X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=1a4a3f9d3c292f4d4aaa85a4415c8420bc1083b7;hb=8ed6a8ebad548bc2c162100c70fa3d952fb0ce17;hp=47137022cabf181f28d2484866e8f8c49681aa17;hpb=44d0e2ac6f2c3b969244054efb2f80498756aeb1;p=nageru diff --git a/mixer.cpp b/mixer.cpp index 4713702..1a4a3f9 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,32 +1,22 @@ -#define WIDTH 1280 -#define HEIGHT 720 - #undef Success #include "mixer.h" #include -#include -#include -#include #include -#include -#include #include -#include -#include -#include -#include -#include +#include +#include +#include +#include +#include +#include #include #include #include #include #include -#include -#include -#include -#include +#include #include #include #include @@ -34,13 +24,16 @@ #include #include #include +#include #include -#include "bmusb.h" +#include "bmusb/bmusb.h" #include "context.h" +#include "defs.h" #include "h264encode.h" #include "pbo_frame_allocator.h" #include "ref_counted_gl_sync.h" +#include "timebase.h" class QOpenGLContext; @@ -50,34 +43,67 @@ using namespace std::placeholders; Mixer *global_mixer = nullptr; -Mixer::Mixer(const QSurfaceFormat &format) - : mixer_surface(create_surface(format)), - h264_encoder_surface(create_surface(format)) +namespace { + +void convert_fixed24_to_fp32(float *dst, size_t out_channels, const uint8_t *src, size_t in_channels, size_t num_samples) +{ + for (size_t i = 0; i < num_samples; ++i) { + for (size_t j = 0; j < out_channels; ++j) { + uint32_t s1 = *src++; + uint32_t s2 = *src++; + uint32_t s3 = *src++; + uint32_t s = s1 | (s1 << 8) | (s2 << 16) | (s3 << 24); + dst[i * out_channels + j] = int(s) * (1.0f / 4294967296.0f); + } + src += 3 * (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 + +Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards) + : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT), + num_cards(num_cards), + mixer_surface(create_surface(format)), + h264_encoder_surface(create_surface(format)), + level_compressor(OUTPUT_FREQUENCY), + limiter(OUTPUT_FREQUENCY), + compressor(OUTPUT_FREQUENCY) { + httpd.start(9095); + CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF)); check_error(); + // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision + // will be halved when sampling them, and we need to compensate here. + movit_texel_subpixel_precision /= 2.0; + resource_pool.reset(new ResourcePool); - theme.reset(new Theme("theme.lua", resource_pool.get())); - output_channel[OUTPUT_LIVE].parent = this; - output_channel[OUTPUT_PREVIEW].parent = this; - output_channel[OUTPUT_INPUT0].parent = this; - output_channel[OUTPUT_INPUT1].parent = this; + theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards)); + for (unsigned i = 0; i < NUM_OUTPUTS; ++i) { + output_channel[i].parent = this; + } ImageFormat inout_format; inout_format.color_space = COLORSPACE_sRGB; inout_format.gamma_curve = GAMMA_sRGB; - YCbCrFormat input_ycbcr_format; - input_ycbcr_format.chroma_subsampling_x = 2; - input_ycbcr_format.chroma_subsampling_y = 1; - input_ycbcr_format.cb_x_position = 0.0; - input_ycbcr_format.cr_x_position = 0.0; - input_ycbcr_format.cb_y_position = 0.5; - input_ycbcr_format.cr_y_position = 0.5; - input_ycbcr_format.luma_coefficients = YCBCR_REC_601; - input_ycbcr_format.full_range = false; - // Display chain; shows the live output produced by the main chain (its RGBA version). display_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get())); check_error(); @@ -87,53 +113,38 @@ Mixer::Mixer(const QSurfaceFormat &format) display_chain->set_dither_bits(0); // Don't bother. display_chain->finalize(); - // Preview chains (always shows just the inputs for now). - preview0_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get())); - check_error(); - preview0_input = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR); - preview0_chain->add_input(preview0_input); - preview0_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED); - preview0_chain->set_dither_bits(0); // Don't bother. - preview0_chain->finalize(); - - preview1_chain.reset(new EffectChain(WIDTH, HEIGHT, resource_pool.get())); - check_error(); - preview1_input = new YCbCrInput(inout_format, input_ycbcr_format, WIDTH, HEIGHT, YCBCR_INPUT_SPLIT_Y_AND_CBCR); - preview1_chain->add_input(preview1_input); - preview1_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED); - preview1_chain->set_dither_bits(0); // Don't bother. - preview1_chain->finalize(); - - h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, "test.mp4")); - - printf("Configuring first card...