrics::TYPE_GAUGE); + global_metrics.add("input_width_pixels", labels, &card->metric_input_width_pixels, Metrics::TYPE_GAUGE); + global_metrics.add("input_height_pixels", labels, &card->metric_input_height_pixels, Metrics::TYPE_GAUGE); + global_metrics.add("input_frame_rate_nom", labels, &card->metric_input_frame_rate_nom, Metrics::TYPE_GAUGE); + global_metrics.add("input_frame_rate_den", labels, &card->metric_input_frame_rate_den, Metrics::TYPE_GAUGE); + global_metrics.add("input_sample_rate_hz", labels, &card->metric_input_sample_rate_hz, Metrics::TYPE_GAUGE); + card->labels = labels; } -float find_peak(const float *samples, size_t num_samples) +void Mixer::set_output_card_internal(int card_index) { - float m = fabs(samples[0]); - for (size_t i = 1; i < num_samples; ++i) { - m = max(m, fabs(samples[i])); + // We don't really need to take card_mutex, since we're in the mixer + // thread and don't mess with any queues (which is the only thing that happens + // from other threads), but it's probably the safest in the long run. + unique_lock lock(card_mutex); + if (output_card_index != -1) { + // Switch the old card from output to input. + CaptureCard *old_card = &cards[output_card_index]; + old_card->output->end_output(); + + // Stop the fake card that we put into place. + // This needs to _not_ happen under the mutex, to avoid deadlock + // (delivering the last frame needs to take the mutex). + CaptureInterface *fake_capture = old_card->capture.get(); + lock.unlock(); + fake_capture->stop_dequeue_thread(); + lock.lock(); + old_card->capture = move(old_card->parked_capture); // TODO: reset the metrics + old_card->is_fake_capture = false; + old_card->capture->start_bm_capture(); } - return m; + if (card_index != -1) { + CaptureCard *card = &cards[card_index]; + CaptureInterface *capture = card->capture.get(); + // TODO: DeckLinkCapture::stop_dequeue_thread can actually take + // several seconds to complete (blocking on DisableVideoInput); + // see if we can maybe do it asynchronously. + lock.unlock(); + capture->stop_dequeue_thread(); + lock.lock(); + card->parked_capture = move(card->capture); + CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio); + configure_card(card_index, fake_capture, CardType::FAKE_CAPTURE, card->output.release()); + card->queue_length_policy.reset(card_index); + card->capture->start_bm_capture(); + desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode); + card->output->start_output(desired_output_video_mode, pts_int); + } + output_card_index = card_index; + output_jitter_history.clear(); } -void deinterleave_samples(const vector &in, vector *out_l, vector *out_r) -{ - size_t num_samples = in.size() / 2; - out_l->resize(num_samples); - out_r->resize(num_samples); +namespace { - const float *inptr = in.data(); - float *lptr = &(*out_l)[0]; - float *rptr = &(*out_r)[0]; - for (size_t i = 0; i < num_samples; ++i) { - *lptr++ = *inptr++; - *rptr++ = *inptr++; +int unwrap_timecode(uint16_t current_wrapped, int last) +{ + uint16_t last_wrapped = last & 0xffff; + if (current_wrapped > last_wrapped) { + return (last & ~0xffff) | current_wrapped; + } else { + return 0x10000 + ((last & ~0xffff) | current_wrapped); } } } // namespace void Mixer::bm_frame(unsigned card_index, uint16_t timecode, - FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format, - FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format) + FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format, + FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format) { + DeviceSpec device; + if (card_index >= num_cards) { + device = DeviceSpec{InputSourceType::FFMPEG_VIDEO_INPUT, card_index - num_cards}; + } else { + device = DeviceSpec{InputSourceType::CAPTURE_CARD, card_index}; + } CaptureCard *card = &cards[card_index]; - unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom; - bool interlaced; + ++card->metric_input_received_frames; + card->metric_input_has_signal_bool = video_format.has_signal; + card->metric_input_is_connected_bool = video_format.is_connected; + card->metric_input_interlaced_bool = video_format.interlaced; + card->metric_input_width_pixels = video_format.width; + card->metric_input_height_pixels = video_format.height; + card->metric_input_frame_rate_nom = video_format.frame_rate_nom; + card->metric_input_frame_rate_den = video_format.frame_rate_den; + card->metric_input_sample_rate_hz = audio_format.sample_rate; + + if (is_mode_scanning[card_index]) { + if (video_format.has_signal) { + // Found a stable signal, so stop scanning. + is_mode_scanning[card_index] = false; + } else { + static constexpr double switch_time_s = 0.1; // Should be enough time for the signal to stabilize. + steady_clock::time_point now = steady_clock::now(); + double sec_since_last_switch = duration(steady_clock::now() - last_mode_scan_change[card_index]).count(); + if (sec_since_last_switch > switch_time_s) { + // It isn't this mode; try the next one. + mode_scanlist_index[card_index]++; + mode_scanlist_index[card_index] %= mode_scanlist[card_index].size(); + cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]); + last_mode_scan_change[card_index] = now; + } + } + } - decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom, - &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now. - int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom; + int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom; + assert(frame_length > 0); - size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0; - if (num_samples > OUTPUT_FREQUENCY / 10) { - printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n", - card_index, int(audio_frame.len), int(audio_offset), - timecode, int(video_frame.len), int(video_offset), video_format); + 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 && card->type != CardType::FFMPEG_INPUT) { + printf("%s: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n", + spec_to_string(device).c_str(), int(audio_frame.len), int(audio_offset), + timecode, int(video_frame.len), int(video_offset), video_format.id); if (video_frame.owner) { video_frame.owner->release_frame(video_frame); } @@ -271,71 +749,71 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, return; } - int64_t local_pts = card->next_local_pts; int dropped_frames = 0; if (card->last_timecode != -1) { dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1; } - // Convert the audio to stereo fp32 and add it. - vector audio; - audio.resize(num_samples * 2); - convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples); - - // Add the audio. - { - unique_lock lock(card->audio_mutex); - - // Number of samples per frame if we