[nageru] / mixer.cpp

#undef Success

#include "mixer.h"

#include <assert.h>
#include <endian.h>
#include <epoxy/egl.h>
#include <movit/effect_chain.h>
#include <movit/effect_util.h>
#include <movit/flat_input.h>
#include <movit/image_format.h>
#include <movit/init.h>
#include <movit/resource_pool.h>
#include <movit/util.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/resource.h>
#include <sys/time.h>
#include <time.h>
#include <algorithm>
#include <chrono>
#include <cmath>
#include <condition_variable>
#include <cstddef>
#include <memory>
#include <mutex>
#include <string>
#include <thread>
#include <utility>
#include <vector>

#include "bmusb/bmusb.h"
#include "bmusb/fake_capture.h"
#include "context.h"
#include "db.h"
#include "decklink_capture.h"
#include "defs.h"
#include "disk_space_estimator.h"
#include "flags.h"
#include "pbo_frame_allocator.h"
#include "ref_counted_gl_sync.h"
#include "timebase.h"
#include "video_encoder.h"

class QOpenGLContext;

using namespace movit;
using namespace std;
using namespace std::chrono;
using namespace std::placeholders;
using namespace bmusb;

Mixer *global_mixer = nullptr;
bool uses_mlock = false;

namespace {

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

void QueueLengthPolicy::update_policy(int queue_length)
{
        if (queue_length < 0) {  // Starvation.
                if (been_at_safe_point_since_last_starvation && safe_queue_length < 5) {
                        ++safe_queue_length;
                        fprintf(stderr, "Card %u: Starvation, increasing safe limit to %u frames\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)) {
                        been_at_safe_point_since_last_starvation = true;
                }
                if (++frames_with_at_least_one >= 1000 && safe_queue_length > 0) {
                        --safe_queue_length;
                        fprintf(stderr, "Card %u: Spare frames for more than 1000 frames, reducing safe limit to %u frames\n",
                                card_index, safe_queue_length);
                        frames_with_at_least_one = 0;
                }
        } else {
                frames_with_at_least_one = 0;
        }
}

Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
        : httpd(),
          num_cards(num_cards),
          mixer_surface(create_surface(format)),
          h264_encoder_surface(create_surface(format)),
          audio_mixer(num_cards),
          correlation(OUTPUT_FREQUENCY)
{
        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(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
        for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
                output_channel[i].parent = this;
                output_channel[i].channel = i;
        }

        ImageFormat inout_format;
        inout_format.color_space = COLORSPACE_sRGB;
        inout_format.gamma_curve = GAMMA_sRGB;

        // 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();
        display_input = new FlatInput(inout_format, FORMAT_RGB, GL_UNSIGNED_BYTE, WIDTH, HEIGHT);  // FIXME: GL_UNSIGNED_BYTE is really wrong.
        display_chain->add_input(display_input);
        display_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
        display_chain->set_dither_bits(0);  // Don't bother.
        display_chain->finalize();

        video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, WIDTH, HEIGHT, &httpd, global_disk_space_estimator));

        // Start listening for clients only once VideoEncoder has written its header, if any.
        httpd.start(9095);

        // First try initializing the then PCI devices, then USB, then
        // fill up with fake cards until we have the desired number of cards.
        unsigned num_pci_devices = 0;
        unsigned card_index = 0;

        {
                IDeckLinkIterator *decklink_iterator = CreateDeckLinkIteratorInstance();
                if (decklink_iterator != nullptr) {
                        for ( ; card_index < num_cards; ++card_index) {
                                IDeckLink *decklink;
                                if (decklink_iterator->Next(&decklink) != S_OK) {
                                        break;
                                }

                                configure_card(card_index, new DeckLinkCapture(decklink, card_index), /*is_fake_capture=*/false);
                                ++num_pci_devices;
                        }
                        decklink_iterator->Release();
                        fprintf(stderr, "Found %u DeckLink PCI card(s).\n", num_pci_devices);
                } else {
                        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);
        }
        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(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
                configure_card(card_index, capture, /*is_fake_capture=*/true);
        }

        if (num_fake_cards > 0) {
                fprintf(stderr, "Initialized %u fake cards.\n", num_fake_cards);
        }

        BMUSBCapture::set_card_connected_callback(bind(&Mixer::bm_hotplug_add, this, _1));
        BMUSBCapture::start_bm_thread();

        for (card_index = 0; card_index < num_cards; ++card_index) {
                cards[card_index].queue_length_policy.reset(card_index);
                cards[card_index].capture->start_bm_capture();
        }

        // Set up stuff for NV12 conversion.

