X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=04a35530cc7af36d7681d1aca2ebef695be960e0;hb=b22d8d6b38d060ccc5dfa591712211caf9ca3968;hp=0e362f069395874c837dac1bf732c9e1d32d25a0;hpb=bc7b6954cc294b1b26e5ef3409e39e6443ff120d;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 0e362f0..04a3553 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,7 +1,3 @@ -#define WIDTH 1280 -#define HEIGHT 720 -#define EXTRAHEIGHT 30 - #undef Success #include "mixer.h" @@ -14,12 +10,12 @@ #include #include #include +#include #include #include #include #include #include -#include #include #include #include @@ -66,7 +62,7 @@ void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src } // namespace Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) - : httpd("test.ts", WIDTH, HEIGHT), + : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT), num_cards(num_cards), mixer_surface(create_surface(format)), h264_encoder_surface(create_surface(format)), @@ -109,7 +105,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) 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->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB. card->usb->set_video_frame_allocator(card->frame_allocator.get()); card->surface = create_surface(format); card->usb->set_dequeue_thread_callbacks( @@ -157,6 +153,8 @@ 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); + + alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); } Mixer::~Mixer() @@ -172,6 +170,8 @@ Mixer::~Mixer() } cards[card_index].usb->stop_dequeue_thread(); } + + h264_encoder.reset(nullptr); } namespace { @@ -218,7 +218,15 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { CaptureCard *card = &cards[card_index]; - if (audio_frame.len - audio_offset > 30000) { + unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom; + bool interlaced; + + decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom, + &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now. + int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom; + + 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); @@ -231,16 +239,13 @@ 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); @@ -249,24 +254,38 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { 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); + // Number of samples per frame if we need to insert silence. + // (Could be nonintegral, but resampling will save us then.) + int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom; + + if (dropped_frames > MAX_FPS * 2) { + fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n", + card_index, card->last_timecode, timecode); card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + dropped_frames = 0; } else if (dropped_frames > 0) { // Insert silence as needed. fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n", card_index, dropped_frames, timecode); - vector silence; - silence.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); @@ -279,7 +298,8 @@ 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) { + if (video_frame.len - video_offset == 0 || + video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) { if (video_frame.len != 0) { printf("Card %d: Dropping video frame with wrong length (%ld)\n", card_index, video_frame.len - video_offset); @@ -294,6 +314,8 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, unique_lock lock(bmusb_mutex); card->new_data_ready = true; card->new_frame = RefCountedFrame(FrameAllocator::Frame()); + card->new_frame_length = frame_length; + card->new_frame_interlaced = false; card->new_data_ready_fence = nullptr; card->dropped_frames = dropped_frames; card->new_data_ready_changed.notify_all(); @@ -301,38 +323,109 @@ 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 = interlaced ? 2 : 1; + timespec frame_upload_start; + if (interlaced) { + // NOTE: This isn't deinterlacing. This is just sending the two fields along + // as separate frames without considering anything like the half-field offset. + // We'll need to add a proper deinterlacer on the receiving side to get this right. + assert(height % 2 == 0); + height /= 2; + assert(frame_length % 2 == 0); + frame_length /= 2; + num_fields = 2; + clock_gettime(CLOCK_MONOTONIC, &frame_upload_start); + } + 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 = width / 2; + size_t cbcr_offset = video_offset / 2; + size_t y_offset = video_frame.size / 2 + video_offset / 2; + + for (unsigned field = 0; field < num_fields; ++field) { + unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22); + + if (userdata->tex_y[field] == 0 || + userdata->tex_cbcr[field] == 0 || + width != userdata->last_width[field] || + height != userdata->last_height[field]) { + // We changed resolution since last use of this texture, so we need to create + // a new object. Note that this each card has its own PBOFrameAllocator, + // we don't need to worry about these flip-flopping between resolutions. + glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); + check_error(); + glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + check_error(); + userdata->last_width[field] = width; + userdata->last_height[field] = 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_ARB, pbo); + check_error(); + glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size); + 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, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t))); + check_error(); + glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); + check_error(); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line)); + check_error(); + glBindTexture(GL_TEXTURE_2D, 