X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=6846d83ffdbaef73420d638cc1b282ce11cb02a3;hb=96cb6414f85e0ef4d660b7bd56267303e80fcd05;hp=ce3ceb4bb5b8e58158eb468f5ea16e77ce40823b;hpb=884f2c044c4ab008ce9295edd343ee851eba4576;p=nageru diff --git a/mixer.cpp b/mixer.cpp index ce3ceb4..6846d83 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -10,77 +10,61 @@ #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 "bmusb/bmusb.h" +#include "bmusb/fake_capture.h" +#include "chroma_subsampler.h" #include "context.h" #include "decklink_capture.h" +#include "decklink_output.h" #include "defs.h" -#include "fake_capture.h" +#include "disk_space_estimator.h" +#include "ffmpeg_capture.h" #include "flags.h" -#include "video_encoder.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" +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; bool uses_mlock = false; namespace { -void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples) -{ - assert(in_channels >= out_channels); - 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 convert_fixed32_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples) -{ - assert(in_channels >= out_channels); - for (size_t i = 0; i < num_samples; ++i) { - for (size_t j = 0; j < out_channels; ++j) { - // Note: Assumes little-endian. - int32_t s = *(int32_t *)src; - dst[i * out_channels + j] = s * (1.0f / 4294967296.0f); - src += 4; - } - src += 4 * (in_channels - out_channels); - } -} - void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state) { if (interlaced) { @@ -96,31 +80,153 @@ void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced } } +void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned cbcr_width, unsigned cbcr_height, unsigned v210_width) +{ + 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); + } + + 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; + } + 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; + } +} + +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 QueueLengthPolicy::update_policy(int queue_length) +void QueueLengthPolicy::register_metrics(const vector> &labels) +{ + global_metrics.add("input_queue_length_frames", labels, &metric_input_queue_length_frames, Metrics::TYPE_GAUGE); + global_metrics.add("input_queue_safe_length_frames", labels, &metric_input_queue_safe_length_frames, Metrics::TYPE_GAUGE); + global_metrics.add("input_queue_duped_frames", labels, &metric_input_duped_frames); +} + +void QueueLengthPolicy::update_policy(unsigned queue_length) { - if (queue_length < 0) { // Starvation. - if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) { + if (queue_length == 0) { // Starvation. + if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) { ++safe_queue_length; - fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n", + fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n", card_index, safe_queue_length); } frames_with_at_least_one = 0; been_at_safe_point_since_last_starvation = false; + ++metric_input_duped_frames; + metric_input_queue_safe_length_frames = safe_queue_length; return; } - if (queue_length > 0) { - if (queue_length >= int(safe_queue_length)) { - been_at_safe_point_since_last_starvation = true; - } - if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) { - --safe_queue_length; - 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; - } - } else { + if (queue_length >= safe_queue_length) { + been_at_safe_point_since_last_starvation = true; + } + if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) { + --safe_queue_length; + metric_input_queue_safe_length_frames = safe_queue_length; + fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n", + card_index, safe_queue_length); frames_with_at_least_one = 0; } } @@ -130,20 +236,23 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) num_cards(num_cards), mixer_surface(create_surface(format)), h264_encoder_surface(create_surface(format)), - correlation(OUTPUT_FREQUENCY), - level_compressor(OUTPUT_FREQUENCY), - limiter(OUTPUT_FREQUENCY), - compressor(OUTPUT_FREQUENCY) + decklink_output_surface(create_surface(format)), + ycbcr_interpretation(global_flags.ycbcr_interpretation), + audio_mixer(num_cards) { 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(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; @@ -153,38 +262,46 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) 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(); - video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, 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)); + + // 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)); // Start listening for clients only once VideoEncoder has written its header, if any. httpd.start(9095); - // First try initializing the fake devices, then PCI devices, then USB, - // until we have the desired number of cards. - unsigned num_pci_devices = 0, num_usb_devices = 0; + // 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; - assert(global_flags.num_fake_cards >= 0); // Enforced in flags.cpp. - unsigned num_fake_cards = global_flags.num_fake_cards; - - assert(num_fake_cards <= num_cards); // Enforced in flags.cpp. - for ( ; card_index < num_fake_cards; ++card_index) { - configure_card(card_index, format, new FakeCapture(card_index)); - } - - if (global_flags.num_fake_cards > 0) { - fprintf(stderr, "Initialized %d fake cards.