        // Cb/Cr shader.
        string cbcr_vert_shader =
                "#version 130 \n"
                " \n"
                "in vec2 position; \n"
                "in vec2 texcoord; \n"
                "out vec2 tc0; \n"
                "uniform vec2 foo_chroma_offset_0; \n"
                " \n"
                "void main() \n"
                "{ \n"
                "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
                "    // \n"
                "    //   2.000  0.000  0.000 -1.000 \n"
                "    //   0.000  2.000  0.000 -1.000 \n"
                "    //   0.000  0.000 -2.000 -1.000 \n"
                "    //   0.000  0.000  0.000  1.000 \n"
                "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
                "    vec2 flipped_tc = texcoord; \n"
                "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
                "} \n";
        string cbcr_frag_shader =
                "#version 130 \n"
                "in vec2 tc0; \n"
                "uniform sampler2D cbcr_tex; \n"
                "out vec4 FragColor; \n"
                "void main() { \n"
                "    FragColor = texture(cbcr_tex, tc0); \n"
                "} \n";
        vector<string> frag_shader_outputs;
        cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);

        float vertices[] = {
                0.0f, 2.0f,
                0.0f, 0.0f,
                2.0f, 0.0f
        };
        cbcr_vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
        cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
        cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");

        r128.init(2, OUTPUT_FREQUENCY);
        r128.integr_start();

        // 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, /*frel=*/1.0);

        if (global_flags.enable_alsa_output) {
                alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
        }
}

Mixer::~Mixer()
{
        resource_pool->release_glsl_program(cbcr_program_num);
        glDeleteBuffers(1, &cbcr_vbo);
        BMUSBCapture::stop_bm_thread();

        for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                {
                        unique_lock<mutex> lock(bmusb_mutex);
                        cards[card_index].should_quit = true;  // Unblock thread.
                        cards[card_index].new_frames_changed.notify_all();
                }
                cards[card_index].capture->stop_dequeue_thread();
        }

        video_encoder.reset(nullptr);
}

void Mixer::configure_card(unsigned card_index, CaptureInterface *capture, bool is_fake_capture)
{
        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->is_fake_capture = is_fake_capture;
        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, WIDTH, HEIGHT));  // 8 MB.
        }
        card->capture->set_video_frame_allocator(card->frame_allocator.get());
        if (card->surface == nullptr) {
                card->surface = create_surface_with_same_format(mixer_surface);
        }
        audio_mixer.reset_card(card_index);
        while (!card->new_frames.empty()) card->new_frames.pop();
        card->fractional_samples = 0;
        card->last_timecode = -1;
        card->capture->configure_card();
}


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 deinterleave_samples(const vector<float> &in, vector<float> *out_l, vector<float> *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, VideoFormat video_format,
                     FrameAllocator::Frame audio_frame, size_t audio_offset, AudioFormat audio_format)
{
        CaptureCard *card = &cards[card_index];

        if (is_mode_scanning[card_index]) {
                if (video_format.has_signal) {
                        // Found a stable signal, so stop scanning.
                        is_mode_scanning[card_index] = false;
                } else {
                        static constexpr double switch_time_s = 0.5;  // Should be enough time for the signal to stabilize.
                        steady_clock::time_point now = steady_clock::now();
                        double sec_since_last_switch = duration<double>(steady_clock::now() - last_mode_scan_change[card_index]).count();
                        if (sec_since_last_switch > switch_time_s) {
                                // It isn't this mode; try the next one.
                                mode_scanlist_index[card_index]++;
                                mode_scanlist_index[card_index] %= mode_scanlist[card_index].size();
                                cards[card_index].capture->set_video_mode(mode_scanlist[card_index][mode_scanlist_index[card_index]]);
                                last_mode_scan_change[card_index] = now;
                        }
                }
        }

