]> git.sesse.net Git - nageru/blobdiff - mixer.cpp
Make the API for begin_frame()/end_frame() in VideoEncoder a bit more sensible.
[nageru] / mixer.cpp
index 1be5660fb7f9219d17a1a39a15841f6d0c5c47b5..79d8e97eb9b9391cda7ceaf2ad686719a9f00caf 100644 (file)
--- a/mixer.cpp
+++ b/mixer.cpp
@@ -10,6 +10,7 @@
 #include <movit/image_format.h>
 #include <movit/init.h>
 #include <movit/resource_pool.h>
+#include <pthread.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -35,6 +36,7 @@
 #include "chroma_subsampler.h"
 #include "context.h"
 #include "decklink_capture.h"
+#include "decklink_output.h"
 #include "defs.h"
 #include "disk_space_estimator.h"
 #include "flags.h"
@@ -43,6 +45,8 @@
 #include "ref_counted_gl_sync.h"
 #include "resampling_queue.h"
 #include "timebase.h"
+#include "timecode_renderer.h"
+#include "v210_converter.h"
 #include "video_encoder.h"
 
 class IDeckLink;
@@ -74,27 +78,101 @@ void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced
        }
 }
 
+void ensure_texture_resolution(PBOFrameAllocator::Userdata *userdata, unsigned field, unsigned width, unsigned height, unsigned v210_width)
+{
+       bool first;
+       if (global_flags.ten_bit_input) {
+               first = userdata->tex_v210[field] == 0 || userdata->tex_444[field] == 0;
+       } else {
+               first = userdata->tex_y[field] == 0 || userdata->tex_cbcr[field] == 0;
+       }
+
+       if (first ||
+           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.
+               if (global_flags.ten_bit_input) {
+                       glBindTexture(GL_TEXTURE_2D, userdata->tex_444[field]);
+                       check_error();
+                       glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
+                       check_error();
+               } else {
+                       size_t cbcr_width = width / 2;
+
+                       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;
+       }
+       if (global_flags.ten_bit_input &&
+           (first || v210_width != userdata->last_v210_width[field])) {
+               // Same as above; we need to recreate the texture.
+               glBindTexture(GL_TEXTURE_2D, userdata->tex_v210[field]);
+               check_error();
+               glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
+               check_error();
+               userdata->last_v210_width[field] = v210_width;
+       }
+}
+
+void upload_texture(GLuint tex, GLuint width, GLuint height, GLuint stride, bool interlaced_stride, GLenum format, GLenum type, GLintptr offset)
+{
+       if (interlaced_stride) {
+               stride *= 2;
+       }
+       if (global_flags.flush_pbos) {
+               glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, offset, stride * height);
+               check_error();
+       }
+
+       glBindTexture(GL_TEXTURE_2D, tex);
+       check_error();
+       if (interlaced_stride) {
+               glPixelStorei(GL_UNPACK_ROW_LENGTH, width * 2);
+               check_error();
+       } else {
+               glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+               check_error();
+       }
+
+       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, format, type, BUFFER_OFFSET(offset));
+       check_error();
+       glBindTexture(GL_TEXTURE_2D, 0);
+       check_error();
+       glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
+       check_error();
+}
+
 }  // namespace
 
-void QueueLengthPolicy::update_policy(int queue_length)
+void QueueLengthPolicy::update_policy(unsigned queue_length)
 {
-       if (queue_length < 0) {  // Starvation.
-               if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
+       if (queue_length == 0) {  // Starvation.
+               if (been_at_safe_point_since_last_starvation && safe_queue_length < unsigned(global_flags.max_input_queue_frames)) {
                        ++safe_queue_length;
-                       fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\n",
+                       fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frame(s)\n",
                                card_index, safe_queue_length);
                }
                frames_with_at_least_one = 0;
                been_at_safe_point_since_last_starvation = false;
                return;
        }
-       if (queue_length > 0) {
-               if (queue_length >= int(safe_queue_length)) {
+       if (queue_length >= 1) {
+               if (queue_length >= safe_queue_length) {
                        been_at_safe_point_since_last_starvation = true;
                }
-               if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
+               if (++frames_with_at_least_one >= 1000 && safe_queue_length > 1) {
                        --safe_queue_length;
-                       fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
+                       fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frame(s)\n",
                                card_index, safe_queue_length);
                        frames_with_at_least_one = 0;
                }
@@ -108,6 +186,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
          num_cards(num_cards),
          mixer_surface(create_surface(format)),
          h264_encoder_surface(create_surface(format)),
+         decklink_output_surface(create_surface(format)),
          audio_mixer(num_cards)
 {
        CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
@@ -156,7 +235,10 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
                                        break;
                                }
 