\n"); - cards[0].usb = new BMUSBCapture(0x1edb, 0xbd3b); // 0xbd4f - cards[0].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 0, _1, _2, _3, _4, _5, _6, _7)); - cards[0].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720)); - cards[0].usb->set_video_frame_allocator(cards[0].frame_allocator.get()); - cards[0].usb->configure_card(); - cards[0].surface = create_surface(format); -#if NUM_CARDS == 2 - cards[1].surface = create_surface(format); -#endif - - if (NUM_CARDS == 2) { - printf("Configuring second card...\n"); - cards[1].usb = new BMUSBCapture(0x1edb, 0xbd4f); - cards[1].usb->set_frame_callback(std::bind(&Mixer::bm_frame, this, 1, _1, _2, _3, _4, _5, _6, _7)); - cards[1].frame_allocator.reset(new PBOFrameAllocator(1280 * 750 * 2 + 44, 1280, 720)); - cards[1].usb->set_video_frame_allocator(cards[1].frame_allocator.get()); - cards[1].usb->configure_card(); + h264_encoder.reset(new H264Encoder(h264_encoder_surface, 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(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( + [card]{ + eglBindAPI(EGL_OPENGL_API); + card->context = create_context(card->surface); + if (!make_current(card->context, card->surface)) { + printf("failed to create bmusb context\n"); + exit(1); + } + }, + [this]{ + resource_pool->clean_context(); + }); + card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2)); + card->usb->configure_card(); } BMUSBCapture::start_bm_thread(); - for (int card_index = 0; card_index < NUM_CARDS; ++card_index) { + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { cards[card_index].usb->start_bm_capture(); } - //chain->enable_phase_timing(true); - // Set up stuff for NV12 conversion. // Cb/Cr shader. @@ -145,148 +156,298 @@ Mixer::Mixer(const QSurfaceFormat &format) "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(); + + locut.init(FILTER_HPF, 2); + + // 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() { resource_pool->release_glsl_program(cbcr_program_num); 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].usb->stop_dequeue_thread(); + } + + h264_encoder.reset(nullptr); +} + +namespace { + +int unwrap_timecode(uint16_t current_wrapped, int last) +{ + uint16_t last_wrapped = last & 0xffff; + if (current_wrapped > last_wrapped) { + return (last & ~0xffff) | current_wrapped; + } else { + return 0x10000 + ((last & ~0xffff) | current_wrapped); + } +} + +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 = max(m, fabs(samples[i])); + } + return m; } -void Mixer::bm_frame(int card_index, uint16_t timecode, +void deinterleave_samples(const vector &in, vector *out_l, vector *out_r) +{ + size_t num_samples = in.size() / 2; + out_l->resize(num_samples); + out_r->resize(num_samples); + + const float *inptr = in.data(); + float *lptr = &(*out_l)[0]; + float *rptr = &(*out_r)[0]; + for (size_t i = 0; i < num_samples; ++i) { + *lptr++ = *inptr++; + *rptr++ = *inptr++; + } +} + +} // 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) { CaptureCard *card = &cards[card_index]; - if (!card->thread_initialized) { - printf("initializing context for bmusb thread %d\n", card_index); - eglBindAPI(EGL_OPENGL_API); - card->context = create_context(); - if (!make_current(card->context, card->surface)) { - printf("failed to create bmusb context\n"); - exit(1); + + 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); + if (video_frame.owner) { + video_frame.owner->release_frame(video_frame); + } + if (audio_frame.owner) { + audio_frame.owner->release_frame(audio_frame); } - card->thread_initialized = true; - } - - if (video_frame.len - video_offset != 1280 * 750 * 2) { - printf("dropping frame with wrong length (%ld)\n", video_frame.len - video_offset); - FILE *fp = fopen("frame.raw", "wb"); - fwrite(video_frame.data, video_frame.len, 