        int64_t frame_length = int64_t(TIMEBASE) * video_format.frame_rate_den / video_format.frame_rate_nom;
        assert(frame_length > 0);

        size_t num_samples = (audio_frame.len > audio_offset) ? (audio_frame.len - audio_offset) / audio_format.num_channels / (audio_format.bits_per_sample / 8) : 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.id);
                if (video_frame.owner) {
                        video_frame.owner->release_frame(video_frame);
                }
                if (audio_frame.owner) {
                        audio_frame.owner->release_frame(audio_frame);
                }
                return;
        }

        int dropped_frames = 0;
        if (card->last_timecode != -1) {
                dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
        }

        // Number of samples per frame if we need to insert silence.
        // (Could be nonintegral, but resampling will save us then.)
        const int silence_samples = OUTPUT_FREQUENCY * video_format.frame_rate_den / video_format.frame_rate_nom;

        if (dropped_frames > MAX_FPS * 2) {
                fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
                        card_index, card->last_timecode, timecode);
                audio_mixer.reset_card(card_index);
                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);

                audio_mixer.add_silence(card_index, silence_samples, dropped_frames, frame_length);
        }

        audio_mixer.add_audio(card_index, audio_frame.data + audio_offset, num_samples, audio_format, frame_length);

        // Done with the audio, so release it.
        if (audio_frame.owner) {
                audio_frame.owner->release_frame(audio_frame);
        }

        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;
        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<mutex> lock(bmusb_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));
                        card->new_frames_changed.notify_all();
                }
                return;
        }

        PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;

        unsigned num_fields = video_format.interlaced ? 2 : 1;
        steady_clock::time_point frame_upload_start;
        if (video_format.interlaced) {
                // Send the two fields along as separate frames; the other side will need to add
                // a deinterlacer to actually get this right.
                assert(video_format.height % 2 == 0);
                video_format.height /= 2;
                assert(frame_length % 2 == 0);
                frame_length /= 2;
                num_fields = 2;
                frame_upload_start = steady_clock::now();
        }
        userdata->last_interlaced = video_format.interlaced;
        userdata->last_has_signal = video_format.has_signal;
        userdata->last_is_connected = video_format.is_connected;
        userdata->last_frame_rate_nom = video_format.frame_rate_nom;
        userdata->last_frame_rate_den = video_format.frame_rate_den;
        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;

        for (unsigned field = 0; field < num_fields; ++field) {
                // Put the actual texture upload in a lambda that is executed in the main thread.
                // It is entirely possible to do this in the same thread (and it might even be
                // faster, depending on the GPU and driver), but it appears to be trickling
                // driver bugs very easily.
                //
                // 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, userdata]() {
                        unsigned field_start_line = (field == 1) ? video_format.second_field_start : video_format.extra_lines_top + field * (video_format.height + 22);

                        if (userdata->tex_y[field] == 0 ||
                            userdata->tex_cbcr[field] == 0 ||
                            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;
                        }

                        GLuint pbo = userdata->pbo;
                        check_error();
                        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 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.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();
                        }

                        glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
                        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();
                        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();
                };

                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.
                        steady_clock::time_point second_field_start = frame_upload_start +
                                nanoseconds(frame_length * 1000000000 / TIMEBASE);
                        this_thread::sleep_until(second_field_start);
                }

                {
                        unique_lock<mutex> lock(bmusb_mutex);
                        CaptureCard::NewFrame new_frame;
                        new_frame.frame = frame;
                        new_frame.length = frame_length;
                        new_frame.field = field;
                        new_frame.interlaced = video_format.interlaced;
                        new_frame.upload_func = upload_func;
                        new_frame.dropped_frames = dropped_frames;
                        card->new_frames.push(move(new_frame));
                        card->new_frames_changed.notify_all();
                }
        }
}

void Mixer::bm_hotplug_add(libusb_device *dev)
{
        lock_guard<mutex> lock(hotplug_mutex);
        hotplugged_cards.push_back(dev);
}