-                               configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
+                               DeckLinkCapture *capture = new DeckLinkCapture(decklink, card_index);
+                               DeckLinkOutput *output = new DeckLinkOutput(resource_pool.get(), decklink_output_surface, global_flags.width, global_flags.height, card_index);
+                               output->set_device(decklink);
+                               configure_card(card_index, capture, /*is_fake_capture=*/false, output);
                                ++num_pci_devices;
                        }
                        decklink_iterator->Release();
@@ -165,18 +247,19 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
                        fprintf(stderr, "DeckLink drivers not found. Probing for USB cards only.\n");
                }
        }
+
        unsigned num_usb_devices = BMUSBCapture::num_cards();
        for (unsigned usb_card_index = 0; usb_card_index < num_usb_devices && card_index < num_cards; ++usb_card_index, ++card_index) {
                BMUSBCapture *capture = new BMUSBCapture(usb_card_index);
                capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, card_index));
-               configure_card(card_index, capture, /*is_fake_capture=*/false);
+               configure_card(card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
        }
        fprintf(stderr, "Found %u USB card(s).\n", num_usb_devices);
 
        unsigned num_fake_cards = 0;
        for ( ; card_index < num_cards; ++card_index, ++num_fake_cards) {
                FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
-               configure_card(card_index, capture, /*is_fake_capture=*/true);
+               configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
        }
 
        if (num_fake_cards > 0) {
@@ -192,9 +275,35 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 
        chroma_subsampler.reset(new ChromaSubsampler(resource_pool.get()));
 
+       if (global_flags.ten_bit_input) {
+               if (!v210Converter::has_hardware_support()) {
+                       fprintf(stderr, "ERROR: --ten-bit-input requires support for OpenGL compute shaders\n");
+                       fprintf(stderr, "       (OpenGL 4.3, or GL_ARB_compute_shader + GL_ARB_shader_image_load_store).\n");
+                       exit(1);
+               }
+               v210_converter.reset(new v210Converter());
+
+               // These are all the widths listed in the Blackmagic SDK documentation
+               // (section 2.7.3, “Display Modes”).
+               v210_converter->precompile_shader(720);
+               v210_converter->precompile_shader(1280);
+               v210_converter->precompile_shader(1920);
+               v210_converter->precompile_shader(2048);
+               v210_converter->precompile_shader(3840);
+               v210_converter->precompile_shader(4096);
+       }
+
+       timecode_renderer.reset(new TimecodeRenderer(resource_pool.get(), global_flags.width, global_flags.height));
+       display_timecode_in_stream = global_flags.display_timecode_in_stream;
+       display_timecode_on_stdout = global_flags.display_timecode_on_stdout;
+
        if (global_flags.enable_alsa_output) {
                alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
        }
+       if (global_flags.output_card != -1) {
+               desired_output_card_index = global_flags.output_card;
+               set_output_card_internal(global_flags.output_card);
+       }
 }
 
 Mixer::~Mixer()
@@ -203,27 +312,33 @@ Mixer::~Mixer()
 
        for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                {
-                       unique_lock<mutex> lock(bmusb_mutex);
+                       unique_lock<mutex> lock(card_mutex);
                        cards[card_index].should_quit = true;  // Unblock thread.
                        cards[card_index].new_frames_changed.notify_all();
                }
                cards[card_index].capture->stop_dequeue_thread();
+               if (cards[card_index].output) {
+                       cards[card_index].output->end_output();
+                       cards[card_index].output.reset();
+               }
        }
 
        video_encoder.reset(nullptr);
 }
 
-void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
+void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture, DeckLinkOutput *output)
 {
        printf("Configuring card %d...\n", card_index);
 
