X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=mixer.cpp;h=21c3f696100eec8407bd8b19ff22becdd7be4882;hb=a4b842506735938ae8c4c7ad19fc9b8623e89e93;hp=618ce31eccebf9ff8a6d40303fb187768dc611bf;hpb=c9dffdd52a4847c6cf5590e8c45d1cf5ee6177d3;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 618ce31..21c3f69 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -4,11 +4,11 @@ #include #include -#include #include #include #include #include +#include #include #include #include @@ -30,6 +30,7 @@ #include "bmusb/bmusb.h" #include "context.h" #include "defs.h" +#include "flags.h" #include "h264encode.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" @@ -98,6 +99,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) num_cards(num_cards), mixer_surface(create_surface(format)), h264_encoder_surface(create_surface(format)), + correlation(OUTPUT_FREQUENCY), level_compressor(OUTPUT_FREQUENCY), limiter(OUTPUT_FREQUENCY), compressor(OUTPUT_FREQUENCY) @@ -131,7 +133,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) display_chain->set_dither_bits(0); // Don't bother. display_chain->finalize(); - h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd)); + h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); for (unsigned card_index = 0; card_index < num_cards; ++card_index) { printf("Configuring card %d...\n", card_index); @@ -166,17 +168,46 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) // 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"; 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(); @@ -184,7 +215,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise, // and there's a limit to how important the peak meter is. - peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16); + peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16, /*frel=*/1.0); alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2)); } @@ -192,6 +223,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) Mixer::~Mixer() { resource_pool->release_glsl_program(cbcr_program_num); + glDeleteBuffers(1, &cbcr_vbo); BMUSBCapture::stop_bm_thread(); for (unsigned card_index = 0; card_index < num_cards; ++card_index) { @@ -245,23 +277,18 @@ 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 video_frame, size_t video_offset, VideoFormat video_format, FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format) { CaptureCard *card = &cards[card_index]; - 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; + int64_t frame_length = int64_t(TIMEBASE * video_format.frame_rate_den) / video_format.frame_rate_nom; size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0; if (num_samples > OUTPUT_FREQUENCY / 10) { printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n", card_index, int(audio_frame.len), int(audio_offset), - timecode, int(video_frame.len), int(video_offset), video_format); + timecode, int(video_frame.len), int(video_offset), video_format.id); if (video_frame.owner) { video_frame.owner->release_frame(video_frame); } @@ -288,7 +315,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, // 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; + 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", @@ -330,7 +357,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, if (card->should_quit) return; } - size_t expected_length = width * (height + extra_lines_top + extra_lines_bottom) * 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) { @@ -358,48 +385,48 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata; - unsigned num_fields = interlaced ? 2 : 1; + unsigned num_fields = video_format.interlaced ? 2 : 1; timespec frame_upload_start; - if (interlaced) { + 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(height % 2 == 0); - height /= 2; + 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 = interlaced; - userdata->last_frame_rate_nom = frame_rate_nom; - userdata->last_frame_rate_den = frame_rate_den; + userdata->last_interlaced = video_format.interlaced; + userdata->last_frame_rate_nom = video_format.frame_rate_nom; + userdata->last_frame_rate_den = video_format.frame_rate_den; RefCountedFrame new_frame(video_frame); // Upload the textures. - size_t cbcr_width = width / 2; + size_t cbcr_width = video_format.width / 2; size_t cbcr_offset = video_offset / 2; size_t y_offset = video_frame.size / 2 + video_offset / 2; for (unsigned field = 0; field < num_fields; ++field) { - unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22); + 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 || - width != userdata->last_width[field] || - height != userdata->last_height[field]) { + 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, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr); + 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, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr); + 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] = width; - userdata->last_height[field] = height; + userdata->last_width[field] = video_format.width; + userdata->last_height[field] = video_format.height; } GLuint pbo = userdata->pbo; @@ -411,14 +438,16 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, 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))); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t))); check_error(); glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]); check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line)); + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + video_format.width * field_start_line)); check_error(); glBindTexture(GL_TEXTURE_2D, 0); check_error(); + glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0); + check_error(); GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); check_error(); assert(fence != nullptr); @@ -448,7 +477,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, card->new_frame = new_frame; card->new_frame_length = frame_length; card->new_frame_field = field; - card->new_frame_interlaced = interlaced; + card->new_frame_interlaced = video_format.interlaced; card->new_data_ready_fence = fence; card->dropped_frames = dropped_frames; card->new_data_ready_changed.notify_all(); @@ -490,7 +519,6 @@ void Mixer::thread_func() for (unsigned card_index = 0; card_index < num_cards; ++card_index) { CaptureCard *card = &cards[card_index]; - card_copy[card_index].usb = card->usb; card_copy[card_index].new_data_ready = card->new_data_ready; card_copy[card_index].new_frame = card->new_frame; card_copy[card_index].new_frame_length = card->new_frame_length; @@ -509,6 +537,7 @@ void Mixer::thread_func() } // Resample the audio as needed, including from previously dropped frames. + assert(num_cards > 0); for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) { { // Signal to the audio thread to process this frame. @@ -526,7 +555,7 @@ void Mixer::thread_func() } if (audio_level_callback != nullptr) { - unique_lock lock(r128_mutex); + unique_lock lock(compressor_mutex); double loudness_s = r128.loudness_S(); double loudness_i = r128.integrated(); double loudness_range_low = r128.range_min(); @@ -534,7 +563,8 @@ void Mixer::thread_func() audio_level_callback(loudness_s, 20.0 * log10(peak), loudness_i, loudness_range_low, loudness_range_high, - last_gain_staging_db); + gain_staging_db, 20.0 * log10(final_makeup_gain), + correlation.get_correlation()); } for (unsigned card_index = 1; card_index < num_cards; ++card_index) { @@ -650,7 +680,7 @@ void Mixer::thread_func() 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)); + h264_encoder.reset(new H264Encoder(h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd)); } #if 0 @@ -689,6 +719,11 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) { vector samples_card; vector samples_out; + + // TODO: Allow mixing audio from several sources. + unsigned selected_audio_card = theme->map_signal(audio_source_channel); + assert(selected_audio_card < num_cards); + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { samples_card.resize(num_samples * 2); { @@ -697,8 +732,7 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) printf("Card %d reported previous underrun.\n", card_index); } } - // TODO: Allow using audio from the other card(s) as well. - if (card_index == 0) { + if (card_index == selected_audio_card) { samples_out = move(samples_card); } } @@ -707,7 +741,9 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) // we don't need it for voice, and it will reduce headroom // and confuse the compressor. (In particular, any hums at 50 or 60 Hz // should be dampened.) - locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f); + if (locut_enabled) { + locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f); + } // Apply a level compressor to get the general level right. // Basically, if it's over about -40 dBFS, we squeeze it down to that level @@ -716,13 +752,22 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) // 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 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 @@ -771,15 +816,57 @@ void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples) peak_resampler.process(); size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count; peak = max(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2)); + peak_resampler.out_data = nullptr; } - // Find R128 levels. + // At this point, we are most likely close to +0 LU, but all of our + // measurements have been on raw sample values, not R128 values. + // So we have a final makeup gain to get us to +0 LU; the gain + // adjustments required should be relatively small, and also, the + // offset shouldn't change much (only if the type of audio changes + // significantly). Thus, we shoot for updating this value basically + // “whenever we process buffers”, since the R128 calculation isn't exactly + // something we get out per-sample. + // + // Note that there's a feedback loop here, so we choose a very slow filter + // (half-time of 100 seconds). + double target_loudness_factor, alpha; + { + unique_lock lock(compressor_mutex); + double loudness_lu = r128.loudness_M() - ref_level_lufs; + double current_makeup_lu = 20.0f * log10(final_makeup_gain); + target_loudness_factor = pow(10.0f, -loudness_lu / 20.0f); + + // If we're outside +/- 5 LU uncorrected, we don't count it as + // a normal signal (probably silence) and don't change the + // correction factor; just apply what we already have. + if (fabs(loudness_lu - current_makeup_lu) >= 5.0 || !final_makeup_gain_auto) { + alpha = 0.0; + } else { + // Formula adapted from + // https://en.wikipedia.org/wiki/Low-pass_filter#Simple_infinite_impulse_response_filter. + const double half_time_s = 100.0; + const double fc_mul_2pi_delta_t = 1.0 / (half_time_s * OUTPUT_FREQUENCY); + alpha = fc_mul_2pi_delta_t / (fc_mul_2pi_delta_t + 1.0); + } + + double m = final_makeup_gain; + for (size_t i = 0; i < samples_out.size(); i += 2) { + samples_out[i + 0] *= m; + samples_out[i + 1] *= m; + m += (target_loudness_factor - m) * alpha; + } + final_makeup_gain = m; + } + + // Find R128 levels and L/R correlation. vector left, right; deinterleave_samples(samples_out, &left, &right); float *ptrs[] = { left.data(), right.data() }; { - unique_lock lock(r128_mutex); + unique_lock lock(compressor_mutex); r128.process(left.size(), ptrs); + correlation.process_samples(samples_out); } // Send the samples to the sound card. @@ -797,12 +884,6 @@ 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(); @@ -829,17 +910,28 @@ 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); @@ -884,6 +976,7 @@ void Mixer::reset_meters() peak = 0.0f; r128.reset(); r128.integr_start(); + correlation.reset(); } Mixer::OutputChannel::~OutputChannel()