need to insert silence. - // (Could be nonintegral, but resampling will save us then.) - int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom; - - 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)); - dropped_frames = 0; - } else if (dropped_frames > 0) { - // Insert silence as needed. - fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n", - card_index, dropped_frames, timecode); - vector silence(silence_samples * 2, 0.0f); - for (int i = 0; i < dropped_frames; ++i) { - card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples); - // Note that if the format changed in the meantime, we have - // no way of detecting that; we just have to assume the frame length - // is always the same. - local_pts += frame_length; - } - } - if (num_samples == 0) { - audio.resize(silence_samples * 2); - num_samples = silence_samples; - } - card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples); - card->next_local_pts = local_pts + frame_length; + // Number of samples per frame if we need to insert silence. + // (Could be nonintegral, but resampling will save us then.) + const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom; + + if (dropped_frames > MAX_FPS * 2) { + fprintf(stderr, "%s lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n", + spec_to_string(device).c_str(), card->last_timecode, timecode); + audio_mixer->reset_resampler(device); + dropped_frames = 0; + ++card->metric_input_resets; + } else if (dropped_frames > 0) { + // Insert silence as needed. + fprintf(stderr, "%s dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n", + spec_to_string(device).c_str(), dropped_frames, timecode); + card->metric_input_dropped_frames_error += dropped_frames; + + bool success; + do { + success = audio_mixer->add_silence(device, silence_samples, dropped_frames, frame_length); + } while (!success); } - card->last_timecode = timecode; + if (num_samples > 0) { + audio_mixer->add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp); + } // Done with the audio, so release it. if (audio_frame.owner) { audio_frame.owner->release_frame(audio_frame); } - { - // Wait until the previous frame was consumed. - unique_lock lock(bmusb_mutex); - card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; }); - if (card->should_quit) return; - } + card->last_timecode = timecode; + + PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; - size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 2; + size_t cbcr_width, cbcr_height, cbcr_offset, y_offset; + size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom); + if (userdata != nullptr && userdata->pixel_format == PixelFormat_8BitYCbCrPlanar) { + // The calculation above is wrong for planar Y'CbCr, so just override it. + assert(card->type == CardType::FFMPEG_INPUT); + assert(video_offset == 0); + expected_length = video_frame.len; + + userdata->ycbcr_format = (static_cast(card->capture.get()))->get_current_frame_ycbcr_format(); + cbcr_width = video_format.width / userdata->ycbcr_format.chroma_subsampling_x; + cbcr_height = video_format.height / userdata->ycbcr_format.chroma_subsampling_y; + cbcr_offset = video_format.width * video_format.height; + y_offset = 0; + } else { + // All the other Y'CbCr formats are 4:2:2. + cbcr_width = video_format.width / 2; + cbcr_height = video_format.height; + cbcr_offset = video_offset / 2; + y_offset = video_frame.size / 2 + video_offset / 2; + } if (video_frame.len - video_offset == 0 || video_frame.len - video_offset != expected_length) { if (video_frame.len != 0) { - printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n", - card_index, video_frame.len - video_offset, expected_length); + printf("%s: Dropping video frame with wrong length (%ld; expected %ld)\n", + spec_to_string(device).c_str(), video_frame.len - video_offset, expected_length); } if (video_frame.owner) { video_frame.owner->release_frame(video_frame); @@ -344,84 +822,116 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, // Still send on the information that we _had_ a frame, even though it's corrupted, // so that pts can go up accordingly. { - unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = RefCountedFrame(FrameAllocator::Frame()); - card->new_frame_length = frame_length; - card->new_frame_interlaced = false; - card->new_data_ready_fence = nullptr; - card->dropped_frames = dropped_frames; - card->new_data_ready_changed.notify_all(); + unique_lock lock(card_mutex); + CaptureCard::NewFrame new_frame; + new_frame.frame = RefCountedFrame(FrameAllocator::Frame()); + new_frame.length = frame_length; + new_frame.interlaced = false; + new_frame.dropped_frames = dropped_frames; + new_frame.received_timestamp = video_frame.received_timestamp; + card->new_frames.push_back(move(new_frame)); + card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames); } + card->new_frames_changed.notify_all(); return; } - PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; - - unsigned num_fields = interlaced ? 2 : 1; - timespec frame_upload_start; - if (interlaced) { + unsigned num_fields = video_format.interlaced ? 2 : 1; + steady_clock::time_point frame_upload_start; + bool interlaced_stride = false; + if (video_format.interlaced) { // Send the two fields along as separate frames; the other side will need to add // a deinterlacer to actually get this right. - assert(height % 2 == 0); - height /= 2; + assert(video_format.height % 2 == 0); + video_format.height /= 2; + cbcr_height /= 2; assert(frame_length % 2 == 0); frame_length /= 2; num_fields = 2; - clock_gettime(CLOCK_MONOTONIC, &frame_upload_start); + if (video_format.second_field_start == 1) { + interlaced_stride = true; + } + frame_upload_start = steady_clock::now(); } - userdata->last_interlaced = interlaced; - userdata->last_frame_rate_nom = frame_rate_nom; - userdata->last_frame_rate_den = frame_rate_den; - RefCountedFrame new_frame(video_frame); + userdata->last_interlaced = video_format.interlaced; + userdata->last_has_signal = video_format.has_signal; + userdata->last_is_connected = video_format.is_connected; + userdata->last_frame_rate_nom = video_format.frame_rate_nom; + userdata->last_frame_rate_den = video_format.frame_rate_den; + RefCountedFrame frame(video_frame); // Upload the textures. - size_t cbcr_width = width / 2; - size_t cbcr_offset = video_offset / 2; - 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) ? second_field_start : extra_lines_top + field * (height + 22); - - if (userdata->tex_y[field] == 0 || - userdata->tex_cbcr[field] == 0 || - width != userdata->last_width[field] || - 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, 