1, fp); - fclose(fp); - //exit(1); - card->usb->get_video_frame_allocator()->release_frame(video_frame); - card->usb->get_audio_frame_allocator()->release_frame(audio_frame); return; } + + int64_t local_pts = card->next_local_pts; + int dropped_frames = 0; + if (card->last_timecode != -1) { + dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1; + } + + // Convert the audio to stereo fp32 and add it. + vector audio; + audio.resize(num_samples * 2); + convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples); + + // Add the audio. { - // Wait until the previous frame was consumed. - std::unique_lock lock(bmusb_mutex); - card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready; }); + unique_lock lock(card->audio_mutex); + + // Number of samples per frame if we need to insert silence. + // (Could be nonintegral, but resampling will save us then.) + int silence_samples = OUTPUT_FREQUENCY * 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_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 + // no way of detecting that; we just have to assume the frame length + // is always the same. + local_pts += frame_length; + } + } + if (num_samples == 0) { + audio.resize(silence_samples * 2); + num_samples = silence_samples; + } + card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples); + card->next_local_pts = local_pts + frame_length; + } + + card->last_timecode = timecode; + + // Done with the audio, so release it. + if (audio_frame.owner) { + audio_frame.owner->release_frame(audio_frame); } - 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(); - // Upload the textures. - glBindTexture(GL_TEXTURE_2D, userdata->tex_y); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280, 720, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((1280 * 750 * 2 + 44) / 2 + 1280 * 25 + 22)); - check_error(); - glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr); - check_error(); - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 1280/2, 720, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(1280 * 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); { - std::unique_lock lock(bmusb_mutex); - card->new_data_ready = true; - card->new_frame = RefCountedFrame(video_frame); - card->new_data_ready_fence = fence; - card->new_data_ready_changed.notify_all(); + // 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; } - // Video frame will be released when last user of card->new_frame goes out of scope. - card->usb->get_audio_frame_allocator()->release_frame(audio_frame); -} - -void Mixer::place_rectangle(Effect *resample_effect, Effect *padding_effect, float x0, float y0, float x1, float y1) -{ - float srcx0 = 0.0f; - float srcx1 = 1.0f; - float srcy0 = 0.0f; - float srcy1 = 1.0f; - - // Cull. - if (x0 > 1280.0 || x1 < 0.0 || y0 > 720.0 || y1 < 0.0) { - CHECK(resample_effect->set_int("width", 1)); - CHECK(resample_effect->set_int("height", 1)); - CHECK(resample_effect->set_float("zoom_x", 1280.0)); - CHECK(resample_effect->set_float("zoom_y", 720.0)); - CHECK(padding_effect->set_int("left", 2000)); - CHECK(padding_effect->set_int("top", 2000)); - return; - } - - // Clip. (TODO: Clip on upper/left sides, too.) - if (x1 > 1280.0) { - srcx1 = (1280.0 - x0) / (x1 - x0); - x1 = 1280.0; - } - if (y1 > 720.0) { - srcy1 = (720.0 - y0) / (y1 - y0); - y1 = 720.0; - } - - float x_subpixel_offset = x0 - floor(x0); - float y_subpixel_offset = y0 - floor(y0); - - // Resampling must be to an integral number of pixels. Round up, - // and then add an extra pixel so we have some leeway for the border. - int width = int(ceil(x1 - x0)) + 1; - int height = int(ceil(y1 - y0)) + 1; - CHECK(resample_effect->set_int("width", width)); - CHECK(resample_effect->set_int("height", height)); - - // Correct the discrepancy with zoom. (This will leave a small - // excess edge of pixels and subpixels, which we'll correct for soon.) - float zoom_x = (x1 - x0) / (width * (srcx1 - srcx0)); - float zoom_y = (y1 - y0) / (height * (srcy1 - srcy0)); - CHECK(resample_effect->set_float("zoom_x", zoom_x)); - CHECK(resample_effect->set_float("zoom_y", zoom_y)); - CHECK(resample_effect->set_float("zoom_center_x", 