        CaptureCard *card = &cards[card_index];
        if (card->capture != nullptr) {
                card->capture->stop_dequeue_thread();
-               delete card->capture;
        }
-       card->capture = capture;
+       card->capture.reset(capture);
        card->is_fake_capture = is_fake_capture;
+       if (card->output.get() != output) {
+               card->output.reset(output);
+       }
        card->capture->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
        if (card->frame_allocator == nullptr) {
                card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, global_flags.width, global_flags.height));  // 8 MB.
@@ -232,16 +347,60 @@ void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool
        if (card->surface == nullptr) {
                card->surface = create_surface_with_same_format(mixer_surface);
        }
-       while (!card->new_frames.empty()) card->new_frames.pop();
+       while (!card->new_frames.empty()) card->new_frames.pop_front();
        card->last_timecode = -1;
+       card->capture->set_pixel_format(global_flags.ten_bit_input ? PixelFormat_10BitYCbCr : PixelFormat_8BitYCbCr);
        card->capture->configure_card();
 
+       // NOTE: start_bm_capture() happens in thread_func().
+
        DeviceSpec device{InputSourceType::CAPTURE_CARD, card_index};
        audio_mixer.reset_resampler(device);
        audio_mixer.set_display_name(device, card->capture->get_description());
        audio_mixer.trigger_state_changed_callback();
 }
 
+void Mixer::set_output_card_internal(int card_index)
+{
+       // We don't really need to take card_mutex, since we're in the mixer
+       // thread and don't mess with any queues (which is the only thing that happens
+       // from other threads), but it's probably the safest in the long run.
+       unique_lock<mutex> lock(card_mutex);
+       if (output_card_index != -1) {
+               // Switch the old card from output to input.
+               CaptureCard *old_card = &cards[output_card_index];
+               old_card->output->end_output();
+
+               // Stop the fake card that we put into place.
+               // This needs to _not_ happen under the mutex, to avoid deadlock
+               // (delivering the last frame needs to take the mutex).
+               bmusb::CaptureInterface *fake_capture = old_card->capture.get();
+               lock.unlock();
+               fake_capture->stop_dequeue_thread();
+               lock.lock();
+               old_card->capture = move(old_card->parked_capture);
+               old_card->is_fake_capture = false;
+               old_card->capture->start_bm_capture();
+       }
+       if (card_index != -1) {
+               CaptureCard *card = &cards[card_index];
+               bmusb::CaptureInterface *capture = card->capture.get();
+               // TODO: DeckLinkCapture::stop_dequeue_thread can actually take
+               // several seconds to complete (blocking on DisableVideoInput);
+               // see if we can maybe do it asynchronously.
+               lock.unlock();
+               capture->stop_dequeue_thread();
+               lock.lock();
+               card->parked_capture = move(card->capture);
+               bmusb::CaptureInterface *fake_capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
+               configure_card(card_index, fake_capture, /*is_fake_capture=*/true, card->output.release());
+               card->queue_length_policy.reset(card_index);
+               card->capture->start_bm_capture();
+               desired_output_video_mode = output_video_mode = card->output->pick_video_mode(desired_output_video_mode);
+               card->output->start_output(desired_output_video_mode, pts_int);
+       }
+       output_card_index = card_index;
+}
 
 namespace {
 
@@ -324,7 +483,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                } while (!success);
        }
 
-       audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);
+       audio_mixer.add_audio(device, audio_frame.data + audio_offset, num_samples, audio_format, frame_length, audio_frame.received_timestamp);
 
        // Done with the audio, so release it.
        if (audio_frame.owner) {
@@ -333,7 +492,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
 
        card->last_timecode = timecode;
 
-       size_t expected_length = video_format.width * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom) * 2;
+       size_t expected_length = video_format.stride * (video_format.height + video_format.extra_lines_top + video_format.extra_lines_bottom);
        if (video_frame.len - video_offset == 0 ||
            video_frame.len - video_offset != expected_length) {
                if (video_frame.len != 0) {
@@ -347,13 +506,14 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                // Still send on the information that we _had_ a frame, even though it's corrupted,
                // so that pts can go up accordingly.
                {
-                       unique_lock<mutex> lock(bmusb_mutex);
+                       unique_lock<mutex> lock(card_mutex);
                        CaptureCard::NewFrame new_frame;
                        new_frame.frame = RefCountedFrame(FrameAllocator::Frame());
                        new_frame.length = frame_length;
                        new_frame.interlaced = false;
                        new_frame.dropped_frames = dropped_frames;
-                       card->new_frames.push(move(new_frame));
+                       new_frame.received_timestamp = video_frame.received_timestamp;
+                       card->new_frames.push_back(move(new_frame));
                        card->new_frames_changed.notify_all();
                }
                return;
@@ -385,9 +545,9 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
        RefCountedFrame frame(video_frame);
 