void Mixer::bm_hotplug_remove(unsigned card_index)
{
        cards[card_index].new_frames_changed.notify_all();
}

void Mixer::thread_func()
{
        eglBindAPI(EGL_OPENGL_API);
        QOpenGLContext *context = create_context(mixer_surface);
        if (!make_current(context, mixer_surface)) {
                printf("oops\n");
                exit(1);
        }

        steady_clock::time_point start, now;
        start = steady_clock::now();

        int frame = 0;
        int stats_dropped_frames = 0;

        while (!should_quit) {
                CaptureCard::NewFrame new_frames[MAX_CARDS];
                bool has_new_frame[MAX_CARDS] = { false };
                int num_samples[MAX_CARDS] = { 0 };

                unsigned master_card_index = theme->map_signal(master_clock_channel);
                assert(master_card_index < num_cards);

                get_one_frame_from_each_card(master_card_index, new_frames, has_new_frame, num_samples);
                schedule_audio_resampling_tasks(new_frames[master_card_index].dropped_frames, num_samples[master_card_index], new_frames[master_card_index].length);
                stats_dropped_frames += new_frames[master_card_index].dropped_frames;
                send_audio_level_callback();

                handle_hotplugged_cards();

                for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                        if (card_index == master_card_index || !has_new_frame[card_index]) {
                                continue;
                        }
                        if (new_frames[card_index].frame->len == 0) {
                                ++new_frames[card_index].dropped_frames;
                        }
                        if (new_frames[card_index].dropped_frames > 0) {
                                printf("Card %u dropped %d frames before this\n",
                                        card_index, int(new_frames[card_index].dropped_frames));
                        }
                }

                // 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) {
                        ++stats_dropped_frames;
                        pts_int += new_frames[master_card_index].length;
                        continue;
                }

                for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                        if (!has_new_frame[card_index] || new_frames[card_index].frame->len == 0)
                                continue;

                        CaptureCard::NewFrame *new_frame = &new_frames[card_index];
                        assert(new_frame->frame != nullptr);
                        insert_new_frame(new_frame->frame, new_frame->field, new_frame->interlaced, card_index, &input_state);
                        check_error();

                        // The new texture might need uploading before use.
                        if (new_frame->upload_func) {
                                new_frame->upload_func();
                                new_frame->upload_func = nullptr;
                        }
                }

                int64_t frame_duration = new_frames[master_card_index].length;
                render_one_frame(frame_duration);
                ++frame;
                pts_int += frame_duration;

                now = steady_clock::now();
                double elapsed = duration<double>(now - start).count();
                if (frame % 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);
                //      chain->print_phase_timing();

                        // Check our memory usage, to see if we are close to our mlockall()
                        // limit (if at all set).
                        rusage used;
                        if (getrusage(RUSAGE_SELF, &used) == -1) {
                                perror("getrusage(RUSAGE_SELF)");
                                assert(false);
                        }

                        if (uses_mlock) {
                                rlimit limit;
                                if (getrlimit(RLIMIT_MEMLOCK, &limit) == -1) {
                                        perror("getrlimit(RLIMIT_MEMLOCK)");
                                        assert(false);
                                }

                                printf(", using %ld / %ld MB lockable memory (%.1f%%)",
                                        long(used.ru_maxrss / 1024),
                                        long(limit.rlim_cur / 1048576),
                                        float(100.0 * (used.ru_maxrss * 1024.0) / limit.rlim_cur));
                        } else {
                                printf(", using %ld MB memory (not locked)",
                                        long(used.ru_maxrss / 1024));
                        }

                        printf("\n");
                }


                if (should_cut.exchange(false)) {  // Test and clear.
                        video_encoder->do_cut(frame);
                }

#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
                // remaining duration of the program.
                if (frame == 10000) {
                        frame = 0;
                        start = now;
                }
#endif
                check_error();
        }

        resource_pool->clean_context();
}

void Mixer::get_one_frame_from_each_card(unsigned master_card_index, CaptureCard::NewFrame new_frames[MAX_CARDS], bool has_new_frame[MAX_CARDS], int num_samples[MAX_CARDS])
{
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(); });