\n", global_flags.num_fake_cards); - } - - if (card_index < num_cards) { + { IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance(); if (decklink_iterator != nullptr) { for ( ; card_index < num_cards; ++card_index) { @@ -193,130 +310,244 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) break; } - configure_card(card_index, format, new DeckLinkCapture(decklink, card_index - num_fake_cards)); + 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); + output->set_device(decklink); + configure_card(card_index, capture, CardType::LIVE_CARD, output); ++num_pci_devices; } decklink_iterator->Release(); - fprintf(stderr, "Found %d DeckLink PCI card(s).\n", num_pci_devices); + 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"); } } - for ( ; card_index < num_cards; ++card_index) { - configure_card(card_index, format, new BMUSBCapture(card_index - num_pci_devices - num_fake_cards)); - ++num_usb_devices; + + 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); - if (num_usb_devices > 0) { - has_bmusb_thread = true; - BMUSBCapture::start_bm_thread(); + 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); } - for (card_index = 0; card_index < num_cards; ++card_index) { - cards[card_index].queue_length_policy.reset(card_index); - cards[card_index].capture->start_bm_capture(); - } - - // Set up stuff for NV12 conversion. - - // Cb/Cr shader. - string cbcr_vert_shader = - "#version 130 \n" - " \n" - "in vec2 position; \n" - "in vec2 texcoord; \n" - "out vec2 tc0; \n" - "uniform vec2 foo_chroma_offset_0; \n" - " \n" - "void main() \n" - "{ \n" - " // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n" - " // \n" - " // 2.000 0.000 0.000 -1.000 \n" - " // 0.000 2.000 0.000 -1.000 \n" - " // 0.000 0.000 -2.000 -1.000 \n" - " // 0.000 0.000 0.000 1.000 \n" - " gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n" - " vec2 flipped_tc = texcoord; \n" - " tc0 = flipped_tc + foo_chroma_offset_0; \n" - "} \n"; - string cbcr_frag_shader = - "#version 130 \n" - "in vec2 tc0; \n" - "uniform sampler2D cbcr_tex; \n" - "out vec4 FragColor; \n" - "void main() { \n" - " FragColor = texture(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); - - float vertices[] = { - 0.0f, 2.0f, - 0.0f, 0.0f, - 2.0f, 0.0f - }; - cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices); - cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position"); - cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord"); + if (num_fake_cards > 0) { + fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards); + } - r128.init(2, OUTPUT_FREQUENCY); - r128.integr_start(); + // Initialize all video inputs the theme asked for. Note that these are + // all put _after_ the regular cards, which stop at - 1. + std::vector video_inputs = theme->get_video_inputs(); + 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(); - locut.init(FILTER_HPF, 2); + BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1)); + BMUSBCapture::start_bm_thread(); - // If --flat-audio is given, turn off everything that messes with the sound, - // except the final makeup gain. - if (global_flags.flat_audio) { - set_locut_enabled(false); - set_gain_staging_auto(false); - set_limiter_enabled(false); - set_compressor_enabled(false); + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) { + cards[card_index].queue_length_policy.reset(card_index); } - // 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, /*frel=*/1.0); + 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); + } + } + + 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; 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); + } + + metric_start_time_seconds = get_timestamp_for_metrics(); + + global_metrics.add("frames_output_total", &metric_frames_output_total); + global_metrics.add("frames_output_dropped", &metric_frames_output_dropped); + global_metrics.add("start_time_seconds", &metric_start_time_seconds, Metrics::TYPE_GAUGE); + global_metrics.add("memory_used_bytes", &metrics_memory_used_bytes); + global_metrics.add("metrics_memory_locked_limit_bytes", &metrics_memory_locked_limit_bytes); } Mixer::~Mixer() { - resource_pool->release_glsl_program(cbcr_program_num); - glDeleteBuffers(1, &cbcr_vbo); - if (has_bmusb_thread) { - BMUSBCapture::stop_bm_thread(); - } + 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; ++card_index) { { - unique_lock lock(bmusb_mutex); + unique_lock lock(card_mutex); cards[card_index].should_quit = true; // Unblock thread. 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(); + } } video_encoder.reset(nullptr); } -void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture) +void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, CardType card_type, DeckLinkOutput *output) { printf("Configuring card %d...\n", card_index); CaptureCard *card = &cards[card_index]; - card->capture = capture; + if (card->capture != nullptr) { + card->capture->stop_dequeue_thread(); + } + card->capture.reset(capture); + card->is_fake_capture = (card_type == CardType::FAKE_CAPTURE); + 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 (global_flags.ten_bit_input) { + pixel_format = PixelFormat_10BitYCbCr; + } else { + pixel_format = PixelFormat_8BitYCbCr; + } + card->capture->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. + 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()); - card->surface = create_surface(format); - card->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + 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{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(); + + // 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; + default: + assert(false); + } + card->queue_length_policy.register_metrics(labels); + 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_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); } +void Mixer::set_output_card_internal(int