0.0f)); - CHECK(resample_effect->set_float("zoom_center_y", 0.0f)); - - // Padding must also be to a whole-pixel offset. - CHECK(padding_effect->set_int("left", floor(x0))); - CHECK(padding_effect->set_int("top", floor(y0))); - - // Correct _that_ discrepancy by subpixel offset in the resampling. - CHECK(resample_effect->set_float("left", -x_subpixel_offset / zoom_x)); - CHECK(resample_effect->set_float("top", -y_subpixel_offset / zoom_y)); - - // Finally, adjust the border so it is exactly where we want it. - CHECK(padding_effect->set_float("border_offset_left", x_subpixel_offset)); - CHECK(padding_effect->set_float("border_offset_right", x1 - (floor(x0) + width))); - CHECK(padding_effect->set_float("border_offset_top", y_subpixel_offset)); - CHECK(padding_effect->set_float("border_offset_bottom", y1 - (floor(y0) + height))); + 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; expected %ld)\n", + card_index, video_frame.len - video_offset, expected_length); + } + if (video_frame.owner) { + video_frame.owner->release_frame(video_frame); + } + + // Still send on the information that we _had_ a frame, even though it's corrupted, + // 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_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(); + } + return; + } + + 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. + 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; + } + + 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; + } + } + } } - + void Mixer::thread_func() { eglBindAPI(EGL_OPENGL_API); - QOpenGLContext *context = create_context(); + QOpenGLContext *context = create_context(mixer_surface); if (!make_current(context, mixer_surface)) { printf("oops\n"); exit(1); @@ -295,114 +456,111 @@ void Mixer::thread_func() struct timespec start, now; clock_gettime(CLOCK_MONOTONIC, &start); - while (!should_quit) { - ++frame; - - //int width0 = lrintf(848 * (1.0 + 0.2 * sin(frame * 0.02))); - int width0 = 848; - int height0 = lrintf(width0 * 9.0 / 16.0); - - //float top0 = 96 + 48 * sin(frame * 0.005); - //float left0 = 96 + 48 * cos(frame * 0.006); - float top0 = 48; - float left0 = 16; - float bottom0 = top0 + height0; - float right0 = left0 + width0; - - int width1 = 384; - int height1 = 216; - - float bottom1 = 720 - 48; - float right1 = 1280 - 16; - float top1 = bottom1 - height1; - float left1 = right1 - width1; - - if (current_source == SOURCE_INPUT1) { - top0 = 0.0; - bottom0 = HEIGHT; - left0 = 0.0; - right0 = WIDTH; - - top1 = HEIGHT + 10; - bottom1 = HEIGHT + 20; - left1 = WIDTH + 10; - right1 = WIDTH + 20; - } else if (current_source == SOURCE_INPUT2) { - top1 = 0.0; - bottom1 = HEIGHT; - left1 = 0.0; - right1 = WIDTH; - - top0 = HEIGHT + 10; - bottom0 = HEIGHT + 20; - left0 = WIDTH + 10; - right0 = WIDTH + 20; - } else { - float t = 0.5 + 0.5 * cos(frame * 0.006); - float scale0 = 1.0 + t * (1280.0 / 848.0 - 1.0); - float tx0 = 0.0 + t * (-16.0 * scale0); - float ty0 = 0.0 + t * (-48.0 * scale0); - - top0 = top0 * scale0 + ty0; - bottom0 = bottom0 * scale0 + ty0; - left0 = left0 * scale0 + tx0; - right0 = right0 * scale0 + tx0; - - top1 = top1 * scale0 + ty0; - bottom1 = bottom1 * scale0 + ty0; - left1 = left1 * scale0 + tx0; - right1 = right1 * scale0 + tx0; - } - -#if 0 - place_rectangle(resample_effect, padding_effect, left0, top0, right0, bottom0); - place_rectangle(resample2_effect, padding2_effect, left1, top1, right1, bottom1); -#endif + int frame = 0; + int stats_dropped_frames = 0; - CaptureCard card_copy[NUM_CARDS]; + while (!should_quit) { + CaptureCard card_copy[MAX_CARDS]; + int num_samples[MAX_CARDS]; { - std::unique_lock lock(bmusb_mutex); + unique_lock lock(bmusb_mutex); // 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 (int card_index = 0; card_index < NUM_CARDS; ++card_index) { + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { CaptureCard *card = &cards[card_index]; card_copy[card_index].usb = card->usb; card_copy[card_index].new_data_ready = card->new_data_ready; card_copy[card_index].new_frame = card->new_frame; + card_copy[card_index].new_frame_length = card->new_frame_length; + card_copy[card_index].new_frame_field = card->new_frame_field; + card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced; card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence; + card_copy[card_index].dropped_frames = card->dropped_frames; card->new_data_ready = false; 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) { + { + // 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. 