        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());
                handle_hotplugged_cards();
                goto start;
        }

        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();

                int num_samples_times_timebase = OUTPUT_FREQUENCY * new_frames[card_index].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);

                if (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);
                } 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();
                        }
                }
        }
}

void Mixer::handle_hotplugged_cards()
{
        // Check for cards that have been disconnected since last frame.
        for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
                CaptureCard *card = &cards[card_index];
                if (card->capture->get_disconnected()) {
                        fprintf(stderr, "Card %u went away, replacing with a fake card.\n", card_index);
                        FakeCapture *capture = new FakeCapture(WIDTH, HEIGHT, FAKE_FPS, OUTPUT_FREQUENCY, card_index, global_flags.fake_cards_audio);
                        configure_card(card_index, capture, /*is_fake_capture=*/true);
                        card->queue_length_policy.reset(card_index);
                        card->capture->start_bm_capture();
                }
        }

        // Check for cards that have been connected since last frame.
        vector<libusb_device *> hotplugged_cards_copy;
        {
                lock_guard<mutex> lock(hotplug_mutex);
                swap(hotplugged_cards, hotplugged_cards_copy);
        }
        for (libusb_device *new_dev : hotplugged_cards_copy) {
                // Look for a fake capture card where we can stick this in.
                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);
                                break;
                        }
                }

                if (free_card_index == -1) {
                        fprintf(stderr, "New card plugged in, but no free slots -- ignoring.\n");
                        libusb_unref_device(new_dev);
                } else {
                        // BMUSBCapture takes ownership.
                        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);
                        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();
                }
        }
}


void Mixer::schedule_audio_resampling_tasks(unsigned dropped_frames, int num_samples_per_frame, int length_per_frame)
{
        // Resample the audio as needed, including from previously dropped frames.
        assert(num_cards > 0);
        for (unsigned frame_num = 0; frame_num < dropped_frames + 1; ++frame_num) {
                const bool dropped_frame = (frame_num != dropped_frames);
                {
                        // Signal to the audio thread to process this frame.
                        // Note that if the frame is a dropped frame, we signal that
                        // we don't want to use this frame as base for adjusting
                        // the resampler rate. The reason for this is that the timing
                        // of these frames is often way too late; they typically don't
                        // “arrive” before we synthesize them. Thus, we could end up
                        // in a situation where we have inserted e.g. five audio frames
                        // into the queue before we then start pulling five of them
                        // back out. This makes ResamplingQueue overestimate the delay,
                        // causing undue resampler changes. (We _do_ use the last,
                        // non-dropped frame; perhaps we should just discard that as well,
                        // 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});
                        audio_task_queue_changed.notify_one();
                }
                if (dropped_frame) {
                        // 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.
                        pts_int += length_per_frame;
                }
        }
}

void Mixer::render_one_frame(int64_t duration)
{
        // Get the main chain from the theme, and set its state immediately.
        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 = video_encoder->begin_frame(&y_tex, &cbcr_tex);
        assert(got_frame);

        // Render main chain.
        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);

        subsample_chroma(cbcr_full_tex, cbcr_tex);
        resource_pool->release_2d_texture(cbcr_full_tex);

        // Set the right state for rgba_tex.
        glBindFramebuffer(GL_FRAMEBUFFER, 0);
        glBindTexture(GL_TEXTURE_2D, rgba_tex);
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
        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 = TIMEBASE / 10;  // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
        RefCountedGLsync fence = video_encoder->end_frame(pts_int + av_delay, duration, theme_main_chain.input_frames);

        // The live frame just shows the RGBA texture we just rendered.
        // It owns rgba_tex now.
        DisplayFrame live_frame;
        live_frame.chain = display_chain.get();
        live_frame.setup_chain = [this, rgba_tex]{
                display_input->set_texture_num(rgba_tex);
        };
        live_frame.ready_fence = fence;
        live_frame.input_frames = {};
        live_frame.temp_textures = { rgba_tex };
        output_channel[OUTPUT_LIVE].output_frame(live_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);
        }
}

void Mixer::send_audio_level_callback()
{
        if (audio_level_callback == nullptr) {
                return;
        }

        unique_lock<mutex> lock(audio_measure_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, to_db(peak),
                loudness_i, loudness_range_low, loudness_range_high,
                audio_mixer.get_gain_staging_db(),
                audio_mixer.get_final_makeup_gain_db(),
                correlation.get_correlation());
}

void Mixer::audio_thread_func()
{
        while (!should_quit) {
                AudioTask task;