card_index) +{ + // 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); + old_card->is_fake_capture = false; + old_card->capture->start_bm_capture(); + } + 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; +} namespace { @@ -330,48 +561,32 @@ int unwrap_timecode(uint16_t current_wrapped, int last) } } -float find_peak(const float *samples, size_t num_samples) -{ - float m = fabs(samples[0]); - for (size_t i = 1; i < num_samples; ++i) { - m = max(m, fabs(samples[i])); - } - return m; -} - -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); - - 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++; - } -} - } // namespace void Mixer::bm_frame(unsigned card_index, uint16_t timecode, FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format, FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format) { + DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index}; CaptureCard *card = &cards[card_index]; + 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 = 0.5; // Should be enough time for the signal to stabilize. - timespec now; - clock_gettime(CLOCK_MONOTONIC, &now); - double sec_since_last_switch = (now.tv_sec - last_mode_scan_change[card_index].tv_sec) + - 1e-9 * (now.tv_nsec - last_mode_scan_change[card_index].tv_nsec); + 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]++; @@ -399,72 +614,64 @@ 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); - switch (audio_format.bits_per_sample) { - case 0: - assert(num_samples == 0); - break; - case 24: - convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples); - break; - case 32: - convert_fixed32_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, audio_format.num_channels, num_samples); - break; - default: - fprintf(stderr, "Cannot handle audio with %u bits per sample\n", audio_format.bits_per_sample); - assert(false); + // 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, "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); + audio_mixer.reset_resampler(device); + dropped_frames = 0; + ++card->metric_input_resets; + } 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); + 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); } - // 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 * video_format.frame_rate_den / video_format.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(card_index, 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; - } - - card->last_timecode = timecode; + 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); } - size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2; + card->last_timecode = timecode; + + PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; + + 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) { @@ -478,31 +685,35 @@ 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); + 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; - card->new_frames.push(move(new_frame)); + new_frame.received_timestamp = video_frame.received_timestamp; + card->new_frames.push_back(move(new_frame)); card->new_frames_changed.notify_all(); } return; } - PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; - unsigned num_fields = video_format.interlaced ? 2 : 1; - timespec frame_upload_start; + 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(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 = video_format.interlaced; userdata->last_has_signal = video_format.has_signal; @@ -512,10 +723,6 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, RefCountedFrame frame(video_frame); // Upload the textures. - size_t cbcr_width = video_format.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) { // 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 @@ -525,53 +732,65 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, // 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; + 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; } - GLuint pbo = userdata->pbo; - check_error(); - glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo); - check_error(); + // 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); - 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); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo); + check_error(); - if (global_flags.flush_pbos) { - glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height); + 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(); - glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t)); + 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, 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(); }; @@ -583,20 +802,13 @@ 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); + unique_lock lock(card_mutex); CaptureCard::NewFrame new_frame; new_frame.frame = frame; new_frame.length = frame_length; @@ -604,14 +816,28 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, new_frame.interlaced = video_format.interlaced; new_frame.upload_func = upload_func; new_frame.dropped_frames = dropped_frames; - card->new_frames.push(move(new_frame)); + new_frame.received_timestamp = video_frame.received_timestamp; // Ignore the audio timestamp. + card->new_frames.push_back(move(new_frame)); 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)) { @@ -619,27 +845,52 @@ 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; ++card_index) { + if (int(card_index) != output_card_index) { + cards[card_index].capture->start_bm_capture(); + } + } + + steady_clock::time_point start, now; + start = steady_clock::now(); - int frame = 0; int stats_dropped_frames = 0; while (!should_quit) { - CaptureCard::NewFrame new_frames[MAX_CARDS]; - bool