        // Upload the textures.
-       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;
+       const size_t cbcr_width = video_format.width / 2;
+       const size_t cbcr_offset = video_offset / 2;
+       const size_t y_offset = video_frame.size / 2 + video_offset / 2;
 
        for (unsigned field = 0; field < num_fields; ++field) {
                // Put the actual texture upload in a lambda that is executed in the main thread.
@@ -398,7 +558,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                // Note that this means we must hold on to the actual frame data in <userdata>
                // until the upload command is run, but we hold on to <frame> much longer than that
                // (in fact, all the way until we no longer use the texture in rendering).
-               auto upload_func = [field, video_format, y_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
+               auto upload_func = [this, field, video_format, y_offset, video_offset, cbcr_offset, cbcr_width, interlaced_stride, userdata]() {
                        unsigned field_start_line;
                        if (field == 1) {
                                field_start_line = video_format.second_field_start;
@@ -406,68 +566,28 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                                field_start_line = video_format.extra_lines_top;
                        }
 
-                       if (userdata->tex_y[field] == 0 ||
-                           userdata->tex_cbcr[field] == 0 ||
-                           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, 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, video_format.width, video_format.height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
-                               check_error();
-                               userdata->last_width[field] = video_format.width;
-                               userdata->last_height[field] = video_format.height;
-                       }
+                       // For 8-bit input, v210_width will be nonsensical but not used.
+                       size_t v210_width = video_format.stride / sizeof(uint32_t);
+                       ensure_texture_resolution(userdata, field, video_format.width, video_format.height, v210_width);
 
-                       GLuint pbo = userdata->pbo;
-                       check_error();
-                       glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo);
+                       glBindBuffer(GL_PIXEL_UNPACK_BUFFER, userdata->pbo);
                        check_error();
 
-                       size_t field_y_start = y_offset + video_format.width * field_start_line;
-                       size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
+                       if (global_flags.ten_bit_input) {
+                               size_t field_start = video_offset + video_format.stride * field_start_line;
+                               upload_texture(userdata->tex_v210[field], v210_width, video_format.height, video_format.stride, interlaced_stride, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, field_start);
+                               v210_converter->convert(userdata->tex_v210[field], userdata->tex_444[field], video_format.width, video_format.height);
+                       } else {
+                               size_t field_y_start = y_offset + video_format.width * field_start_line;
+                               size_t field_cbcr_start = cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t);
 
-                       if (global_flags.flush_pbos) {
-                               glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_y_start, video_format.width * video_format.height);
-                               check_error();
-                               glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, field_cbcr_start, cbcr_width * video_format.height * sizeof(uint16_t));
-                               check_error();
+                               // Make up our own strides, since we are interleaving.
+                               upload_texture(userdata->tex_y[field], video_format.width, video_format.height, video_format.width, interlaced_stride, GL_RED, GL_UNSIGNED_BYTE, field_y_start);
+                               upload_texture(userdata->tex_cbcr[field], cbcr_width, video_format.height, cbcr_width * sizeof(uint16_t), interlaced_stride, GL_RG, GL_UNSIGNED_BYTE, field_cbcr_start);
                        }
 
-                       glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
-                       check_error();
-                       if (interlaced_stride) {
-                               glPixelStorei(GL_UNPACK_ROW_LENGTH, cbcr_width * 2);
-                               check_error();
-                       } else {
-                               glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-                               check_error();
-                       }
-                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, video_format.height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_cbcr_start));
-                       check_error();
-                       glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
-                       check_error();
-                       if (interlaced_stride) {
-                               glPixelStorei(GL_UNPACK_ROW_LENGTH, video_format.width * 2);
-                               check_error();
-                       } else {
-                               glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-                               check_error();
-                       }
-                       glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, video_format.width, video_format.height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(field_y_start));
-                       check_error();
-                       glBindTexture(GL_TEXTURE_2D, 0);
-                       check_error();
                        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
                        check_error();
-                       glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-                       check_error();
                };
 
                if (field == 1) {
@@ -483,7 +603,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                }
 