                {
                        unique_lock<mutex> lock(audio_mutex);
                        audio_task_queue_changed.wait(lock, [this]{ return should_quit || !audio_task_queue.empty(); });
                        if (should_quit) {
                                return;
                        }
                        task = audio_task_queue.front();
                        audio_task_queue.pop();
                }

                ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy =
                        task.adjust_rate ? ResamplingQueue::ADJUST_RATE : ResamplingQueue::DO_NOT_ADJUST_RATE;
                process_audio_one_frame(task.pts_int, task.num_samples, rate_adjustment_policy);
        }
}

void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
{
        vector<float> samples_out = audio_mixer.get_output(double(frame_pts_int) / TIMEBASE, num_samples, rate_adjustment_policy);

        // Upsample 4x to find interpolated peak.
        peak_resampler.inp_data = samples_out.data();
        peak_resampler.inp_count = samples_out.size() / 2;

        vector<float> interpolated_samples_out;
        interpolated_samples_out.resize(samples_out.size());
        {
                unique_lock<mutex> lock(audio_measure_mutex);

                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<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
                        peak_resampler.out_data = nullptr;
                }
        }

        // Find R128 levels and L/R correlation.
        vector<float> left, right;
        deinterleave_samples(samples_out, &left, &right);
        float *ptrs[] = { left.data(), right.data() };
        {
                unique_lock<mutex> lock(audio_measure_mutex);
                r128.process(left.size(), ptrs);
                audio_mixer.set_current_loudness(r128.loudness_M());
                correlation.process_samples(samples_out);
        }

        // Send the samples to the sound card.
        if (alsa) {
                alsa->write(samples_out);
        }

        // And finally add them to the output.
        video_encoder->add_audio(frame_pts_int, move(samples_out));
}

void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
{
        GLuint vao;
        glGenVertexArrays(1, &vao);
        check_error();

        glBindVertexArray(vao);
        check_error();

        // Extract Cb/Cr.
        GLuint fbo = resource_pool->create_fbo(dst_tex);
        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
        glViewport(0, 0, WIDTH/2, HEIGHT/2);
        check_error();

        glUseProgram(cbcr_program_num);
        check_error();

        glActiveTexture(GL_TEXTURE0);
        check_error();
        glBindTexture(GL_TEXTURE_2D, src_tex);
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
        check_error();
        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
        check_error();

        float chroma_offset_0[] = { -0.5f / WIDTH, 0.0f };
        set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);

        glBindBuffer(GL_ARRAY_BUFFER, cbcr_vbo);
        check_error();

        for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
                glEnableVertexAttribArray(attr_index);
                check_error();
                glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
                check_error();
        }

        glDrawArrays(GL_TRIANGLES, 0, 3);
        check_error();

        for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
                glDisableVertexAttribArray(attr_index);
                check_error();
        }

        glUseProgram(0);
        check_error();
        glBindFramebuffer(GL_FRAMEBUFFER, 0);
        check_error();

        resource_pool->release_fbo(fbo);
        glDeleteVertexArrays(1, &vao);
}

void Mixer::release_display_frame(DisplayFrame *frame)
{
        for (GLuint texnum : frame->temp_textures) {
                resource_pool->release_2d_texture(texnum);
        }
        frame->temp_textures.clear();
        frame->ready_fence.reset();
        frame->input_frames.clear();
}

void Mixer::start()
{
        mixer_thread = thread(&Mixer::thread_func, this);
        audio_thread = thread(&Mixer::audio_thread_func, this);
}

void Mixer::quit()
{
        should_quit = true;
        audio_task_queue_changed.notify_one();
        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()
{
        unique_lock<mutex> lock(audio_measure_mutex);
        peak_resampler.reset();
        peak = 0.0f;
        r128.reset();
        r128.integr_start();
        correlation.reset();
}

