X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=mixer.cpp;h=13fa29928ad8ecd9efc7f0d706c84621d51e761c;hb=0abd47d9bad672e4ffd882ed3597a3ea67d80eb8;hp=db5980d2d79bbdf30d3b005bd3394644120309a3;hpb=f4652550f7952d935640ca97bcd7bdbaa9820762;p=nageru diff --git a/mixer.cpp b/mixer.cpp index db5980d..13fa299 100644 --- a/mixer.cpp +++ b/mixer.cpp @@ -1,5 +1,3 @@ -#define EXTRAHEIGHT 30 - #undef Success #include "mixer.h" @@ -107,7 +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 + 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( @@ -212,57 +210,6 @@ void deinterleave_samples(const vector &in, vector *out_l, vector< } } -// Returns length of a frame with the given format, in TIMEBASE units. -int64_t find_frame_length(uint16_t video_format) -{ - if (video_format == 0x0800) { - // No video signal. These green pseudo-frames seem to come at about 30.13 Hz. - // It's a strange thing, but what can you do. - return TIMEBASE * 100 / 3013; - } - if ((video_format & 0xe800) != 0xe800) { - printf("Video format 0x%04x does not appear to be a video format. Assuming 60 Hz.\n", - video_format); - return TIMEBASE / 60; - } - - // 0x8 seems to be a flag about availability of deep color on the input, - // except when it's not (e.g. it's the only difference between NTSC 23.98 - // and PAL). Rather confusing. But we clear it here nevertheless, because - // usually it doesn't mean anything. - // - // We don't really handle interlaced formats at all yet. - uint16_t normalized_video_format = video_format & ~0xe808; - if (normalized_video_format == 0x0143) { // 720p50. - return TIMEBASE / 50; - } else if (normalized_video_format == 0x0103) { // 720p60. - return TIMEBASE / 60; - } else if (normalized_video_format == 0x0121) { // 720p59.94. - return TIMEBASE * 1001 / 60000; - } else if (normalized_video_format == 0x01c3 || // 1080p30. - normalized_video_format == 0x0003) { // 1080i60. - return TIMEBASE / 30; - } else if (normalized_video_format == 0x01e1 || // 1080p29.97. - normalized_video_format == 0x0021 || // 1080i59.94. - video_format == 0xe901 || // NTSC (480i59.94, I suppose). - video_format == 0xe9c1 || // Ditto. - video_format == 0xe801) { // Ditto. - return TIMEBASE * 1001 / 30000; - } else if (normalized_video_format == 0x0063 || // 1080p25. - normalized_video_format == 0x0043 || // 1080i50. - video_format == 0xe909) { // PAL (576i50, I suppose). - return TIMEBASE / 25; - } else if (normalized_video_format == 0x008e) { // 1080p24. - return TIMEBASE / 24; - } else if (normalized_video_format == 0x00a1) { // 1080p23.98. - return TIMEBASE * 1001 / 24000; - return TIMEBASE / 25; - } else { - printf("Unknown video format 0x%04x. Assuming 60 Hz.\n", video_format); - return TIMEBASE / 60; - } -} - } // namespace void Mixer::bm_frame(unsigned card_index, uint16_t timecode, @@ -271,9 +218,15 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { CaptureCard *card = &cards[card_index]; - int64_t frame_length = find_frame_length(video_format); + 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; - if (audio_frame.len - audio_offset > 30000) { + 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); @@ -291,10 +244,8 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, if (card->last_timecode != -1) { dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1; } - card->last_timecode = 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); @@ -303,29 +254,39 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode, { 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), resetting resampler\n", - card_index); + 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. (The number of samples could be nonintegral, - // but resampling will save us then.) + // 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 * frame_length / TIMEBASE) * 2); + silence.resize(silence_samples * 2); for (int i = 0; i < dropped_frames; ++i) { - card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence.size() / 2); + 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); @@ -338,7 +299,8 @@ 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); @@ -361,39 +323,108 @@ 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_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(); + 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_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; + } + } } } @@ -410,7 +441,7 @@ 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]; @@ -429,6 +460,7 @@ 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_data_ready_fence = card->new_data_ready_fence; card_copy[card_index].dropped_frames = card->dropped_frames; card->new_data_ready = false; @@ -437,6 +469,7 @@ void Mixer::thread_func() 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); } } @@ -452,7 +485,7 @@ void Mixer::thread_func() // 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. - ++dropped_frames; + ++stats_dropped_frames; pts_int += card_copy[0].new_frame_length; } } @@ -481,7 +514,7 @@ 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; + ++stats_dropped_frames; pts_int += card_copy[0].new_frame_length; continue; } @@ -504,7 +537,12 @@ void Mixer::thread_func() 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); + theme->set_input_textures( + card_index, + userdata->tex_y[card->new_frame_field], + userdata->tex_cbcr[card->new_frame_field], + userdata->last_width[card->new_frame_field], + userdata->last_height[card->new_frame_field]); } // Get the main chain from the theme, and set its state immediately. @@ -582,7 +620,7 @@ 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(); }