has_new_frame[MAX_CARDS] = { false }; - int num_samples[MAX_CARDS] = { 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); + } - // TODO: Add a timeout. - unsigned master_card_index = theme->map_signal(master_clock_channel); - assert(master_card_index < num_cards); + CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS]; + bool has_new_frame[MAX_VIDEO_CARDS] = { false }; - get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples); - schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length); - stats_dropped_frames += new_frames[master_card_index].dropped_frames; - send_audio_level_callback(); + 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); + } - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + 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; + + handle_hotplugged_cards(); + + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) { if (card_index == master_card_index || !has_new_frame[card_index]) { continue; } @@ -654,13 +905,13 @@ void Mixer::thread_func() // 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 (new_frames[master_card_index].frame->len == 0) { + if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) { ++stats_dropped_frames; pts_int += new_frames[master_card_index].length; continue; } - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) { if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0) continue; @@ -676,61 +927,62 @@ void Mixer::thread_func() } } - int64_t duration = new_frames[master_card_index].length; - render_one_frame(duration); - ++frame; - pts_int += duration; - - 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) { - // check our memory usage, to see if we are close to our mlockall() - // limit (if at all set). - rusage used; - if (getrusage(RUSAGE_SELF, &used) == -1) { - perror("getrusage(RUSAGE_SELF)"); - assert(false); - } + int64_t frame_duration = output_frame_info.frame_duration; + render_one_frame(frame_duration); + ++frame_num; + pts_int += frame_duration; - rlimit limit; - if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) { - perror("getrlimit(RLIMIT_MEMLOCK)"); - assert(false); - } + now = steady_clock::now(); + double elapsed = duration(now - start).count(); + + metric_frames_output_total = frame_num; + metric_frames_output_dropped = stats_dropped_frames; + if (frame_num % 100 == 0) { printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)", - frame, stats_dropped_frames, elapsed, frame / elapsed, - 1e3 * elapsed / frame); + frame_num, stats_dropped_frames, elapsed, frame_num / elapsed, + 1e3 * elapsed / frame_num); // chain->print_phase_timing(); - if (uses_mlock) { - // Check our memory usage, to see if we are close to our mlockall() - // limit (if at all set). - rusage used; - if (getrusage(RUSAGE_SELF, &used) == -1) { - perror("getrusage(RUSAGE_SELF)"); - assert(false); - } + // Check our memory usage, to see if we are close to our mlockall() + // limit (if at all set). + rusage used; + if (getrusage(RUSAGE_SELF, &used) == -1) { + perror("getrusage(RUSAGE_SELF)"); + assert(false); + } + if (uses_mlock) { rlimit limit; if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) { perror("getrlimit(RLIMIT_MEMLOCK)"); assert(false); } - printf(", using %ld / %ld MB lockable memory (%.1f%%)", - long(used.ru_maxrss / 1024), - long(limit.rlim_cur / 1048576), - float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur)); + if (limit.rlim_cur == 0) { + printf(", using %ld MB memory (locked)", + long(used.ru_maxrss / 1024)); + } else { + printf(", using %ld / %ld MB lockable memory (%.1f%%)", + long(used.ru_maxrss / 1024), + long(limit.rlim_cur / 1048576), + float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur)); + } + metrics_memory_locked_limit_bytes = limit.rlim_cur; + } else { + printf(", using %ld MB memory (not locked)", + long(used.ru_maxrss / 1024)); + metrics_memory_locked_limit_bytes = 0.0 / 0.0; } printf("\n"); + + metrics_memory_used_bytes = used.ru_maxrss; } if (should_cut.exchange(false)) { // Test and clear. - video_encoder->do_cut(frame); + video_encoder->do_cut(frame_num); } #if 0 @@ -749,57 +1001,201 @@ void Mixer::thread_func() resource_pool->clean_context(); } -void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_CARDS], bool has_new_frame[MAX_CARDS], int num_samples[MAX_CARDS]) +bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const { - // The first card is the master timer, so wait for it to have a new frame. - unique_lock lock(bmusb_mutex); - cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty(); }); + 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); +} - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - CaptureCard *card = &cards[card_index]; - if (card->new_frames.empty()) { - assert(card_index != master_card_index); - card->queue_length_policy.update_policy(-1); - continue; +void Mixer::trim_queue(CaptureCard *card, unsigned card_index) +{ + // 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; + } + card->queue_length_policy.update_policy(queue_length); + + // 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 > card->queue_length_policy.get_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 <= card->queue_length_policy.get_safe_queue_length()) { + // No need to drop anything. + break; } - new_frames[card_index] = move(card->new_frames.front()); - has_new_frame[card_index] = true; - card->new_frames.pop(); + + card->new_frames.pop_front(); card->new_frames_changed.notify_all(); + --queue_length; + ++dropped_frames; + } + + metric_input_queue_length_frames = queue_length; + card->metric_input_dropped_frames_jitter += dropped_frames; + +#if 0 + if (dropped_frames > 0) { + fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n", + card_index, dropped_frames); + } +#endif +} - int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].