                {
-                       unique_lock<mutex> lock(bmusb_mutex);
+                       unique_lock<mutex> lock(card_mutex);
                        CaptureCard::NewFrame new_frame;
                        new_frame.frame = frame;
                        new_frame.length = frame_length;
@@ -492,7 +612,7 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                        new_frame.upload_func = upload_func;
                        new_frame.dropped_frames = dropped_frames;
                        new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
-                       card->new_frames.push(move(new_frame));
+                       card->new_frames.push_back(move(new_frame));
                        card->new_frames_changed.notify_all();
                }
        }
@@ -511,6 +631,8 @@ void Mixer::bm_hotplug_remove(unsigned card_index)
 
 void Mixer::thread_func()
 {
+       pthread_setname_np(pthread_self(), "Mixer_OpenGL");
+
        eglBindAPI(EGL_OPENGL_API);
        QOpenGLContext *context = create_context(mixer_surface);
        if (!make_current(context, mixer_surface)) {
@@ -521,24 +643,44 @@ void Mixer::thread_func()
        // Start the actual capture. (We don't want to do it before we're actually ready
        // to process output frames.)
        for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
-               cards[card_index].capture->start_bm_capture();
+               if (int(card_index) != output_card_index) {
+                       cards[card_index].capture->start_bm_capture();
+               }
        }
 
        steady_clock::time_point start, now;
        start = steady_clock::now();
 
-       int frame = 0;
        int stats_dropped_frames = 0;
 
        while (!should_quit) {
+               if (desired_output_card_index != output_card_index) {
+                       set_output_card_internal(desired_output_card_index);
+               }
+               if (output_card_index != -1 &&
+                   desired_output_video_mode != output_video_mode) {
+                       DeckLinkOutput *output = cards[output_card_index].output.get();
+                       output->end_output();
+                       desired_output_video_mode = output_video_mode = output->pick_video_mode(desired_output_video_mode);
+                       output->start_output(desired_output_video_mode, pts_int);
+               }
+
                CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS];
                bool has_new_frame[MAX_VIDEO_CARDS] = { false };
 
-               unsigned master_card_index = theme->map_signal(master_clock_channel);
-               assert(master_card_index < num_cards);
+               bool master_card_is_output;
+               unsigned master_card_index;
+               if (output_card_index != -1) {
+                       master_card_is_output = true;
+                       master_card_index = output_card_index;
+               } else {
+                       master_card_is_output = false;
+                       master_card_index = theme->map_signal(master_clock_channel);
+                       assert(master_card_index < num_cards);
+               }
 
-               OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame);
-               schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration);
+               OutputFrameInfo output_frame_info = get_one_frame_from_each_card(master_card_index, master_card_is_output, new_frames, has_new_frame);
+               schedule_audio_resampling_tasks(output_frame_info.dropped_frames, output_frame_info.num_samples, output_frame_info.frame_duration, output_frame_info.is_preroll, output_frame_info.frame_timestamp);
                stats_dropped_frames += output_frame_info.dropped_frames;
 
                handle_hotplugged_cards();
@@ -558,7 +700,7 @@ 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 (new_frames[master_card_index].frame->len == 0) {
+               if (!master_card_is_output && new_frames[master_card_index].frame->len == 0) {
                        ++stats_dropped_frames;
                        pts_int += new_frames[master_card_index].length;
                        continue;
@@ -582,15 +724,15 @@ void Mixer::thread_func()
 
                int64_t frame_duration = output_frame_info.frame_duration;
                render_one_frame(frame_duration);
-               ++frame;
+               ++frame_num;
                pts_int += frame_duration;
 
                now = steady_clock::now();
                double elapsed = duration<double>(now - start).count();
-               if (frame % 100 == 0) {
+               if (frame_num % 100 == 0) {
                        printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)",
-                               frame, stats_dropped_frames, elapsed, frame / elapsed,
-                               1e3 * elapsed / frame);
+                               frame_num, stats_dropped_frames, elapsed, frame_num / elapsed,
+                               1e3 * elapsed / frame_num);
                //      chain->print_phase_timing();
 
