X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=f3ff4c385c1d46b4f3b03dc5832a6ba816a8587a;hb=8fe6a683cb5bc9f04555c8cb9257f33c4d356ded;hp=1603b8dc47f8ea894ef48c51232c4b6f7c6edcb3;hpb=e1a81d67edc8e3634ee8f3c30e558e171367f628;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 1603b8d..f3ff4c3 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,5 +1,3 @@ -#define EXTRAHEIGHT 30 - #undef Success #include "mixer.h" @@ -12,12 +10,12 @@ #include #include #include +#include #include #include #include #include #include -#include #include #include #include @@ -61,10 +59,42 @@ 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)), @@ -72,6 +102,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) 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)); @@ -107,7 +138,7 @@ 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 + 1, 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( @@ -132,8 +163,6 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) cards[card_index].usb->start_bm_capture(); } - //chain->enable_phase_timing(true); - // Set up stuff for NV12 conversion. // Cb/Cr shader. @@ -145,7 +174,8 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) "void main() { \n" " gl_FragColor = texture2D(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); r128.init(2, OUTPUT_FREQUENCY); r128.integr_start(); @@ -192,7 +222,7 @@ 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; } @@ -220,10 +250,11 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { CaptureCard *card = &cards[card_index]; - int width, height, frame_rate_nom, frame_rate_den; + 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, &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now. + decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom, + &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now. int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom; size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0; @@ -263,12 +294,12 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, 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(silence_samples * 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_samples); // Note that if the format changed in the meantime, we have @@ -299,10 +330,12 @@ 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 = width * (height + extra_lines_top + 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); @@ -315,6 +348,7 @@ 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(); @@ -322,39 +356,109 @@ 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) { + // Send the two fields along as separate frames; the other side will need to add + // a deinterlacer to actually get this right. + assert(height % 2 == 0); + height /= 2; + assert(frame_length % 2 == 0); + frame_length /= 2; + num_fields = 2; + clock_gettime(CLOCK_MONOTONIC, &frame_upload_start); + } + userdata->last_interlaced = interlaced; + userdata->last_frame_rate_nom = frame_rate_nom; + userdata->last_frame_rate_den = frame_rate_den; + RefCountedFrame new_frame(video_frame); // 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_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, 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; + } + } } } @@ -390,6 +494,8 @@ void Mixer::thread_func() 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; @@ -420,6 +526,7 @@ void Mixer::thread_func() } 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(); @@ -427,7 +534,7 @@ 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)); } for (unsigned card_index = 1; card_index < num_cards; ++card_index) { @@ -454,7 +561,7 @@ 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, @@ -465,14 +572,13 @@ 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); @@ -499,15 +605,8 @@ 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; @@ -526,15 +625,11 @@ 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); } @@ -549,6 +644,15 @@ void Mixer::thread_func() // chain->print_phase_timing(); } + if (should_cut.exchange(false)) { // Test and clear. + string filename = generate_local_dump_filename(frame); + printf("Starting new recording: %s\n", filename.c_str()); + h264_encoder->shutdown(); + httpd.close_output_file(); + httpd.open_output_file(filename.c_str()); + h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd)); + } + #if 0 // Reset every 100 frames, so that local variations in frame times // (especially for the first few frames, when the shaders are @@ -611,15 +715,23 @@ 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 @@ -670,11 +782,54 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) peak = max(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2)); } + // 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. 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); + } // Send the samples to the sound card. if (alsa) {