X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=436b761d3efd433a2b6addf93c315c05faeaf35c;hb=e3bb32331aa8f0c63a43fb5434082134a6fbed12;hp=2eb037c4b6d2903ae60399e4c5cba95a34fb2aab;hpb=89744513e07c29e54a3d773734a111081f24a96c;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 2eb037c..436b761 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,23 +1,21 @@ -#define EXTRAHEIGHT 30 - #undef Success #include "mixer.h" #include #include -#include #include #include #include #include +#include #include +#include #include #include #include #include #include -#include #include #include #include @@ -28,10 +26,13 @@ #include #include #include +#include #include "bmusb/bmusb.h" #include "context.h" +#include "decklink_capture.h" #include "defs.h" +#include "flags.h" #include "h264encode.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" @@ -61,17 +62,51 @@ void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src } } +void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state) +{ + if (interlaced) { + for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-) + input_state->buffered_frames[card_index][frame_num] = + input_state->buffered_frames[card_index][frame_num - 1]; + } + input_state->buffered_frames[card_index][0] = { frame, field_num }; + } else { + for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) { + input_state->buffered_frames[card_index][frame_num] = { frame, field_num }; + } + } +} + +string generate_local_dump_filename(int frame) +{ + time_t now = time(NULL); + tm now_tm; + localtime_r(&now, &now_tm); + + char timestamp[256]; + strftime(timestamp, sizeof(timestamp), "%F-%T%z", &now_tm); + + // Use the frame number to disambiguate between two cuts starting + // on the same second. + char filename[256]; + snprintf(filename, sizeof(filename), "%s%s-f%02d%s", + LOCAL_DUMP_PREFIX, timestamp, frame % 100, LOCAL_DUMP_SUFFIX); + return filename; +} + } // namespace Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) - : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT), + : httpd(WIDTH, HEIGHT), 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) { + httpd.open_output_file(generate_local_dump_filename(/*frame=*/0).c_str()); httpd.start(9095); CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF)); @@ -100,52 +135,83 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) display_chain->set_dither_bits(0); // Don't bother. display_chain->finalize(); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd)); + h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); - 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(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44 + 1, WIDTH, HEIGHT)); - 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); - } - }, - [this]{ - resource_pool->clean_context(); - }); - card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); - card->usb->configure_card(); - } - - BMUSBCapture::start_bm_thread(); + // First try initializing the PCI devices, then USB, until we have the desired number of cards. + unsigned num_pci_devices = 0, num_usb_devices = 0; + unsigned card_index = 0; - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - cards[card_index].usb->start_bm_capture(); + 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; + } + + configure_card(card_index, format, new DeckLinkCapture(decklink, card_index)); + ++num_pci_devices; + } + decklink_iterator->Release(); + fprintf(stderr, "Found %d 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_usb_devices; + } + + if (num_usb_devices > 0) { + BMUSBCapture::start_bm_thread(); } - //chain->enable_phase_timing(true); + for (card_index = 0; card_index < num_cards; ++card_index) { + cards[card_index].capture->start_bm_capture(); + } // Set up stuff for NV12 conversion. // Cb/Cr shader. - string cbcr_vert_shader = read_file("vs-cbcr.130.vert"); + 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" - " gl_FragColor = texture2D(cbcr_tex, tc0); \n" + " FragColor = texture(cbcr_tex, tc0); \n" "} \n"; - cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader); + 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"); r128.init(2, OUTPUT_FREQUENCY); r128.integr_start(); @@ -154,7 +220,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) // 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); + peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0); alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); } @@ -162,6 +228,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) Mixer::~Mixer() { resource_pool->release_glsl_program(cbcr_program_num); + glDeleteBuffers(1, &cbcr_vbo); BMUSBCapture::stop_bm_thread(); for (unsigned card_index = 0; card_index < num_cards; ++card_index) { @@ -170,12 +237,39 @@ Mixer::~Mixer() cards[card_index].should_quit = true; // Unblock thread. cards[card_index].new_data_ready_changed.notify_all(); } - cards[card_index].usb->stop_dequeue_thread(); + cards[card_index].capture->stop_dequeue_thread(); } h264_encoder.reset(nullptr); } +void Mixer::configure_card(unsigned card_index, const QSurfaceFormat &format, CaptureInterface *capture) +{ + printf("Configuring card %d...