                        // Check our memory usage, to see if we are close to our mlockall()
@@ -627,7 +769,7 @@ void Mixer::thread_func()
 
 
                if (should_cut.exchange(false)) {  // Test and clear.
-                       video_encoder->do_cut(frame);
+                       video_encoder->do_cut(frame_num);
                }
 
 #if 0
@@ -646,17 +788,77 @@ void Mixer::thread_func()
        resource_pool->clean_context();
 }
 
-Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS])
+bool Mixer::input_card_is_master_clock(unsigned card_index, unsigned master_card_index) const
 {
-       OutputFrameInfo output_frame_info;
+       if (output_card_index != -1) {
+               // The output card (ie., cards[output_card_index].output) is the master clock,
+               // so no input card (ie., cards[card_index].capture) is.
+               return false;
+       }
+       return (card_index == master_card_index);
+}
+
+void Mixer::trim_queue(CaptureCard *card, unsigned card_index)
+{
+       // Count the number of frames in the queue, including any frames
+       // we dropped. It's hard to know exactly how we should deal with
+       // dropped (corrupted) input frames; they don't help our goal of
+       // avoiding starvation, but they still add to the problem of latency.
+       // Since dropped frames is going to mean a bump in the signal anyway,
+       // we err on the side of having more stable latency instead.
+       unsigned queue_length = 0;
+       for (const CaptureCard::NewFrame &frame : card->new_frames) {
+               queue_length += frame.dropped_frames + 1;
+       }
+       card->queue_length_policy.update_policy(queue_length);
+
+       // If needed, drop frames until the queue is below the safe limit.
+       // We prefer to drop from the head, because all else being equal,
+       // we'd like more recent frames (less latency).
+       unsigned dropped_frames = 0;
+       while (queue_length > card->queue_length_policy.get_safe_queue_length()) {
+               assert(!card->new_frames.empty());
+               assert(queue_length > card->new_frames.front().dropped_frames);
+               queue_length -= card->new_frames.front().dropped_frames;
+
+               if (queue_length <= card->queue_length_policy.get_safe_queue_length()) {
+                       // No need to drop anything.
+                       break;
+               }
+
+               card->new_frames.pop_front();
+               card->new_frames_changed.notify_all();
+               --queue_length;
+               ++dropped_frames;
+       }
+
+       if (dropped_frames > 0) {
+               fprintf(stderr, "Card %u dropped %u frame(s) to keep latency down.\n",
+                       card_index, dropped_frames);
+       }
+}
 
+
+Mixer::OutputFrameInfo Mixer::get_one_frame_from_each_card(unsigned master_card_index, bool master_card_is_output, CaptureCard::NewFrame new_frames[MAX_VIDEO_CARDS], bool has_new_frame[MAX_VIDEO_CARDS])
+{
+       OutputFrameInfo output_frame_info;
 start:
-       // The first card is the master timer, so wait for it to have a new frame.
-       // TODO: Add a timeout.
-       unique_lock<mutex> lock(bmusb_mutex);
-       cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
+       unique_lock<mutex> lock(card_mutex, defer_lock);
+       if (master_card_is_output) {
+               // Clocked to the output, so wait for it to be ready for the next frame.
+               cards[master_card_index].output->wait_for_frame(pts_int, &output_frame_info.dropped_frames, &output_frame_info.frame_duration, &output_frame_info.is_preroll, &output_frame_info.frame_timestamp);
+               lock.lock();
+       } else {
+               // Wait for the master card to have a new frame.
+               // TODO: Add a timeout.
+               output_frame_info.is_preroll = false;
+               lock.lock();
+               cards[master_card_index].new_frames_changed.wait(lock, [this, master_card_index]{ return !cards[master_card_index].new_frames.empty() || cards[master_card_index].capture->get_disconnected(); });
+       }
 
-       if (cards[master_card_index].new_frames.empty()) {
+       if (master_card_is_output) {
+               handle_hotplugged_cards();
+       } else if (cards[master_card_index].new_frames.empty()) {
                // We were woken up, but not due to a new frame. Deal with it
                // and then restart.
                assert(cards[master_card_index].capture->get_disconnected());
@@ -664,35 +866,34 @@ start:
                goto start;
        }
 