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 (card_index == master_card_index) { +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(); + goto start; + } + + if (!master_card_is_output) { + output_frame_info.frame_timestamp = + cards[master_card_index].new_frames.front().received_timestamp; + } + + for (unsigned card_index = 0; card_index < num_cards + num_video_inputs; ++card_index) { + CaptureCard *card = &cards[card_index]; + if (input_card_is_master_clock(card_index, master_card_index)) { // We don't use the queue length policy for the master card, // but we will if it stops being the master. Thus, clear out // the policy in case we switch in the future. card->queue_length_policy.reset(card_index); + assert(!card->new_frames.empty()); } else { - // If we have excess frames compared to the policy for this card, - // drop frames from the head. - card->queue_length_policy.update_policy(card->new_frames.size()); - while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) { - card->new_frames.pop(); + trim_queue(card, card_index); + } + if (!card->new_frames.empty()) { + 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(); + } + } + + if (!master_card_is_output) { + output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames; + output_frame_info.frame_duration = new_frames[master_card_index].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); + + 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; + { + 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; } } + + 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(); + } } } -void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame) + +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); - audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame}); + 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 (frame_num != dropped_frames) { + 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. @@ -810,53 +1206,145 @@ void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_sam void Mixer::render_one_frame(int64_t duration) { + // 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()); + } + + // 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; + } + } + // 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); + 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); - GLuint y_tex, cbcr_tex; - bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex); + // The theme can't (or at least shouldn't!) call connect_signal() on + // each FFmpeg input, so we'll do it here. + for (const pair &conn : theme->get_signal_connections()) { + conn.first->connect_signal_raw(conn.second->get_card_index()); + } + + // 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; + } + + // 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); + + y_display_tex = y_tex; + cbcr_display_tex = cbcr_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); - // 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); + 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, WIDTH, HEIGHT); + 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); - subsample_chroma(cbcr_full_tex, cbcr_tex); + 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 rgba_tex. + // Set the right state for the Y' and CbCr textures we use for display. glBindFramebuffer(GL_FRAMEBUFFER, 0); - glBindTexture(GL_TEXTURE_2D, rgba_tex); + glBindTexture(GL_TEXTURE_2D, y_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); - const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded. - RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames); + 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); - // The live frame just shows the RGBA texture we just rendered. - // It owns rgba_tex now. + RefCountedGLsync fence = video_encoder->end_frame(); + + // 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, rgba_tex]{ - display_input->set_texture_num(rgba_tex); + 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 = { rgba_tex }; + live_frame.temp_textures = { y_display_tex, cbcr_display_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 + Theme::Chain chain = theme->get_chain(i, pts(), global_flags.width, global_flags.height, input_state); // FIXME: dimensions display_frame.chain = chain.chain; display_frame.setup_chain = chain.setup_chain; display_frame.ready_fence = fence; @@ -866,26 +1354,10 @@ void Mixer::render_one_frame(int64_t duration) } } -void Mixer::send_audio_level_callback() -{ - if (audio_level_callback == nullptr) { - return; - } - - unique_lock lock(compressor_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(); - - audio_level_callback(loudness_s, 20.0 * log10(peak), - loudness_i, loudness_range_low, loudness_range_high, - gain_staging_db, 20.0 * log10(final_makeup_gain), - correlation.get_correlation()); -} - void Mixer::audio_thread_func() { + pthread_setname_np(pthread_self(), "Mixer_Audio"); + while (!should_quit) { AudioTask task; @@ -899,228 +1371,23 @@ void Mixer::audio_thread_func() audio_task_queue.pop(); } - process_audio_one_frame(task.pts_int, task.num_samples); - } -} - -void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) -{ - vector samples_card; - vector samples_out; - - // TODO: Allow mixing audio from several sources. - unsigned selected_audio_card = theme->map_signal(audio_source_channel); - assert(selected_audio_card < num_cards); + 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); - 