X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=80b0df1ea0df60d852bc530cb0d5535e99e2f92a;hb=8d575ed1bf8c01daaeb4bdf87de53d089448ba93;hp=8ccd7fa0e5d71fbe24a26fd0e8c47b8092df5044;hpb=3b77fb6a74f7a58033168202a772f3630fb9a30f;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 8ccd7fa..80b0df1 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,25 +1,21 @@ -#define WIDTH 1280 -#define HEIGHT 720 -#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 @@ -30,11 +26,14 @@ #include #include #include +#include #include "bmusb/bmusb.h" #include "context.h" +#include "decklink_capture.h" #include "defs.h" -#include "h264encode.h" +#include "flags.h" +#include "video_encoder.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" #include "timebase.h" @@ -51,6 +50,7 @@ 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++; @@ -63,19 +63,74 @@ void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src } } +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) { + 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 }; + } + } +} + } // namespace +void QueueLengthPolicy::update_policy(int queue_length) +{ + if (queue_length < 0) { // Starvation. + if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) { + ++safe_queue_length; + fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n", + card_index, safe_queue_length); + } + frames_with_at_least_one = 0; + been_at_safe_point_since_last_starvation = false; + 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 { + frames_with_at_least_one = 0; + } +} + Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) - : httpd("test.ts", WIDTH, HEIGHT), + : httpd(), 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.start(9095); - CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF)); check_error(); @@ -84,9 +139,10 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) movit_texel_subpixel_precision /= 2.0; resource_pool.reset(new ResourcePool); - theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards)); + theme.reset(new Theme(global_flags.theme_filename.c_str(), resource_pool.get(), num_cards)); for (unsigned i = 0; i < NUM_OUTPUTS; ++i) { output_channel[i].parent = this; + output_channel[i].channel = i; } ImageFormat inout_format; @@ -102,78 +158,143 @@ 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)); + video_encoder.reset(new VideoEncoder(resource_pool.get(), 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, 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(); - } + // Start listening for clients only once VideoEncoder has written its header, if any. + httpd.start(9095); - 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; } - //chain->enable_phase_timing(true); + if (num_usb_devices > 0) { + BMUSBCapture::start_bm_thread(); + } + + 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 = 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(); locut.init(FILTER_HPF, 2); + // 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_limiter_enabled(false); + set_compressor_enabled(false); + } + // 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); + + if (global_flags.enable_alsa_output) { + alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); + } } 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) { { unique_lock lock(bmusb_mutex); cards[card_index].should_quit = true; // Unblock thread. - cards[card_index].new_data_ready_changed.notify_all(); + cards[card_index].new_frames_changed.notify_all(); } - cards[card_index].usb->stop_dequeue_thread(); + cards[card_index].capture->stop_dequeue_thread(); } + + video_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->resampling_queue.reset(new ResamplingQueue(card_index, OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + card->capture->configure_card(); } + namespace { int unwrap_timecode(uint16_t current_wrapped, int last) @@ -190,7 +311,7 @@ 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; } @@ -213,15 +334,38 @@ void deinterleave_samples(const vector &in, vector *out_l, vector< } // namespace void Mixer::bm_frame(unsigned card_index, uint16_t timecode, - FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format, - FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format) + FrameAllocator::Frame video_frame, size_t video_offset, VideoFormat video_format, + FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format) { CaptureCard *card = &cards[card_index]; - if (audio_frame.len - audio_offset > 30000) { + 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); + if (sec_since_last_switch > switch_time_s) { + // It isn't this mode; try the next one. + mode_scanlist_index[card_index]++; + mode_scanlist_index[card_index] %= mode_scanlist[card_index].size(); + cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]); + last_mode_scan_change[card_index] = now; + } + } + } + + int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom; + + size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 0; + if (num_samples > OUTPUT_FREQUENCY / 10) { 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); } @@ -231,58 +375,77 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, return; } - int unwrapped_timecode = timecode; + int64_t local_pts = card->next_local_pts; int dropped_frames = 0; if (card->last_timecode != -1) { - unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode); - dropped_frames = unwrapped_timecode - card->last_timecode - 1; + dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1; } - card->last_timecode = unwrapped_timecode; // Convert the audio to stereo fp32 and add it. - size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0; vector audio; audio.resize(num_samples * 2); - convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples); + 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); + } // Add the audio. { unique_lock lock(card->audio_mutex); - int unwrapped_timecode = timecode; - if (dropped_frames > FPS * 2) { - fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n", - card_index); - card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + // 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.resize((OUTPUT_FREQUENCY / FPS) * 2); + vector silence(silence_samples * 2, 0.0f); for (int i = 0; i < dropped_frames; ++i) { - card->resampling_queue->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS)); + 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; } } - card->resampling_queue->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples); + 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); } - { - // Wait until the previous frame was consumed. - unique_lock lock(bmusb_mutex); - card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; }); - if (card->should_quit) return; - } - - 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); @@ -292,47 +455,133 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, // so that pts can go up accordingly. { unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = RefCountedFrame(FrameAllocator::Frame()); - card->new_data_ready_fence = nullptr; - card->dropped_frames = dropped_frames; - card->new_data_ready_changed.notify_all(); + 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)); + card->new_frames_changed.notify_all(); } 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_has_signal = video_format.has_signal; + userdata->last_frame_rate_nom = video_format.frame_rate_nom; + userdata->last_frame_rate_den = video_format.frame_rate_den; + RefCountedFrame frame(video_frame); // Upload the textures. - 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) { + // Put the actual texture upload in a lambda that is executed in the main thread. + // It is entirely possible to do this in the same thread (and it might even be + // faster, depending on the GPU and driver), but it appears to be trickling + // driver bugs very easily. + // + // Note that this means we must hold on to the actual frame data in + // until the upload command is run, but we hold on to much longer than that + // (in fact, all the way until we no longer use the texture in rendering). + auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, userdata]() { + unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22); + + if (userdata->tex_y[field] == 0 || + userdata->tex_cbcr[field] == 0 || + video_format.width != userdata->last_width[field] || + video_format.height != userdata->last_height[field]) { + // We changed resolution since last use of this texture, so we need to create + // a new object. Note that this each card has its own PBOFrameAllocator, + // we don't need to worry about these flip-flopping between resolutions. + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, video_format.height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + userdata->last_width[field] = video_format.width; + userdata->last_height[field] = video_format.height; + } - { - unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = RefCountedFrame(video_frame); - 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, pbo); + check_error(); + + size_t field_y_start = y_offset + video_format.width * field_start_line; + size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t); + + if (global_flags.flush_pbos) { + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height); + check_error(); + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t)); + check_error(); + } + + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + 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(); + }; + + 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); + CaptureCard::NewFrame new_frame; + new_frame.frame = frame; + new_frame.length = frame_length; + new_frame.field = field; + new_frame.interlaced = video_format.interlaced; + new_frame.upload_func = upload_func; + new_frame.dropped_frames = dropped_frames; + card->new_frames.push(move(new_frame)); + card->new_frames_changed.notify_all(); + } } } @@ -349,170 +598,78 @@ 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]; + CaptureCard::NewFrame new_frames[MAX_CARDS]; + bool has_new_frame[MAX_CARDS] = { false }; + int num_samples[MAX_CARDS] = { 0 }; - { - unique_lock lock(bmusb_mutex); + // TODO: Add a timeout. + unsigned master_card_index = theme->map_signal(master_clock_channel); + assert(master_card_index < num_cards); - // The first card is the master timer, so wait for it to have a new frame. - // TODO: Make configurable, and with a timeout. - cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; }); - - for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - CaptureCard *card = &cards[card_index]; - card_copy[card_index].usb = card->usb; - card_copy[card_index].new_data_ready = card->new_data_ready; - card_copy[card_index].new_frame = card->new_frame; - card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence; - card_copy[card_index].dropped_frames = card->dropped_frames; - card->new_data_ready = false; - card->new_data_ready_changed.notify_all(); - } - } + 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(); - // Resample the audio as needed, including from previously dropped frames. - for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) { - process_audio_one_frame(); - if (frame_num != card_copy[0].dropped_frames) { - // For dropped frames, increase the pts. - ++dropped_frames; - pts_int += TIMEBASE / FPS; + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { + if (card_index == master_card_index || !has_new_frame[card_index]) { + continue; } - } - - if (audio_level_callback != nullptr) { - 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, - last_gain_staging_db); - } - - for (unsigned card_index = 1; card_index < num_cards; ++card_index) { - if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) { - ++card_copy[card_index].dropped_frames; + if (new_frames[card_index].frame->len == 0) { + ++new_frames[card_index].dropped_frames; } - if (card_copy[card_index].dropped_frames > 0) { + if (new_frames[card_index].dropped_frames > 0) { printf("Card %u dropped %d frames before