void Mixer::start_mode_scanning(unsigned card_index)
{
        assert(card_index < num_cards);
        if (is_mode_scanning[card_index]) {
                return;
        }
        is_mode_scanning[card_index] = true;
        mode_scanlist[card_index].clear();
        for (const auto &mode : cards[card_index].capture->get_available_video_modes()) {
                mode_scanlist[card_index].push_back(mode.first);
        }
        assert(!mode_scanlist[card_index].empty());
        mode_scanlist_index[card_index] = 0;
        cards[card_index].capture->set_video_mode(mode_scanlist[card_index][0]);
        last_mode_scan_change[card_index] = steady_clock::now();
}

Mixer::OutputChannel::~OutputChannel()
{
        if (has_current_frame) {
                parent->release_display_frame(&current_frame);
        }
        if (has_ready_frame) {
                parent->release_display_frame(&ready_frame);
        }
}

void Mixer::OutputChannel::output_frame(DisplayFrame frame)
{
        // Store this frame for display. Remove the ready frame if any
        // (it was seemingly never used).
        {
                unique_lock<mutex> lock(frame_mutex);
                if (has_ready_frame) {
                        parent->release_display_frame(&ready_frame);
                }
                ready_frame = frame;
                has_ready_frame = true;
        }

        if (new_frame_ready_callback) {
                new_frame_ready_callback();
        }

        // Reduce the number of callbacks by filtering duplicates. The reason
        // why we bother doing this is that Qt seemingly can get into a state
        // where its builds up an essentially unbounded queue of signals,
        // consuming more and more memory, and there's no good way of collapsing
        // user-defined signals or limiting the length of the queue.
        if (transition_names_updated_callback) {
                vector<string> transition_names = global_mixer->get_transition_names();
                bool changed = false;
                if (transition_names.size() != last_transition_names.size()) {
                        changed = true;
                } else {
                        for (unsigned i = 0; i < transition_names.size(); ++i) {
                                if (transition_names[i] != last_transition_names[i]) {
                                        changed = true;
                                        break;
                                }
                        }
                }
                if (changed) {
                        transition_names_updated_callback(transition_names);
                        last_transition_names = transition_names;
                }
        }
        if (name_updated_callback) {
                string name = global_mixer->get_channel_name(channel);
                if (name != last_name) {
                        name_updated_callback(name);
                        last_name = name;
                }
        }
        if (color_updated_callback) {
                string color = global_mixer->get_channel_color(channel);
                if (color != last_color) {
                        color_updated_callback(color);
                        last_color = color;
                }
        }
}

bool Mixer::OutputChannel::get_display_frame(DisplayFrame *frame)
{
        unique_lock<mutex> lock(frame_mutex);
        if (!has_current_frame && !has_ready_frame) {
                return false;
        }

        if (has_current_frame && has_ready_frame) {
                // We have a new ready frame. Toss the current one.
                parent->release_display_frame(&current_frame);
                has_current_frame = false;
        }
        if (has_ready_frame) {
                assert(!has_current_frame);
                current_frame = ready_frame;
                ready_frame.ready_fence.reset();  // Drop the refcount.
                ready_frame.input_frames.clear();  // Drop the refcounts.
                has_current_frame = true;
                has_ready_frame = false;
        }

        *frame = current_frame;
        return true;
}

void Mixer::OutputChannel::set_frame_ready_callback(Mixer::new_frame_ready_callback_t callback)
{
        new_frame_ready_callback = callback;
}

void Mixer::OutputChannel::set_transition_names_updated_callback(Mixer::transition_names_updated_callback_t callback)
{
        transition_names_updated_callback = callback;
}

void Mixer::OutputChannel::set_name_updated_callback(Mixer::name_updated_callback_t callback)
{
        name_updated_callback = callback;
}

void Mixer::OutputChannel::set_color_updated_callback(Mixer::color_updated_callback_t callback)
{
        color_updated_callback = callback;
}

mutex RefCountedGLsync::fence_lock;