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); - cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples); + // Send the samples to the sound card, then add them to the output. + if (alsa) { + alsa->write(samples_out); } - if (card_index == selected_audio_card) { - samples_out = move(samples_card); + 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)); } - - // 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.) - if (locut_enabled) { - locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f); - } - - // 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. - { - unique_lock lock(compressor_mutex); - 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); - gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain); - } else { - // Just apply the gain we already had. - float g = pow(10.0f, gain_staging_db / 20.0f); - for (size_t i = 0; i < samples_out.size(); ++i) { - samples_out[i] *= g; - } - } - } - -#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 - -// 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)); - peak_resampler.out_data = nullptr; - } - - // At this point, we are most likely close to +0 LU, but all of our - // measurements have been on raw sample values, not R128 values. - // So we have a final makeup gain to get us to +0 LU; the gain - // adjustments required should be relatively small, and also, the - // offset shouldn't change much (only if the type of audio changes - // significantly). Thus, we shoot for updating this value basically - // “whenever we process buffers”, since the R128 calculation isn't exactly - // something we get out per-sample. - // - // Note that there's a feedback loop here, so we choose a very slow filter - // (half-time of 100 seconds). - double target_loudness_factor, alpha; - { - unique_lock lock(compressor_mutex); - double loudness_lu = r128.loudness_M() - ref_level_lufs; - double current_makeup_lu = 20.0f * log10(final_makeup_gain); - target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f); - - // If we're outside +/- 5 LU uncorrected, we don't count it as - // a normal signal (probably silence) and don't change the - // correction factor; just apply what we already have. - if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) { - alpha = 0.0; - } else { - // Formula adapted from - // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter. - const double half_time_s = 100.0; - const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY); - alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0); - } - - double m = final_makeup_gain; - for (size_t i = 0; i < samples_out.size(); i += 2) { - samples_out[i + 0] *= m; - samples_out[i + 1] *= m; - m += (target_loudness_factor - m) * alpha; - } - final_makeup_gain = m; - } - - // Find R128 levels and L/R correlation. - vector left, right; - deinterleave_samples(samples_out, &left, &right); - float *ptrs[] = { left.data(), right.data() }; - { - unique_lock lock(compressor_mutex); - r128.process(left.size(), ptrs); - correlation.process_samples(samples_out); - } - - // Send the samples to the sound card. - if (alsa) { - alsa->write(samples_out); - } - - // And finally add them to the output. - video_encoder->add_audio(frame_pts_int, move(samples_out)); -} - -void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) -{ - GLuint vao; - glGenVertexArrays(1, &vao); - check_error(); - - glBindVertexArray(vao); - check_error(); - - // Extract Cb/Cr. - GLuint fbo = resource_pool->create_fbo(dst_tex); - glBindFramebuffer(GL_FRAMEBUFFER, fbo); - glViewport(0, 0, WIDTH/2, HEIGHT/2); - check_error(); - - glUseProgram(cbcr_program_num); - check_error(); - - glActiveTexture(GL_TEXTURE0); - check_error(); - glBindTexture(GL_TEXTURE_2D, src_tex); - check_error(); - 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); - - glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo); - check_error(); - - for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) { - glEnableVertexAttribArray(attr_index); - check_error(); - glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0)); - check_error(); - } - - glDrawArrays(GL_TRIANGLES, 0, 3); - check_error(); - - for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) { - glDisableVertexAttribArray(attr_index); - check_error(); - } - - glUseProgram(0); - check_error(); - glBindFramebuffer(GL_FRAMEBUFFER, 0); - check_error(); - - resource_pool->release_fbo(fbo); - glDeleteVertexArrays(1, &vao); } void Mixer::release_display_frame(DisplayFrame *frame) @@ -1157,13 +1424,16 @@ 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 { - peak_resampler.reset(); - peak = 0.0f; - r128.reset(); - r128.integr_start(); - correlation.reset(); + 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) @@ -1180,7 +1450,14 @@ void Mixer::start_mode_scanning(unsigned card_index) assert(!mode_scanlist[card_index].empty()); mode_scanlist_index[card_index] = 0; cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]); - clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]); + 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(); } Mixer::OutputChannel::~OutputChannel() @@ -1204,10 +1481,12 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) } ready_frame = frame; has_ready_frame = true; - } - if (new_frame_ready_callback) { - new_frame_ready_callback(); + // 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(); + } } // Reduce the number of callbacks by filtering duplicates. The reason @@ -1274,9 +1553,16 @@ 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; + 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)