+       if (!master_card_is_output) {
+               output_frame_info.frame_timestamp =
+                       cards[master_card_index].new_frames.front().received_timestamp;
+       }
+
        for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                CaptureCard *card = &cards[card_index];
-               if (card->new_frames.empty()) {
-                       assert(card_index != master_card_index);
-                       card->queue_length_policy.update_policy(-1);
-                       continue;
-               }
-               new_frames[card_index] = move(card->new_frames.front());
-               has_new_frame[card_index] = true;
-               card->new_frames.pop();
-               card->new_frames_changed.notify_all();
-
-               if (card_index == master_card_index) {
+               if (input_card_is_master_clock(card_index, master_card_index)) {
                        // We don't use the queue length policy for the master card,
                        // but we will if it stops being the master. Thus, clear out
                        // the policy in case we switch in the future.
                        card->queue_length_policy.reset(card_index);
+                       assert(!card->new_frames.empty());
                } else {
-                       // If we have excess frames compared to the policy for this card,
-                       // drop frames from the head.
-                       card->queue_length_policy.update_policy(card->new_frames.size());
-                       while (card->new_frames.size() > card->queue_length_policy.get_safe_queue_length()) {
-                               card->new_frames.pop();
-                       }
+                       trim_queue(card, card_index);
+               }
+               if (!card->new_frames.empty()) {
+                       new_frames[card_index] = move(card->new_frames.front());
+                       has_new_frame[card_index] = true;
+                       card->new_frames.pop_front();
+                       card->new_frames_changed.notify_all();
                }
        }
 
-       output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
-       output_frame_info.frame_duration = new_frames[master_card_index].length;
+       if (!master_card_is_output) {
+               output_frame_info.dropped_frames = new_frames[master_card_index].dropped_frames;
+               output_frame_info.frame_duration = new_frames[master_card_index].length;
+       }
 
        // This might get off by a fractional sample when changing master card
        // between ones with different frame rates, but that's fine.
@@ -712,7 +913,7 @@ void Mixer::handle_hotplugged_cards()
                if (card->capture->get_disconnected()) {
                        fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
                        FakeCapture *capture = new FakeCapture(global_flags.width, global_flags.height, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
-                       configure_card(card_index, capture, /*is_fake_capture=*/true);
+                       configure_card(card_index, capture, /*is_fake_capture=*/true, /*output=*/nullptr);
                        card->queue_length_policy.reset(card_index);
                        card->capture->start_bm_capture();
                }
@@ -729,7 +930,7 @@ void Mixer::handle_hotplugged_cards()
                int free_card_index = -1;
                for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                        if (cards[card_index].is_fake_capture) {
-                               free_card_index = int(card_index);
+                               free_card_index = card_index;
                                break;
                        }
                }
@@ -742,7 +943,7 @@ void Mixer::handle_hotplugged_cards()
                        fprintf(stderr, "New card plugged in, choosing slot %d.\n", free_card_index);
                        CaptureCard *card = &cards[free_card_index];
                        BMUSBCapture *capture = new BMUSBCapture(free_card_index, new_dev);
-                       configure_card(free_card_index, capture, /*is_fake_capture=*/false);
+                       configure_card(free_card_index, capture, /*is_fake_capture=*/false, /*output=*/nullptr);
                        card->queue_length_policy.reset(free_card_index);
                        capture->set_card_disconnected_callback(bind(&Mixer::bm_hotplug_remove, this, free_card_index));
                        capture->start_bm_capture();
@@ -751,7 +952,7 @@ void Mixer::handle_hotplugged_cards()
 }
 
 
-void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
+void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame, bool is_preroll, steady_clock::time_point frame_timestamp)
 {
        // Resample the audio as needed, including from previously dropped frames.
        assert(num_cards > 0);
@@ -772,8 +973,8 @@ void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_sam
                        // since dropped frames are expected to be rare, and it might be
                        // better to just wait until we have a slightly more normal situation).
                        unique_lock<mutex> lock(audio_mutex);
-                       bool adjust_rate = !dropped_frame;
-                       audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate});
+                       bool adjust_rate = !dropped_frame && !is_preroll;
+                       audio_task_queue.push(AudioTask{pts_int, num_samples_per_frame, adjust_rate, frame_timestamp});
                        audio_task_queue_changed.notify_one();
                }
                if (dropped_frame) {
@@ -787,14 +988,21 @@ void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_sam
 