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, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + // 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 = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, cbcr_height, interlaced_stride, userdata]() { + unsigned field_start_line; + if (field == 1) { + field_start_line = video_format.second_field_start; + } else { + field_start_line = video_format.extra_lines_top; + } + + // For anything not FRAME_FORMAT_YCBCR_10BIT, v210_width will be nonsensical but not used. + size_t v210_width = video_format.stride / sizeof(uint32_t); + ensure_texture_resolution(userdata, field, video_format.width, video_format.height, cbcr_width, cbcr_height, v210_width); + + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo); check_error(); - userdata->last_width[field] = width; - userdata->last_height[field] = height; - } - GLuint pbo = userdata->pbo; - check_error(); - glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo); - check_error(); - glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT); - check_error(); + switch (userdata->pixel_format) { + case PixelFormat_10BitYCbCr: { + size_t field_start = video_offset + video_format.stride * field_start_line; + upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start); + v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height); + break; + } + case PixelFormat_8BitYCbCr: { + 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); + + // Make up our own strides, since we are interleaving. + upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start); + upload_texture(userdata->tex_cbcr[field], cbcr_width, cbcr_height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start); + break; + } + case PixelFormat_8BitYCbCrPlanar: { + assert(field_start_line == 0); // We don't really support interlaced here. + size_t field_y_start = y_offset; + size_t field_cb_start = cbcr_offset; + size_t field_cr_start = cbcr_offset + cbcr_width * cbcr_height; + + // Make up our own strides, since we are interleaving. + upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start); + upload_texture(userdata->tex_cb[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cb_start); + upload_texture(userdata->tex_cr[field], cbcr_width, cbcr_height, cbcr_width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_cr_start); + break; + } + case PixelFormat_8BitBGRA: { + size_t field_start = video_offset + video_format.stride * field_start_line; + upload_texture(userdata->tex_rgba[field], video_format.width, video_format.height, video_format.stride, interlaced_stride, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, field_start); + // These could be asked to deliver mipmaps at any time. + glBindTexture(GL_TEXTURE_2D, userdata->tex_rgba[field]); + check_error(); + glGenerateMipmap(GL_TEXTURE_2D); + check_error(); + glBindTexture(GL_TEXTURE_2D, 0); + check_error(); + break; + } + default: + assert(false); + } - glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t))); - check_error(); - glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line)); - check_error(); - glBindTexture(GL_TEXTURE_2D, 0); - check_error(); - GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); - check_error(); - assert(fence != nullptr); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0); + check_error(); + }; if (field == 1) { // Don't upload the second field as fast as we can; wait until @@ -430,40 +940,44 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, // against the video display, although the latter is not as critical.) // This requires our system clock to be reasonably close to the // video clock, but that's not an unreasonable assumption. - timespec second_field_start; - second_field_start.tv_nsec = frame_upload_start.tv_nsec + - frame_length * 1000000000 / TIMEBASE; - second_field_start.tv_sec = frame_upload_start.tv_sec + - second_field_start.tv_nsec / 1000000000; - second_field_start.tv_nsec %= 1000000000; - - while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, - &second_field_start, nullptr) == -1 && - errno == EINTR) ; + steady_clock::time_point second_field_start = frame_upload_start + + nanoseconds(frame_length * 1000000000 / TIMEBASE); + this_thread::sleep_until(second_field_start); } { - unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = new_frame; - card->new_frame_length = frame_length; - card->new_frame_field = field; - card->new_frame_interlaced = interlaced; - card->new_data_ready_fence = fence; - card->dropped_frames = dropped_frames; - card->new_data_ready_changed.notify_all(); - - if (field != num_fields - 1) { - // Wait until the previous frame was consumed. - card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; }); - if (card->should_quit) return; - } + unique_lock lock(card_mutex); + CaptureCard::NewFrame new_frame; + new_frame.frame = frame; + new_frame.length = frame_length; + new_frame.field = field; + new_frame.interlaced = video_format.interlaced; + new_frame.upload_func = upload_func; + new_frame.dropped_frames = dropped_frames; + new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp. + card->new_frames.push_back(move(new_frame)); + card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames); + card->may_have_dropped_last_frame = false; } + card->new_frames_changed.notify_all(); } } +void Mixer::bm_hotplug_add(libusb_device *dev) +{ + lock_guard lock(hotplug_mutex); + hotplugged_cards.push_back(dev); +} + +void Mixer::bm_hotplug_remove(unsigned card_index) +{ + cards[card_index].new_frames_changed.notify_all(); +} + void Mixer::thread_func() { + pthread_setname_np(pthread_self(), "Mixer_OpenGL"); + eglBindAPI(EGL_OPENGL_API); QOpenGLContext *context = create_context(mixer_surface); if (!make_current(context, mixer_surface)) { @@ -471,186 +985,100 @@ void Mixer::thread_func() exit(1); } - struct timespec start, now; - clock_gettime(CLOCK_MONOTONIC, &start); + // Start the actual capture. (We don't want to do it before we're actually ready + // to process output frames.) + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + if (int(card_index) != output_card_index) { + cards[card_index].capture->start_bm_capture(); + } + } - int frame = 0; + BasicStats basic_stats(/*verbose=*/true, /*use_opengl=*/true); int stats_dropped_frames = 0; while (!should_quit) { - CaptureCard card_copy[MAX_CARDS]; - int num_samples[MAX_CARDS]; - - { - unique_lock lock(bmusb_mutex); - - // The first card is the master timer, so wait for it to have a new frame. - // TODO: Make configurable, and with a timeout. - cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; }); - - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - CaptureCard *card = &cards[card_index]; - card_copy[card_index].usb = card->usb; - card_copy[card_index].new_data_ready = card->new_data_ready; - card_copy[card_index].new_frame = card->new_frame; - card_copy[card_index].new_frame_length = card->new_frame_length; - card_copy[card_index].new_frame_field = card->new_frame_field; - card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced; - card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence; - card_copy[card_index].dropped_frames = card->dropped_frames; - card->new_data_ready = false; - card->new_data_ready_changed.notify_all(); - - int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples; - num_samples[card_index] = num_samples_times_timebase / TIMEBASE; - card->fractional_samples = num_samples_times_timebase % TIMEBASE; - assert(num_samples[card_index] >= 0); - } + if (desired_output_card_index != output_card_index) { + set_output_card_internal(desired_output_card_index); + } + if (output_card_index != -1 && + desired_output_video_mode != output_video_mode) { + DeckLinkOutput *output = cards[output_card_index].output.get(); + output->end_output(); + desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode); + output->start_output(desired_output_video_mode, pts_int); } - // Resample the audio as needed, including from previously dropped frames. - for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) { - { - // Signal to the audio thread to process this frame. - unique_lock lock(audio_mutex); - audio_task_queue.push(AudioTask{pts_int, num_samples[0]}); - audio_task_queue_changed.notify_one(); - } - if (frame_num != card_copy[0].dropped_frames) { - // For dropped frames, increase the pts. Note that if the format changed - // in the meantime, we have no way of detecting that; we just have to - // assume the frame length is always the same. - ++stats_dropped_frames; - pts_int += card_copy[0].new_frame_length; - } + CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS]; + bool has_new_frame[MAX_VIDEO_CARDS] = { false }; + + bool master_card_is_output; + unsigned master_card_index; + if (output_card_index != -1) { + master_card_is_output = true; + master_card_index = output_card_index; + } else { + master_card_is_output = false; + master_card_index = theme->map_signal(master_clock_channel); + assert(master_card_index < num_cards + num_video_inputs); } - if (audio_level_callback != nullptr) { - unique_lock lock(r128_mutex); - double loudness_s = r128.loudness_S(); - double loudness_i = r128.integrated(); - double loudness_range_low = r128.range_min(); - double loudness_range_high = r128.range_max(); + OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame); + schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp); + stats_dropped_frames += output_frame_info.dropped_frames; - audio_level_callback(loudness_s, 20.0 * log10(peak), - loudness_i, loudness_range_low, loudness_range_high, - last_gain_staging_db); - } + handle_hotplugged_cards(); - for (unsigned card_index = 1; card_index < num_cards; ++card_index) { - if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) { - ++card_copy[card_index].dropped_frames; + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + if (card_index == master_card_index || !has_new_frame[card_index]) { + continue; + } + if (new_frames[card_index].frame->len == 0) { + ++new_frames[card_index].dropped_frames; } - if (card_copy[card_index].dropped_frames > 0) { + if (new_frames[card_index].dropped_frames > 0) { printf("Card %u dropped %d frames before this\n", - card_index, int(card_copy[card_index].dropped_frames)); + card_index, int(new_frames[card_index].dropped_frames)); } } // If the first card is reporting a corrupted or otherwise dropped frame, // just increase the pts (skipping over this frame) and don't try to compute anything new. - if (card_copy[0].new_frame->len == 0) { + if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) { ++stats_dropped_frames; - pts_int += card_copy[0].new_frame_length; + pts_int += new_frames[master_card_index].length; continue; } - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - CaptureCard *card = &card_copy[card_index]; - if (!card->new_data_ready || card->new_frame->len == 0) + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0) continue; - assert(card->new_frame != nullptr); - insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state); + CaptureCard::NewFrame *new_frame = &new_frames[card_index]; + assert(new_frame->frame != nullptr); + 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 (card->new_data_ready_fence) { - glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED); - check_error(); - glDeleteSync(card->new_data_ready_fence); - check_error(); + // The new texture might need uploading before use. + if (new_frame->upload_func) { + new_frame->upload_func(); + new_frame->upload_func = nullptr; } } - // 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); - EffectChain *chain = theme_main_chain.chain; - theme_main_chain.setup_chain(); - //theme_main_chain.chain->enable_phase_timing(true); - - GLuint y_tex, cbcr_tex; - bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex); - assert(got_frame); - - // Render main chain. - GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT); - GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up. - GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex); - check_error(); - chain->render_to_fbo(fbo, WIDTH, HEIGHT); - resource_pool->release_fbo(fbo); - - subsample_chroma(cbcr_full_tex, cbcr_tex); - resource_pool->release_2d_texture(cbcr_full_tex); - - // Set the right state for rgba_tex. - glBindFramebuffer(GL_FRAMEBUFFER, 0); - glBindTexture(GL_TEXTURE_2D, rgba_tex); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - - RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); - check_error(); - - const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded. - h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames); - ++frame; - pts_int += card_copy[0].new_frame_length; - - // The live frame just shows the RGBA texture we just rendered. - // It owns rgba_tex now. - DisplayFrame live_frame; - live_frame.chain = display_chain.get(); - live_frame.setup_chain = [this, rgba_tex]{ - display_input->set_texture_num(rgba_tex); - }; - live_frame.ready_fence = fence; - live_frame.input_frames = {}; - live_frame.temp_textures = { rgba_tex }; - output_channel[OUTPUT_LIVE].output_frame(live_frame); - - // Set up preview and any additional channels. - for (int i = 1; i < theme->get_num_channels() + 2; ++i) { - DisplayFrame display_frame; - Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions - display_frame.chain = chain.chain; - display_frame.setup_chain = chain.setup_chain; - display_frame.ready_fence = fence; - display_frame.input_frames = chain.input_frames; - display_frame.temp_textures = {}; - output_channel[i].output_frame(display_frame); - } - - clock_gettime(CLOCK_MONOTONIC, &now); - double elapsed = now.tv_sec - start.tv_sec + - 1e-9 * (now.tv_nsec - start.tv_nsec); - if (frame % 100 == 0) { - printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n", - frame, stats_dropped_frames, elapsed, frame / elapsed, - 1e3 * elapsed / frame); - // chain->print_phase_timing(); + int64_t frame_duration = output_frame_info.frame_duration; + render_one_frame(frame_duration); + { + lock_guard lock(frame_num_mutex); + ++frame_num; } + frame_num_updated.notify_all(); + pts_int += frame_duration; + + basic_stats.update(frame_num, stats_dropped_frames); + // if (frame_num % 100 == 0) chain->print_phase_timing(); 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->shutdown(); - httpd.close_output_file(); - httpd.open_output_file(filename.c_str()); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd)); + video_encoder->do_cut(frame_num); } #if 0 @@ -669,180 +1097,429 @@ void Mixer::thread_func() resource_pool->clean_context(); } -void Mixer::audio_thread_func() +bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const { - while (!should_quit) { - AudioTask task; + if (output_card_index != -1) { + // The output card (ie., cards[output_card_index].output) is the master clock, + // so no input card (ie., cards[card_index].capture) is. + return false; + } + return (card_index == master_card_index); +} - { - unique_lock lock(audio_mutex); - audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); }); - task = audio_task_queue.front(); - audio_task_queue.pop(); +void Mixer::trim_queue(CaptureCard *card, size_t safe_queue_length) +{ + // Count the number of frames in the queue, including any frames + // we dropped. It's hard to know exactly how we should deal with + // dropped (corrupted) input frames; they don't help our goal of + // avoiding starvation, but they still add to the problem of latency. + // Since dropped frames is going to mean a bump in the signal anyway, + // we err on the side of having more stable latency instead. + unsigned queue_length = 0; + for (const CaptureCard::NewFrame &frame : card->new_frames) { + queue_length += frame.dropped_frames + 1; + } + + // If needed, drop frames until the queue is below the safe limit. + // We prefer to drop from the head, because all else being equal, + // we'd like more recent frames (less latency). + unsigned dropped_frames = 0; + while (queue_length > safe_queue_length) { + assert(!card->new_frames.empty()); + assert(queue_length > card->new_frames.front().dropped_frames); + queue_length -= card->new_frames.front().dropped_frames; + + if (queue_length <= safe_queue_length) { + // No need to drop anything. + break; } - process_audio_one_frame(task.pts_int, task.num_samples); + card->new_frames.pop_front(); + card->new_frames_changed.notify_all(); + --queue_length; + ++dropped_frames; + + if (queue_length == 0 && card->is_cef_capture) { + card->may_have_dropped_last_frame = true; + } } + + card->metric_input_dropped_frames_jitter += dropped_frames; + card->metric_input_queue_length_frames = queue_length; + +#if 0 + if (dropped_frames > 0) { + fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n", + card_index, dropped_frames); + } +#endif } -void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) +pair Mixer::get_channels_json() { - vector samples_card; - vector samples_out; - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - 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); + Channels ret; + for (int channel_idx = 2; channel_idx < theme->get_num_channels(); ++channel_idx) { + Channel *channel = ret.add_channel(); + channel->set_index(channel_idx); + channel->set_name(theme->get_channel_name(channel_idx)); + channel->set_color(theme->get_channel_color(channel_idx)); + } + string contents; + google::protobuf::util::MessageToJsonString(ret, &contents); // Ignore any errors. + return make_pair(contents, "text/json"); +} + +pair Mixer::get_channel_color_http(unsigned channel_idx) +{ + return make_pair(theme->get_channel_color(channel_idx), "text/plain"); +} + +Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS]) +{ + OutputFrameInfo output_frame_info; +start: + unique_lock lock(card_mutex, defer_lock); + if (master_card_is_output) { + // Clocked to the output, so wait for it to be ready for the next frame. + cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp); + lock.lock(); + } else { + // Wait for the master card to have a new frame. + // TODO: Add a timeout. + output_frame_info.is_preroll = false; + lock.lock(); + cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); }); + } + + if (master_card_is_output) { + handle_hotplugged_cards(); + } else if (cards[master_card_index].new_frames.empty()) { + // We were woken up, but not due to a new frame. Deal with it + // and then restart. + assert(cards[master_card_index].capture->get_disconnected()); + handle_hotplugged_cards(); + lock.unlock(); + goto start; + } + + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + CaptureCard *card = &cards[card_index]; + if (card->new_frames.empty()) { // Starvation. + ++card->metric_input_duped_frames; +#ifdef HAVE_CEF + if (card->is_cef_capture && card->may_have_dropped_last_frame) { + // Unlike other sources, CEF is not guaranteed to send us a steady + // stream of frames, so we'll have to ask it to repaint the frame + // we dropped. (may_have_dropped_last_frame is set whenever we + // trim the queue completely away, and cleared when we actually + // get a new frame.) + ((CEFCapture *)card->capture.get())->request_new_frame(); } - } - // TODO: Allow using audio from the other card(s) as well. - if (card_index == 0) { - samples_out = move(samples_card); +#endif + } else { + new_frames[card_index] = move(card->new_frames.front()); + has_new_frame[card_index] = true; + card->new_frames.pop_front(); + card->new_frames_changed.notify_all(); } } - // Cut away everything under 120 Hz (or whatever the cutoff is); - // we don't need it for voice, and it will reduce headroom - // and confuse the compressor. (In particular, any hums at 50 or 60 Hz - // should be dampened.) - locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f); + if (!master_card_is_output) { + output_frame_info.frame_timestamp = new_frames[master_card_index].received_timestamp; + output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames; + output_frame_info.frame_duration = new_frames[master_card_index].length; + } - // Apply