\n", card_index); + + CaptureCard *card = &cards[card_index]; + card->capture = capture; + 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. + card->capture->set_video_frame_allocator(card->frame_allocator.get()); + card->surface = create_surface(format); + card->capture->set_dequeue_thread_callbacks( + [card]{ + eglBindAPI(EGL_OPENGL_API); + card->context = create_context(card->surface); + if (!make_current(card->context, card->surface)) { + printf("failed to create bmusb context\n"); + exit(1); + } + }, + [this]{ + resource_pool->clean_context(); + }); + card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + card->capture->configure_card(); +} + + namespace { int unwrap_timecode(uint16_t current_wrapped, int last) @@ -192,7 +286,7 @@ 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 = std::max(m, fabs(samples[i])); + m = max(m, fabs(samples[i])); } return m; } @@ -212,72 +306,21 @@ void deinterleave_samples(const vector &in, vector *out_l, vector< } } -// Returns length of a frame with the given format, in TIMEBASE units. -int64_t find_frame_length(uint16_t video_format) -{ - if (video_format == 0x0800) { - // No video signal. These green pseudo-frames seem to come at about 30.13 Hz. - // It's a strange thing, but what can you do. - return TIMEBASE * 100 / 3013; - } - if ((video_format & 0xe800) != 0xe800) { - printf("Video format 0x%04x does not appear to be a video format. Assuming 60 Hz.\n", - video_format); - return TIMEBASE / 60; - } - - // 0x8 seems to be a flag about availability of deep color on the input, - // except when it's not (e.g. it's the only difference between NTSC 23.98 - // and PAL). Rather confusing. But we clear it here nevertheless, because - // usually it doesn't mean anything. - // - // We don't really handle interlaced formats at all yet. - uint16_t normalized_video_format = video_format & ~0xe808; - if (normalized_video_format == 0x0143) { // 720p50. - return TIMEBASE / 50; - } else if (normalized_video_format == 0x0103) { // 720p60. - return TIMEBASE / 60; - } else if (normalized_video_format == 0x0121) { // 720p59.94. - return TIMEBASE * 1001 / 60000; - } else if (normalized_video_format == 0x01c3 || // 1080p30. - normalized_video_format == 0x0003) { // 1080i60. - return TIMEBASE / 30; - } else if (normalized_video_format == 0x01e1 || // 1080p29.97. - normalized_video_format == 0x0021 || // 1080i59.94. - video_format == 0xe901 || // NTSC (480i59.94, I suppose). - video_format == 0xe9c1 || // Ditto. - video_format == 0xe801) { // Ditto. - return TIMEBASE * 1001 / 30000; - } else if (normalized_video_format == 0x0063 || // 1080p25. - normalized_video_format == 0x0043 || // 1080i50. - video_format == 0xe909) { // PAL (576i50, I suppose). - return TIMEBASE / 25; - } else if (normalized_video_format == 0x008e) { // 1080p24. - return TIMEBASE / 24; - } else if (normalized_video_format == 0x00a1) { // 1080p23.98. - return TIMEBASE * 1001 / 24000; - return TIMEBASE / 25; - } else { - printf("Unknown video format 0x%04x. Assuming 60 Hz.\n", video_format); - return TIMEBASE / 60; - } -} - } // 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 video_frame, size_t video_offset, VideoFormat video_format, FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format) { CaptureCard *card = &cards[card_index]; - int64_t frame_length = find_frame_length(video_format); + int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom; size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 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); + timecode, int(video_frame.len), int(video_offset), video_format.id); if (video_frame.owner) { video_frame.owner->release_frame(video_frame); } @@ -292,7 +335,6 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, if (card->last_timecode != -1) { dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1; } - card->last_timecode = timecode; // Convert the audio to stereo fp32 and add it. vector audio; @@ -303,29 +345,38 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { 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), resetting resampler\n", - card_index); + 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. (The number of samples could be nonintegral, - // but resampling will save us then.) + // 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.resize((OUTPUT_FREQUENCY * frame_length / TIMEBASE) * 2); + 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.size() / 2); + 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; + // Done with the audio, so release it. if (audio_frame.owner) { audio_frame.owner->release_frame(audio_frame); @@ -338,10 +389,12 