 void Mixer::render_one_frame(int64_t duration)
 {
+       // Determine the time code for this frame before we start rendering.
+       string timecode_text = timecode_renderer->get_timecode_text(double(pts_int) / TIMEBASE, frame_num);
+       if (display_timecode_on_stdout) {
+               printf("Timecode: '%s'\n", timecode_text.c_str());
+       }
+
        // Get the main chain from the theme, and set its state immediately.
        Theme::Chain theme_main_chain = theme->get_chain(0, pts(), global_flags.width, global_flags.height, input_state);
        EffectChain *chain = theme_main_chain.chain;
        theme_main_chain.setup_chain();
        //theme_main_chain.chain->enable_phase_timing(true);
 
+       const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
        GLuint y_tex, cbcr_tex;
-       bool got_frame = video_encoder->begin_frame(&y_tex, &cbcr_tex);
+       bool got_frame = video_encoder->begin_frame(pts_int + av_delay, duration, theme_main_chain.input_frames, &y_tex, &cbcr_tex);
        assert(got_frame);
 
        // Render main chain.
@@ -803,9 +1011,18 @@ void Mixer::render_one_frame(int64_t duration)
        GLuint fbo = resource_pool->create_fbo(y_tex, cbcr_full_tex, rgba_tex);
        check_error();
        chain->render_to_fbo(fbo, global_flags.width, global_flags.height);
+
+       if (display_timecode_in_stream) {
+               // Render the timecode on top.
+               timecode_renderer->render_timecode(fbo, timecode_text);
+       }
+
        resource_pool->release_fbo(fbo);
 
        chroma_subsampler->subsample_chroma(cbcr_full_tex, global_flags.width, global_flags.height, cbcr_tex);
+       if (output_card_index != -1) {
+               cards[output_card_index].output->send_frame(y_tex, cbcr_full_tex, theme_main_chain.input_frames, pts_int, duration);
+       }
        resource_pool->release_2d_texture(cbcr_full_tex);
 
        // Set the right state for rgba_tex.
@@ -815,8 +1032,7 @@ void Mixer::render_one_frame(int64_t duration)
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 
-       const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
-       RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);
+       RefCountedGLsync fence = video_encoder->end_frame();
 
        // The live frame just shows the RGBA texture we just rendered.
        // It owns rgba_tex now.
@@ -845,6 +1061,8 @@ void Mixer::render_one_frame(int64_t duration)
 
 void Mixer::audio_thread_func()
 {
+       pthread_setname_np(pthread_self(), "Mixer_Audio");
+
        while (!should_quit) {
                AudioTask task;
 
@@ -861,7 +1079,7 @@ void Mixer::audio_thread_func()
                ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
                        task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
                vector<float> samples_out = audio_mixer.get_output(
-                       double(task.pts_int) / TIMEBASE,
+                       task.frame_timestamp,
                        task.num_samples,
                        rate_adjustment_policy);
 
@@ -869,7 +1087,10 @@ void Mixer::audio_thread_func()
                if (alsa) {
                        alsa->write(samples_out);
                }
-               decklink_output->send_audio(task.pts_int, samples_out);
+               if (output_card_index != -1) {
+                       const int64_t av_delay = lrint(global_flags.audio_queue_length_ms * 0.001 * TIMEBASE);  // Corresponds to the delay in ResamplingQueue.
+                       cards[output_card_index].output->send_audio(task.pts_int + av_delay, samples_out);
+               }
                video_encoder->add_audio(task.pts_int, move(samples_out));
        }
 }
@@ -925,6 +1146,13 @@ void Mixer::start_mode_scanning(unsigned card_index)
        last_mode_scan_change[card_index] = steady_clock::now();
 }
 
+map<uint32_t, bmusb::VideoMode> Mixer::get_available_output_video_modes() const
+{
+       assert(desired_output_card_index != -1);
+       unique_lock<mutex> lock(card_mutex);
+       return cards[desired_output_card_index].output->get_available_video_modes();
+}
+
 Mixer::OutputChannel::~OutputChannel()
 {
        if (has_current_frame) {