this\n", - card_index, int(card_copy[card_index].dropped_frames)); + card_index, int(new_frames[card_index].dropped_frames)); } } // If the first card is reporting a corrupted or otherwise dropped frame, // just increase the pts (skipping over this frame) and don't try to compute anything new. - if (card_copy[0].new_frame->len == 0) { - ++dropped_frames; - pts_int += TIMEBASE / FPS; + if (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) { - CaptureCard *card = &card_copy[card_index]; - if (!card->new_data_ready || card->new_frame->len == 0) + if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0) continue; - assert(card->new_frame != nullptr); - bmusb_current_rendering_frame[card_index] = card->new_frame; + CaptureCard::NewFrame *new_frame = &new_frames[card_index]; + assert(new_frame->frame != nullptr); + insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state); check_error(); - // The new texture might still be uploaded, - // tell the GPU to wait until it's there. - if (card->new_data_ready_fence) { - glWaitSync(card->new_data_ready_fence, /*flags=*/0, GL_TIMEOUT_IGNORED); - check_error(); - glDeleteSync(card->new_data_ready_fence); - check_error(); + // The new texture might need uploading before use. + if (new_frame->upload_func) { + new_frame->upload_func(); + new_frame->upload_func = nullptr; } - 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(); - - GLuint y_tex, cbcr_tex; - bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex); - assert(got_frame); - - // Render main chain. - GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT); - GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up. - GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex); - check_error(); - chain->render_to_fbo(fbo, WIDTH, HEIGHT); - resource_pool->release_fbo(fbo); - - subsample_chroma(cbcr_full_tex, cbcr_tex); - resource_pool->release_2d_texture(cbcr_full_tex); - - // Set the right state for rgba_tex. - glBindFramebuffer(GL_FRAMEBUFFER, 0); - glBindTexture(GL_TEXTURE_2D, rgba_tex); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); - - RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); - check_error(); - - // 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); + int64_t duration = new_frames[master_card_index].length; + render_one_frame(duration); ++frame; - pts_int += TIMEBASE / FPS; - - // The live frame just shows the RGBA texture we just rendered. - // It owns rgba_tex now. - DisplayFrame live_frame; - live_frame.chain = display_chain.get(); - live_frame.setup_chain = [this, rgba_tex]{ - display_input->set_texture_num(rgba_tex); - }; - live_frame.ready_fence = fence; - live_frame.input_frames = {}; - live_frame.temp_textures = { rgba_tex }; - output_channel[OUTPUT_LIVE].output_frame(live_frame); - - // Set up preview and any additional channels. - for (int i = 1; i < theme->get_num_channels() + 2; ++i) { - DisplayFrame display_frame; - pair> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions - display_frame.chain = chain.first; - display_frame.setup_chain = chain.second; - 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.temp_textures = {}; - output_channel[i].output_frame(display_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) { 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. + video_encoder->do_cut(frame); + } + #if 0 // Reset every 100 frames, so that local variations in frame times // (especially for the first few frames, when the shaders are @@ -529,28 +686,187 @@ void Mixer::thread_func() resource_pool->clean_context(); } -void Mixer::process_audio_one_frame() +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]) +{ + // 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(); }); + + 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; + } + new_frames[card_index] = move(card->new_frames.front()); + has_new_frame[card_index] = true; + card->new_frames.pop(); + card->new_frames_changed.notify_all(); + + 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) { + // 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); + } 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(); + } + } + } +} + +void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame) +{ + // 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) { + { + // Signal to the audio thread to process this frame. + unique_lock lock(audio_mutex); + audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame}); + audio_task_queue_changed.notify_one(); + } + if (frame_num != dropped_frames) { + // For dropped frames, increase the pts. Note that if the format changed + // in the meantime, we have no way of detecting that; we just have to + // assume the frame length is always the same. + pts_int += length_per_frame; + } + } +} + +void Mixer::render_one_frame(int64_t duration) +{ + // Get the main chain from the theme, and set its state immediately. + Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state); + 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); + assert(got_frame); + + // Render main chain. + GLuint cbcr_full_tex = resource_pool->create_2d_texture(GL_RG8, WIDTH, HEIGHT); + GLuint rgba_tex = resource_pool->create_2d_texture(GL_RGB565, WIDTH, HEIGHT); // Saves texture bandwidth, although dithering gets messed up. + GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex); + check_error(); + chain->render_to_fbo(fbo, WIDTH, HEIGHT); + resource_pool->release_fbo(fbo); + + subsample_chroma(cbcr_full_tex, cbcr_tex); + resource_pool->release_2d_texture(cbcr_full_tex); + + // Set the right state for rgba_tex. + glBindFramebuffer(GL_FRAMEBUFFER, 0); + glBindTexture(GL_TEXTURE_2D, rgba_tex); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE); + + 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); + + // The live frame just shows the RGBA texture we just rendered. + // It owns rgba_tex now. + DisplayFrame live_frame; + live_frame.chain = display_chain.get(); + live_frame.setup_chain = [this, rgba_tex]{ + display_input->set_texture_num(rgba_tex); + }; + live_frame.ready_fence = fence; + live_frame.input_frames = {}; + live_frame.temp_textures = { rgba_tex }; + output_channel[OUTPUT_LIVE].output_frame(live_frame); + + // Set up preview and any additional channels. + for (int i = 1; i < theme->get_num_channels() + 2; ++i) { + DisplayFrame display_frame; + Theme::Chain chain = theme->get_chain(i, pts(), WIDTH, HEIGHT, input_state); // FIXME: dimensions + display_frame.chain = chain.chain; + display_frame.setup_chain = chain.setup_chain; + display_frame.ready_fence = fence; + display_frame.input_frames = chain.input_frames; + display_frame.temp_textures = {}; + output_channel[i].output_frame(display_frame); + } +} + +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() +{ + while (!should_quit) { + AudioTask task; + + { + unique_lock lock(audio_mutex); + audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); }); + if (should_quit) { + return; + } + task = audio_task_queue.front(); + audio_task_queue.pop(); + } + + 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); + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { - samples_card.resize((OUTPUT_FREQUENCY / FPS) * 2); + samples_card.resize(num_samples * 2); { unique_lock lock(cards[card_index].audio_mutex); - if (!cards[card_index].resampling_queue->get_output_samples(pts(), &samples_card[0], OUTPUT_FREQUENCY / FPS)) { - printf("Card %d reported previous underrun.\n", card_index); - } + cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples); } - // 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); } } - // Cut away everything under 150 Hz; 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); + // 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 @@ -558,15 +874,23 @@ void Mixer::process_audio_one_frame() // 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 @@ -578,9 +902,19 @@ void Mixer::process_audio_one_frame() // float limiter_att, compressor_att; - // Then a limiter at +0 dB (so, -14 dBFS) to take out the worst peaks only. + // 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. - // Probably more tuning is warranted here. if (limiter_enabled) { float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f); float ratio = 30.0f; @@ -591,17 +925,6 @@ void Mixer::process_audio_one_frame() // limiter_att = limiter.get_attenuation(); } - // Finally, 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(); - } - // printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att)); // Upsample 4x to find interpolated peak. @@ -616,16 +939,66 @@ void Mixer::process_audio_one_frame() 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) { + alsa->write(samples_out); + } - // Actually add the samples to the output. - h264_encoder->add_audio(pts_int, move(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) @@ -634,12 +1007,6 @@ 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(); @@ -666,17 +1033,28 @@ 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); @@ -695,12 +1073,15 @@ void Mixer::release_display_frame(DisplayFrame *frame) void Mixer::start() { mixer_thread = thread(&Mixer::thread_func, this); + audio_thread = thread(&Mixer::audio_thread_func, this); } void Mixer::quit() { should_quit = true; + audio_task_queue_changed.notify_one(); mixer_thread.join(); + audio_thread.join(); } void Mixer::transition_clicked(int transition_num) @@ -719,6 +1100,24 @@ void Mixer::reset_meters() peak = 0.0f; r128.reset(); r128.integr_start(); + correlation.reset(); +} + +void Mixer::start_mode_scanning(unsigned card_index) +{ + assert(card_index < num_cards); + if (is_mode_scanning[card_index]) { + return; + } + is_mode_scanning[card_index] = true; + mode_scanlist[card_index].clear(); + for (const auto &mode : cards[card_index].capture->get_available_video_modes()) { + mode_scanlist[card_index].push_back(mode.first); + } + assert(!mode_scanlist[card_index].empty()); + mode_scanlist_index[card_index] = 0; + cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]); + clock_gettime(CLOCK_MONOTONIC, &last_mode_scan_change[card_index]); } Mixer::OutputChannel::~OutputChannel() @@ -744,9 +1143,47 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) has_ready_frame = true; } - if (has_new_frame_ready_callback) { + if (new_frame_ready_callback) { new_frame_ready_callback(); } + + // Reduce the number of callbacks by filtering duplicates. The reason + // why we bother doing this is that Qt seemingly can get into a state + // where its builds up an essentially unbounded queue of signals, + // consuming more and more memory, and there's no good way of collapsing + // user-defined signals or limiting the length of the queue. + if (transition_names_updated_callback) { + vector transition_names = global_mixer->get_transition_names(); + bool changed = false; + if (transition_names.size() != last_transition_names.size()) { + changed = true; + } else { + for (unsigned i = 0; i < transition_names.size(); ++i) { + if (transition_names[i] != last_transition_names[i]) { + changed = true; + break; + } + } + } + if (changed) { + transition_names_updated_callback(transition_names); + last_transition_names = transition_names; + } + } + if (name_updated_callback) { + string name = global_mixer->get_channel_name(channel); + if (name != last_name) { + name_updated_callback(name); + last_name = name; + } + } + if (color_updated_callback) { + string color = global_mixer->get_channel_color(channel); + if (color != last_color) { + color_updated_callback(color); + last_color = color; + } + } } bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) @@ -777,5 +1214,21 @@ bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback) { new_frame_ready_callback = callback; - has_new_frame_ready_callback = true; } + +void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback) +{ + transition_names_updated_callback = callback; +} + +void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback) +{ + name_updated_callback = callback; +} + +void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback) +{ + color_updated_callback = callback; +} + +mutex RefCountedGLsync::fence_lock;