a level compressor to get the general level right. - // Basically, if it's over about -40 dBFS, we squeeze it down to that level - // (or more precisely, near it, since we don't use infinite ratio), - // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course, - // entirely arbitrary, but from practical tests with speech, it seems to - // put ut around -23 LUFS, so it's a reasonable starting point for later use. - if (level_compressor_enabled) { - float threshold = 0.01f; // -40 dBFS. - float ratio = 20.0f; - float attack_time = 0.5f; - float release_time = 20.0f; - float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB. - level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain); - last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain); + if (!output_frame_info.is_preroll) { + output_jitter_history.frame_arrived(output_frame_info.frame_timestamp, output_frame_info.frame_duration, output_frame_info.dropped_frames); } -#if 0 - printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n", - level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()), - level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()), - 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain)); -#endif + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs + num_html_inputs; ++card_index) { + CaptureCard *card = &cards[card_index]; + if (has_new_frame[card_index] && + !input_card_is_master_clock(card_index, master_card_index) && + !output_frame_info.is_preroll) { + card->queue_length_policy.update_policy( + output_frame_info.frame_timestamp, + card->jitter_history.get_expected_next_frame(), + new_frames[master_card_index].length, + output_frame_info.frame_duration, + card->jitter_history.estimate_max_jitter(), + output_jitter_history.estimate_max_jitter()); + trim_queue(card, min(global_flags.max_input_queue_frames, + card->queue_length_policy.get_safe_queue_length())); + } + } + + // This might get off by a fractional sample when changing master card + // between ones with different frame rates, but that's fine. + int num_samples_times_timebase = OUTPUT_FREQUENCY * output_frame_info.frame_duration + fractional_samples; + output_frame_info.num_samples = num_samples_times_timebase / TIMEBASE; + fractional_samples = num_samples_times_timebase % TIMEBASE; + assert(output_frame_info.num_samples >= 0); -// float limiter_att, compressor_att; - - // The real compressor. - if (compressor_enabled) { - float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f); - float ratio = 20.0f; - float attack_time = 0.005f; - float release_time = 0.040f; - float makeup_gain = 2.0f; // +6 dB. - compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain); -// compressor_att = compressor.get_attenuation(); - } - - // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only. - // Note that since ratio is not infinite, we could go slightly higher than this. - if (limiter_enabled) { - float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f); - float ratio = 30.0f; - float attack_time = 0.0f; // Instant. - float release_time = 0.020f; - float makeup_gain = 1.0f; // 0 dB. - limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain); -// limiter_att = limiter.get_attenuation(); - } - -// printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att)); - - // Upsample 4x to find interpolated peak. - peak_resampler.inp_data = samples_out.data(); - peak_resampler.inp_count = samples_out.size() / 2; - - vector interpolated_samples_out; - interpolated_samples_out.resize(samples_out.size()); - while (peak_resampler.inp_count > 0) { // About four iterations. - peak_resampler.out_data = &interpolated_samples_out[0]; - peak_resampler.out_count = interpolated_samples_out.size() / 2; - peak_resampler.process(); - size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count; - peak = max(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2)); - } - - // Find R128 levels. - vector left, right; - deinterleave_samples(samples_out, &left, &right); - float *ptrs[] = { left.data(), right.data() }; + return output_frame_info; +} + +void Mixer::handle_hotplugged_cards() +{ + // Check for cards that have been disconnected since last frame. + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + CaptureCard *card = &cards[card_index]; + if (card->capture->get_disconnected()) { + fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index); + FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio); + configure_card(card_index, capture, CardType::FAKE_CAPTURE, /*output=*/nullptr); + card->queue_length_policy.reset(card_index); + card->capture->start_bm_capture(); + } + } + + // Check for cards that have been connected since last frame. + vector hotplugged_cards_copy; { - unique_lock lock(r128_mutex); - r128.process(left.size(), ptrs); + lock_guard lock(hotplug_mutex); + swap(hotplugged_cards, hotplugged_cards_copy); } + for (libusb_device *new_dev : hotplugged_cards_copy) { + // Look for a fake capture card where we can stick this in. + int free_card_index = -1; + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + if (cards[card_index].is_fake_capture) { + free_card_index = card_index; + break; + } + } - // Send the samples to the sound card. - if (alsa) { - alsa->write(samples_out); + if (free_card_index == -1) { + fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n"); + libusb_unref_device(new_dev); + } else { + // BMUSBCapture takes ownership. + fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index); + CaptureCard *card = &cards[free_card_index]; + BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev); + configure_card(free_card_index, capture, CardType::LIVE_CARD, /*output=*/nullptr); + card->queue_length_policy.reset(free_card_index); + capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index)); + capture->start_bm_capture(); + } } +} + - // And finally add them to the output. - h264_encoder->add_audio(frame_pts_int, move(samples_out)); +void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp) +{ + // Resample the audio as needed, including from previously dropped frames. + assert(num_cards > 0); + for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) { + const bool dropped_frame = (frame_num != dropped_frames); + { + // Signal to the audio thread to process this frame. + // Note that if the frame is a dropped frame, we signal that + // we don't want to use this frame as base for adjusting + // the resampler rate. The reason for this is that the timing + // of these frames is often way too late; they typically don't + // “arrive” before we synthesize them. Thus, we could end up + // in a situation