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, if (card->should_quit) return; } - if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) { + size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 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)\n", - card_index, video_frame.len - video_offset); + printf("Card %d: Dropping video frame with wrong length (%ld; expected %ld)\n", + card_index, video_frame.len - video_offset, expected_length); } if (video_frame.owner) { video_frame.owner->release_frame(video_frame); @@ -354,6 +407,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, 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(); @@ -361,39 +415,111 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, return; } - const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata; - GLuint pbo = userdata->pbo; - check_error(); - glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo); - check_error(); - glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size); - check_error(); - //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT); - //check_error(); + PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; + + unsigned num_fields = video_format.interlaced ? 2 : 1; + timespec frame_upload_start; + 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; + assert(frame_length % 2 == 0); + frame_length /= 2; + num_fields = 2; + clock_gettime(CLOCK_MONOTONIC, &frame_upload_start); + } + userdata->last_interlaced = video_format.interlaced; + userdata->last_frame_rate_nom = video_format.frame_rate_nom; + userdata->last_frame_rate_den = video_format.frame_rate_den; + RefCountedFrame new_frame(video_frame); // Upload the textures. - glBindTexture(GL_TEXTURE_2D, userdata->tex_y); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44) / 2 + WIDTH * 25 + 22)); - check_error(); - glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22)); - check_error(); - glBindTexture(GL_TEXTURE_2D, 0); - check_error(); - GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); - check_error(); - assert(fence != nullptr); + 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) { + 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; + } - { - unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = RefCountedFrame(video_frame); - card->new_frame_length = frame_length; - card->new_data_ready_fence = fence; - card->dropped_frames = dropped_frames; - card->new_data_ready_changed.notify_all(); + GLuint pbo = userdata->pbo; + check_error(); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo); + check_error(); + glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT); + check_error(); + + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(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, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line)); + check_error(); + glBindTexture(GL_TEXTURE_2D, 0); + check_error(); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0); + check_error(); + GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); + check_error(); + assert(fence != nullptr); + + if (field == 1) { + // Don't upload the second field as fast as we can; wait until + // the field time has approximately passed. (Otherwise, we could + // get timing jitter against the other sources, and possibly also + // 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) ; + } + + { + 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 = video_format.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; + } + } } } @@ -410,7 +536,7 @@ void Mixer::thread_func() clock_gettime(CLOCK_MONOTONIC, &start); int frame = 0; - int dropped_frames = 0; + int stats_dropped_frames = 0; while (!should_quit) { CaptureCard card_copy[MAX_CARDS]; @@ -425,10 +551,11 @@ void Mixer::thread_func() 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; @@ -437,10 +564,12 @@ void Mixer::thread_func() 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); } } // Resample the audio as needed, including from previously dropped frames. + assert(num_cards > 0); for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) { { // Signal to the audio thread to process this frame. @@ -452,12 +581,13 @@ void Mixer::thread_func() // 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. - ++dropped_frames; + ++stats_dropped_frames; pts_int += card_copy[0].new_frame_length; } } if (audio_level_callback != nullptr) { + unique_lock lock(compressor_mutex); double loudness_s = r128.loudness_S(); double loudness_i = r128.integrated(); double loudness_range_low = r128.range_min(); @@ -465,7 +595,8 @@ void Mixer::thread_func() audio_level_callback(loudness_s, 20.0 * log10(peak), loudness_i, loudness_range_low, loudness_range_high, - last_gain_staging_db); + gain_staging_db, 20.0 * log10(final_makeup_gain), + correlation.get_correlation()); } for (unsigned card_index = 1; card_index < num_cards; ++card_index) { @@ -481,7 +612,7 @@ 