0); + check_error(); + GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); + check_error(); + assert(fence != nullptr); + + 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 = 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; + } + } } } @@ -349,10 +442,11 @@ 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]; + int num_samples[MAX_CARDS]; { unique_lock lock(bmusb_mutex); @@ -366,24 +460,40 @@ void Mixer::thread_func() card_copy[card_index].usb = card->usb; card_copy[card_index].new_data_ready = card->new_data_ready; card_copy[card_index].new_frame = card->new_frame; + card_copy[card_index].new_frame_length = card->new_frame_length; + card_copy[card_index].new_frame_field = card->new_frame_field; + card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced; card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence; card_copy[card_index].dropped_frames = card->dropped_frames; card->new_data_ready = false; card->new_data_ready_changed.notify_all(); + + int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples; + num_samples[card_index] = num_samples_times_timebase / TIMEBASE; + card->fractional_samples = num_samples_times_timebase % TIMEBASE; + assert(num_samples[card_index] >= 0); } } // 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(); + { + // Signal to the audio thread to process this frame. + unique_lock lock(audio_mutex); + audio_task_queue.push(AudioTask{pts_int, num_samples[0]}); + audio_task_queue_changed.notify_one(); + } if (frame_num != card_copy[0].dropped_frames) { - // For dropped frames, increase the pts. - ++dropped_frames; - pts_int += TIMEBASE / FPS; + // For dropped frames, increase the pts. Note that if the format changed + // in the meantime, we have no way of detecting that; we just have to + // assume the frame length is always the same. + ++stats_dropped_frames; + pts_int += card_copy[0].new_frame_length; } } if (audio_level_callback != nullptr) { + unique_lock lock(r128_mutex); double loudness_s = r128.loudness_S(); double loudness_i = r128.integrated(); double loudness_range_low = r128.range_min(); @@ -407,8 +517,8 @@ 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; - pts_int += TIMEBASE / FPS; + ++stats_dropped_frames; + pts_int += card_copy[0].new_frame_length; continue; } @@ -418,7 +528,16 @@ void Mixer::thread_func() continue; assert(card->new_frame != nullptr); - bmusb_current_rendering_frame[card_index] = card->new_frame; + if (card->new_frame_interlaced) { + for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-) + buffered_frames[card_index][frame_num] = buffered_frames[card_index][frame_num - 1]; + } + buffered_frames[card_index][0] = { card->new_frame, card->new_frame_field }; + } else { + for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) { + buffered_frames[card_index][frame_num] = { card->new_frame, card->new_frame_field }; + } + } check_error(); // The new texture might still be uploaded, @@ -429,14 +548,12 @@ 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); + EffectChain *chain = theme_main_chain.chain; + theme_main_chain.setup_chain(); GLuint y_tex, cbcr_tex; bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex); @@ -463,17 +580,10 @@ 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 += TIMEBASE / FPS; + pts_int += card_copy[0].new_frame_length; // The live frame just shows the RGBA texture we just rendered. // It owns rgba_tex now. @@ -490,15 +600,11 @@ 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); // 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); } @@ -508,7 +614,7 @@ 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(); } @@ -529,15 +635,31 @@ void Mixer::thread_func() resource_pool->clean_context(); } -void Mixer::process_audio_one_frame() +void Mixer::audio_thread_func() +{ + while (!should_quit) { + AudioTask task; + + { + unique_lock lock(audio_mutex); + audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); }); + task = audio_task_queue.front(); + audio_task_queue.pop(); + } + + 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; 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)) { + if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) { printf("Card %d reported previous underrun.\n", card_index); } } @@ -547,9 +669,10 @@ void Mixer::process_audio_one_frame() } } - // 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.) + // Cut away everything under 120 Hz (or whatever the cutoff is); + // we don't need it for voice, and it will reduce headroom + // and confuse the compressor. (In particular, any hums at 50 or 60 Hz + // should be dampened.) locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f); // Apply a level compressor to get the general level right. @@ -621,10 +744,18 @@ void Mixer::process_audio_one_frame() vector left, right; deinterleave_samples(samples_out, &left, &right); float *ptrs[] = { left.data(), right.data() }; - r128.process(left.size(), ptrs); + { + unique_lock lock(r128_mutex); + r128.process(left.size(), ptrs); + } + + // 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. + h264_encoder->add_audio(frame_pts_int, move(samples_out)); } void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex) @@ -694,12 +825,14 @@ 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; mixer_thread.join(); + audio_thread.join(); } void Mixer::transition_clicked(int transition_num)