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(); + 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 (card_copy[card_index].dropped_frames > 0) { + printf("Card %u dropped %d frames before this\n", + card_index, int(card_copy[card_index].dropped_frames)); } } - for (int card_index = 0; card_index < NUM_CARDS; ++card_index) { + // 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) { + ++stats_dropped_frames; + pts_int += card_copy[0].new_frame_length; + continue; + } + + for (unsigned card_index = 0; card_index < num_cards; ++card_index) { CaptureCard *card = &card_copy[card_index]; - if (!card->new_data_ready) + if (!card->new_data_ready || card->new_frame->len == 0) 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, // tell the GPU to wait until it's there. - if (card->new_data_ready_fence) + 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(); - const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata; - theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr); + check_error(); + glDeleteSync(card->new_data_ready_fence); + check_error(); + } } // Get the main chain from the theme, and set its state immediately. - pair> theme_main_chain = theme->get_chain(0, frame / 60.0f, 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); @@ -412,6 +570,7 @@ void Mixer::thread_func() 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); @@ -428,14 +587,10 @@ 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 (int card_index = 0; card_index < NUM_CARDS; ++card_index) { - input_frames.push_back(bmusb_current_rendering_frame[card_index]); - } - h264_encoder->end_frame(fence, input_frames); + 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, theme_main_chain.input_frames); + ++frame; + pts_int += card_copy[0].new_frame_length; // The live frame just shows the RGBA texture we just rendered. // It owns rgba_tex now. @@ -449,53 +604,29 @@ void Mixer::thread_func() live_frame.temp_textures = { rgba_tex }; output_channel[OUTPUT_LIVE].output_frame(live_frame); - // The preview frame shows the first input. Note that the textures - // are owned by the input frame, not the display frame. - if (bmusb_current_rendering_frame[0] != nullptr) { - const PBOFrameAllocator::Userdata *input0_userdata = (const PBOFrameAllocator::Userdata *)bmusb_current_rendering_frame[0]->userdata; - GLuint input0_y_tex = input0_userdata->tex_y; - GLuint input0_cbcr_tex = input0_userdata->tex_cbcr; - DisplayFrame preview0_frame; - preview0_frame.chain = preview0_chain.get(); - preview0_frame.setup_chain = [this, input0_y_tex, input0_cbcr_tex]{ - preview0_input->set_texture_num(0, input0_y_tex); - preview0_input->set_texture_num(1, input0_cbcr_tex); - }; - preview0_frame.ready_fence = fence; - preview0_frame.input_frames = { bmusb_current_rendering_frame[0] }; - preview0_frame.temp_textures = {}; - output_channel[OUTPUT_PREVIEW].output_frame(preview0_frame); - output_channel[OUTPUT_INPUT0].output_frame(preview0_frame); - } - - // Same for the other preview. - // TODO: Use a for loop. Gah. - if (bmusb_current_rendering_frame[1] != nullptr) { - const PBOFrameAllocator::Userdata *input1_userdata = (const PBOFrameAllocator::Userdata *)bmusb_current_rendering_frame[1]->userdata; - GLuint input1_y_tex = input1_userdata->tex_y; - GLuint input1_cbcr_tex = input1_userdata->tex_cbcr; - DisplayFrame preview1_frame; - preview1_frame.chain = preview1_chain.get(); - preview1_frame.setup_chain = [this, input1_y_tex, input1_cbcr_tex]{ - preview1_input->set_texture_num(0, input1_y_tex); - preview1_input->set_texture_num(1, input1_cbcr_tex); - }; - preview1_frame.ready_fence = fence; - preview1_frame.input_frames = { bmusb_current_rendering_frame[1] }; - preview1_frame.temp_textures = {}; - output_channel[OUTPUT_INPUT1].output_frame(preview1_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); } 