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 (card_copy[0].new_frame->len == 0) { - ++dropped_frames; + ++stats_dropped_frames; pts_int += card_copy[0].new_frame_length; continue; } @@ -492,7 +623,7 @@ void Mixer::thread_func() continue; assert(card->new_frame != nullptr); - bmusb_current_rendering_frame[card_index] = card->new_frame; + insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state); check_error(); // The new texture might still be uploaded, @@ -503,14 +634,13 @@ void Mixer::thread_func() glDeleteSync(card->new_data_ready_fence); check_error(); } - const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata; - theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr); } // Get the main chain from the theme, and set its state immediately. - pair> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT); - EffectChain *chain = theme_main_chain.first; - theme_main_chain.second(); + 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); @@ -537,15 +667,8 @@ void Mixer::thread_func() RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); check_error(); - // Make sure the H.264 gets a reference to all the - // input frames needed, so that they are not released back - // until the rendering is done. - vector input_frames; - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - input_frames.push_back(bmusb_current_rendering_frame[card_index]); - } 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, input_frames); + h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames); ++frame; pts_int += card_copy[0].new_frame_length; @@ -564,15 +687,11 @@ void Mixer::thread_func() // Set up preview and any additional channels. for (int i = 1; i < theme->get_num_channels() + 2; ++i) { DisplayFrame display_frame; - pair> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions - display_frame.chain = chain.first; - display_frame.setup_chain = chain.second; + 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; - - // FIXME: possible to do better? - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]); - } + display_frame.input_frames = chain.input_frames; display_frame.temp_textures = {}; output_channel[i].output_frame(display_frame); } @@ -582,11 +701,20 @@ void Mixer::thread_func() 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, dropped_frames, elapsed, frame / elapsed, + frame, stats_dropped_frames, elapsed, frame / elapsed, 1e3 * elapsed / frame); // 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, global_flags.va_display, WIDTH, HEIGHT, &httpd)); + } + #if 0 // Reset every 100 frames, so that local variations in frame times // (especially for the first few frames, when the shaders are @@ -623,6 +751,11 @@ 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); + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { samples_card.resize(num_samples * 2); { @@ -631,8 +764,7 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) printf("Card %d reported previous underrun.\n", card_index); } } - // TODO: Allow using audio from the other card(s) as well. - if (card_index == 0) { + if (card_index == selected_audio_card) { samples_out = move(samples_card); } } @@ -641,7 +773,9 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) // 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 (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 @@ -649,15 +783,23 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) // 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. - float ref_level_dbfs = -14.0f; { - 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); + 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 @@ -706,13 +848,58 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) 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; } - // Find R128 levels. + // 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() }; - r128.process(left.size(), ptrs); + { + unique_lock lock(compressor_mutex); + r128.process(left.size(), ptrs); + correlation.process_samples(samples_out); + } // Send the samples to the sound card. if (alsa) { @@ -729,12 +916,6 @@ void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) glGenVertexArrays(1, &vao); check_error(); - float vertices[] = { - 0.0f, 2.0f, - 0.0f, 0.0f, - 2.0f, 0.0f - }; - glBindVertexArray(vao); check_error(); @@ -761,17 +942,28 @@ void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f }; set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0); - 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. + 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(); - cleanup_vertex_attribute(cbcr_program_num, "position", position_vbo); - cleanup_vertex_attribute(cbcr_program_num, "texcoord", texcoord_vbo); + 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); @@ -816,6 +1008,7 @@ void Mixer::reset_meters() peak = 0.0f; r128.reset(); r128.integr_start(); + correlation.reset(); } Mixer::OutputChannel::~OutputChannel()