where we have inserted e.g. five audio frames + // into the queue before we then start pulling five of them + // back out. This makes ResamplingQueue overestimate the delay, + // causing undue resampler changes. (We _do_ use the last, + // non-dropped frame; perhaps we should just discard that as well, + // since dropped frames are expected to be rare, and it might be + // better to just wait until we have a slightly more normal situation). + unique_lock lock(audio_mutex); + bool adjust_rate = !dropped_frame && !is_preroll; + audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp}); + audio_task_queue_changed.notify_one(); + } + if (dropped_frame) { + // For dropped frames, increase the pts. Note that if the format changed + // in the meantime, we have no way of detecting that; we just have to + // assume the frame length is always the same. + pts_int += length_per_frame; + } + } } -void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) +void Mixer::render_one_frame(int64_t duration) { - GLuint vao; - glGenVertexArrays(1, &vao); - check_error(); + // Determine the time code for this frame before we start rendering. + string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num); + if (display_timecode_on_stdout) { + printf("Timecode: '%s'\n", timecode_text.c_str()); + } - float vertices[] = { - 0.0f, 2.0f, - 0.0f, 0.0f, - 2.0f, 0.0f - }; + // Update Y'CbCr settings for all cards. + { + unique_lock lock(card_mutex); + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + YCbCrInterpretation *interpretation = &ycbcr_interpretation[card_index]; + input_state.ycbcr_coefficients_auto[card_index] = interpretation->ycbcr_coefficients_auto; + input_state.ycbcr_coefficients[card_index] = interpretation->ycbcr_coefficients; + input_state.full_range[card_index] = interpretation->full_range; + } + } - glBindVertexArray(vao); - check_error(); + // Get the main chain from the theme, and set its state immediately. + Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state); + EffectChain *chain = theme_main_chain.chain; + theme_main_chain.setup_chain(); + //theme_main_chain.chain->enable_phase_timing(true); + + // If HDMI/SDI output is active and the user has requested auto mode, + // its mode overrides the existing Y'CbCr setting for the chain. + YCbCrLumaCoefficients ycbcr_output_coefficients; + if (global_flags.ycbcr_auto_coefficients && output_card_index != -1) { + ycbcr_output_coefficients = cards[output_card_index].output->preferred_ycbcr_coefficients(); + } else { + ycbcr_output_coefficients = global_flags.ycbcr_rec709_coefficients ? YCBCR_REC_709 : YCBCR_REC_601; + } - // Extract Cb/Cr. - GLuint fbo = resource_pool->create_fbo(dst_tex); - glBindFramebuffer(GL_FRAMEBUFFER, fbo); - glViewport(0, 0, WIDTH/2, HEIGHT/2); - check_error(); + // TODO: Reduce the duplication against theme.cpp. + YCbCrFormat output_ycbcr_format; + output_ycbcr_format.chroma_subsampling_x = 1; + output_ycbcr_format.chroma_subsampling_y = 1; + output_ycbcr_format.luma_coefficients = ycbcr_output_coefficients; + output_ycbcr_format.full_range = false; + output_ycbcr_format.num_levels = 1 << global_flags.x264_bit_depth; + chain->change_ycbcr_output_format(output_ycbcr_format); + + // Render main chain. If we're using zerocopy Quick Sync encoding + // (the default case), we take an extra copy of the created outputs, + // so that we can display it back to the screen later (it's less memory + // bandwidth than writing and reading back an RGBA texture, even at 16-bit). + // Ideally, we'd like to avoid taking copies and just use the main textures + // for display as well, but they're just views into VA-API memory and must be + // unmapped during encoding, so we can't use them for display, unfortunately. + GLuint y_tex, cbcr_full_tex, cbcr_tex; + GLuint y_copy_tex, cbcr_copy_tex = 0; + GLuint y_display_tex, cbcr_display_tex; + GLenum y_type = (global_flags.x264_bit_depth > 8) ? GL_R16 : GL_R8; + GLenum cbcr_type = (global_flags.x264_bit_depth > 8) ? GL_RG16 : GL_RG8; + const bool is_zerocopy = video_encoder->is_zerocopy(); + if (is_zerocopy) { + cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height); + y_copy_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height); + cbcr_copy_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2); + + y_display_tex = y_copy_tex; + cbcr_display_tex = cbcr_copy_tex; + + // y_tex and cbcr_tex will be given by VideoEncoder. + } else { + cbcr_full_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width, global_flags.height); + y_tex = resource_pool->create_2d_texture(y_type, global_flags.width, global_flags.height); + cbcr_tex = resource_pool->create_2d_texture(cbcr_type, global_flags.width / 2, global_flags.height / 2); - glUseProgram(cbcr_program_num); - check_error(); + y_display_tex = y_tex; + cbcr_display_tex = cbcr_tex; + } - glActiveTexture(GL_TEXTURE0); - check_error(); - glBindTexture(GL_TEXTURE_2D, src_tex); + const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue. + bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, ycbcr_output_coefficients, theme_main_chain.input_frames, &y_tex, &cbcr_tex); + assert(got_frame); + + GLuint fbo; + if (is_zerocopy) { + fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, y_copy_tex); + } else { + fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex); + } check_error(); + chain->render_to_fbo(fbo, global_flags.width, global_flags.height); + + if (display_timecode_in_stream) { + // Render the timecode on top. + timecode_renderer->render_timecode(fbo, timecode_text); + } + + resource_pool->release_fbo(fbo); + + if (is_zerocopy) { + chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex, cbcr_copy_tex); + } else { + chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex); + } + if (output_card_index != -1) { + cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, ycbcr_output_coefficients, theme_main_chain.input_frames, pts_int, duration); + } + resource_pool->release_2d_texture(cbcr_full_tex); + + // Set the right state for the Y' and CbCr textures we use for display. + glBindFramebuffer(GL_FRAMEBUFFER, 0); + glBindTexture(GL_TEXTURE_2D, y_display_tex); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - check_error(); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - check_error(); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - check_error(); - float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f }; - set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0); + glBindTexture(GL_TEXTURE_2D, cbcr_display_tex); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - GLuint position_vbo = fill_vertex_attribute(cbcr_program_num, "position", 