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 in %.3f seconds = %.1f fps (%.1f ms/frame)\n", - frame, elapsed, frame / elapsed, + printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n", + frame, stats_dropped_frames, elapsed, frame / elapsed, 1e3 * elapsed / frame); // chain->print_phase_timing(); } +#if 0 // Reset every 100 frames, so that local variations in frame times // (especially for the first few frames, when the shaders are // compiled etc.) don't make it hard to measure for the entire @@ -504,8 +635,134 @@ void Mixer::thread_func() frame = 0; start = now; } +#endif check_error(); } + + resource_pool->clean_context(); +} + +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(num_samples * 2); + { + unique_lock lock(cards[card_index].audio_mutex); + 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); + } + } + // TODO: Allow using audio from the other card(s) as well. + if (card_index == 0) { + samples_out = move(samples_card); + } + } + + // 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. + // Basically, if it's over about -40 dBFS, we squeeze it down to that level + // (or more precisely, near it, since we don't use infinite ratio), + // 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); + } + +#if 0 + printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n", + level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()), + level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()), + 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain)); +#endif + +// float limiter_att, compressor_att; + + // 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. + if (limiter_enabled) { + float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f); + float ratio = 30.0f; + float attack_time = 0.0f; // Instant. + float release_time = 0.020f; + float makeup_gain = 1.0f; // 0 dB. + limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain); +// limiter_att = limiter.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. + peak_resampler.inp_data = samples_out.data(); + peak_resampler.inp_count = samples_out.size() / 2; + + vector interpolated_samples_out; + interpolated_samples_out.resize(samples_out.size()); + while (peak_resampler.inp_count > 0) { // About four iterations. + peak_resampler.out_data = &interpolated_samples_out[0]; + peak_resampler.out_count = interpolated_samples_out.size() / 2; + 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)); + } + + // Find R128 levels. + vector left, right; + deinterleave_samples(samples_out, &left, &right); + float *ptrs[] = { left.data(), right.data() }; + { + unique_lock lock(r128_mutex); + r128.process(left.size(), ptrs); + } + + // Send the samples to the sound card. + if (alsa) { + alsa->write(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) @@ -574,18 +831,43 @@ void Mixer::release_display_frame(DisplayFrame *frame) void Mixer::start() { - mixer_thread = std::thread(&Mixer::thread_func, this); + 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) +{ + theme->transition_clicked(transition_num, pts()); +} + +void Mixer::channel_clicked(int preview_num) +{ + theme->channel_clicked(preview_num); +} + +void Mixer::reset_meters() +{ + peak_resampler.reset(); + peak = 0.0f; + r128.reset(); + r128.integr_start(); } -void Mixer::cut(Source source) +Mixer::OutputChannel::~OutputChannel() { - current_source = source; + if (has_current_frame) { + parent->release_display_frame(¤t_frame); + } + if (has_ready_frame) { + parent->release_display_frame(&ready_frame); + } } void Mixer::OutputChannel::output_frame(DisplayFrame frame) @@ -593,7 +875,7 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) // Store this frame for display. Remove the ready frame if any // (it was seemingly never used). { - std::unique_lock lock(frame_mutex); + unique_lock lock(frame_mutex); if (has_ready_frame) { parent->release_display_frame(&ready_frame); } @@ -608,7 +890,7 @@ void Mixer::OutputChannel::output_frame(DisplayFrame frame) bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame) { - std::unique_lock lock(frame_mutex); + unique_lock lock(frame_mutex); if (!has_current_frame && !has_ready_frame) { return false; }