2, GL_FLOAT, sizeof(vertices), vertices); - GLuint texcoord_vbo = fill_vertex_attribute(cbcr_program_num, "texcoord", 2, GL_FLOAT, sizeof(vertices), vertices); // Same as vertices. + RefCountedGLsync fence = video_encoder->end_frame(); - glDrawArrays(GL_TRIANGLES, 0, 3); - check_error(); + // The live frame pieces the Y'CbCr texture copies back into RGB and displays them. + // It owns y_display_tex and cbcr_display_tex now (whichever textures they are). + DisplayFrame live_frame; + live_frame.chain = display_chain.get(); + live_frame.setup_chain = [this, y_display_tex, cbcr_display_tex]{ + display_input->set_texture_num(0, y_display_tex); + display_input->set_texture_num(1, cbcr_display_tex); + }; + live_frame.ready_fence = fence; + live_frame.input_frames = {}; + live_frame.temp_textures = { y_display_tex, cbcr_display_tex }; + output_channel[OUTPUT_LIVE].output_frame(move(live_frame)); + + // Set up preview and any additional channels. + for (int i = 1; i < theme->get_num_channels() + 2; ++i) { + DisplayFrame display_frame; + Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions + display_frame.chain = move(chain.chain); + display_frame.setup_chain = move(chain.setup_chain); + display_frame.ready_fence = fence; + display_frame.input_frames = move(chain.input_frames); + display_frame.temp_textures = {}; + output_channel[i].output_frame(move(display_frame)); + } +} - cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo); - cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo); +void Mixer::audio_thread_func() +{ + pthread_setname_np(pthread_self(), "Mixer_Audio"); - glUseProgram(0); - check_error(); + while (!should_quit) { + AudioTask task; - resource_pool->release_fbo(fbo); - glDeleteVertexArrays(1, &vao); + { + unique_lock lock(audio_mutex); + 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(); + } + + ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy = + task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE; + vector samples_out = audio_mixer->get_output( + task.frame_timestamp, + task.num_samples, + rate_adjustment_policy); + + // Send the samples to the sound card, then add them to the output. + if (alsa) { + alsa->write(samples_out); + } + if (output_card_index != -1) { + const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE); // Corresponds to the delay in ResamplingQueue. + cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out); + } + video_encoder->add_audio(task.pts_int, move(samples_out)); + } } void Mixer::release_display_frame(DisplayFrame *frame) @@ -864,6 +1541,7 @@ void Mixer::start() void Mixer::quit() { should_quit = true; + audio_task_queue_changed.notify_one(); mixer_thread.join(); audio_thread.join(); } @@ -878,12 +1556,59 @@ void Mixer::channel_clicked(int preview_num) theme->channel_clicked(preview_num); } -void Mixer::reset_meters() +YCbCrInterpretation Mixer::get_input_ycbcr_interpretation(unsigned card_index) const +{ + unique_lock lock(card_mutex); + return ycbcr_interpretation[card_index]; +} + +void Mixer::set_input_ycbcr_interpretation(unsigned card_index, const YCbCrInterpretation &interpretation) +{ + unique_lock lock(card_mutex); + ycbcr_interpretation[card_index] = interpretation; +} + +void Mixer::start_mode_scanning(unsigned card_index) +{ + assert(card_index < num_cards); + if (is_mode_scanning[card_index]) { + return; + } + is_mode_scanning[card_index] = true; + mode_scanlist[card_index].clear(); + for (const auto &mode : cards[card_index].capture->get_available_video_modes()) { + mode_scanlist[card_index].push_back(mode.first); + } + assert(!mode_scanlist[card_index].empty()); + mode_scanlist_index[card_index] = 0; + cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]); + last_mode_scan_change[card_index] = steady_clock::now(); +} + +map Mixer::get_available_output_video_modes() const +{ + assert(desired_output_card_index != -1); + unique_lock lock(card_mutex); + return cards[desired_output_card_index].output->get_available_video_modes(); +} + +string Mixer::get_ffmpeg_filename(unsigned card_index) const +{ + assert(card_index >= num_cards && card_index < num_cards + num_video_inputs); + return ((FFmpegCapture *)(cards[card_index].capture.get()))->get_filename(); +} + +void Mixer::set_ffmpeg_filename(unsigned card_index, const string &filename) { + assert(card_index >= num_cards && card_index < num_cards + num_video_inputs); + ((FFmpegCapture *)(cards[card_index].capture.get()))->change_filename(filename); +} + +void Mixer::wait_for_next_frame() { - peak_resampler.reset(); - peak = 0.0f; - r128.reset(); - r128.integr_start(); + unique_lock lock(frame_num_mutex); + unsigned old_frame_num = frame_num; + frame_num_updated.wait_for(lock, seconds(1), // Timeout is just in case. + [old_frame_num, this]{ return this->frame_num > old_frame_num; }); } Mixer::OutputChannel::~OutputChannel() @@ -896,7 +1621,7 @@ Mixer::OutputChannel::~OutputChannel() } } -void Mixer::OutputChannel::output_frame(DisplayFrame frame) +void Mixer::OutputChannel::output_frame(DisplayFrame &&frame) { // Store this frame for display. Remove the ready frame if any // (it was seemingly never used). @@ -905,12 +1630,52 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) if (has_ready_frame) { parent->release_display_frame(&ready_frame); } - ready_frame = frame; + ready_frame = move(frame); has_ready_frame = true; + + // Call the callbacks under the mutex (they should be short), + // so that we don't race against a callback removal. + for (const auto &key_and_callback : new_frame_ready_callbacks) { + key_and_callback.second(); + } } - if (has_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; + } } } @@ -928,7 +1693,7 @@ bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) } if (has_ready_frame) { assert(!has_current_frame); - current_frame = ready_frame; + current_frame = move(ready_frame); ready_frame.ready_fence.reset(); // Drop the refcount. ready_frame.input_frames.clear(); // Drop the refcounts. has_current_frame = true; @@ -939,8 +1704,31 @@ bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) return true; } -void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback) +void Mixer::OutputChannel::add_frame_ready_callback(void *key, Mixer::new_frame_ready_callback_t callback) { - new_frame_ready_callback = callback; - has_new_frame_ready_callback = true; + unique_lock lock(frame_mutex); + new_frame_ready_callbacks[key] = callback; } + +void Mixer::OutputChannel::remove_frame_ready_callback(void *key) +{ + unique_lock lock(frame_mutex); + new_frame_ready_callbacks.erase(key); +} + +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;