-#define WIDTH 1280
-#define HEIGHT 720
#define EXTRAHEIGHT 30
#undef Success
#include "mixer.h"
#include <assert.h>
-#include <effect.h>
-#include <effect_chain.h>
-#include <effect_util.h>
#include <epoxy/egl.h>
-#include <features.h>
-#include <image_format.h>
#include <init.h>
-#include <overlay_effect.h>
-#include <padding_effect.h>
-#include <resample_effect.h>
-#include <resource_pool.h>
-#include <saturation_effect.h>
+#include <movit/effect_chain.h>
+#include <movit/effect_util.h>
+#include <movit/flat_input.h>
+#include <movit/image_format.h>
+#include <movit/resource_pool.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <time.h>
#include <util.h>
-#include <white_balance_effect.h>
-#include <ycbcr.h>
-#include <ycbcr_input.h>
+#include <algorithm>
#include <cmath>
#include <condition_variable>
#include <cstddef>
#include <mutex>
#include <string>
#include <thread>
+#include <utility>
#include <vector>
#include "bmusb/bmusb.h"
#include "context.h"
+#include "defs.h"
#include "h264encode.h"
#include "pbo_frame_allocator.h"
#include "ref_counted_gl_sync.h"
} // namespace
-Mixer::Mixer(const QSurfaceFormat &format)
- : httpd("test.ts", WIDTH, HEIGHT),
+Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
+ : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
+ num_cards(num_cards),
mixer_surface(create_surface(format)),
- h264_encoder_surface(create_surface(format))
+ h264_encoder_surface(create_surface(format)),
+ level_compressor(OUTPUT_FREQUENCY),
+ limiter(OUTPUT_FREQUENCY),
+ compressor(OUTPUT_FREQUENCY)
{
httpd.start(9095);
CHECK(init_movit(MOVIT_SHADER_DIR, MOVIT_DEBUG_OFF));
check_error();
+ // Since we allow non-bouncing 4:2:2 YCbCrInputs, effective subpixel precision
+ // will be halved when sampling them, and we need to compensate here.
+ movit_texel_subpixel_precision /= 2.0;
+
resource_pool.reset(new ResourcePool);
- theme.reset(new Theme("theme.lua", resource_pool.get()));
- output_channel[OUTPUT_LIVE].parent = this;
- output_channel[OUTPUT_PREVIEW].parent = this;
- output_channel[OUTPUT_INPUT0].parent = this;
- output_channel[OUTPUT_INPUT1].parent = this;
+ theme.reset(new Theme("theme.lua", resource_pool.get(), num_cards));
+ for (unsigned i = 0; i < NUM_OUTPUTS; ++i) {
+ output_channel[i].parent = this;
+ }
ImageFormat inout_format;
inout_format.color_space = COLORSPACE_sRGB;
h264_encoder.reset(new H264Encoder(h264_encoder_surface, WIDTH, HEIGHT, &httpd));
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
printf("Configuring card %d...\n", card_index);
CaptureCard *card = &cards[card_index];
card->usb = new BMUSBCapture(card_index);
card->usb->set_dequeue_thread_callbacks(
[card]{
eglBindAPI(EGL_OPENGL_API);
- card->context = create_context();
+ card->context = create_context(card->surface);
if (!make_current(card->context, card->surface)) {
printf("failed to create bmusb context\n");
exit(1);
[this]{
resource_pool->clean_context();
});
- card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
+ card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
card->usb->configure_card();
}
BMUSBCapture::start_bm_thread();
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
cards[card_index].usb->start_bm_capture();
}
"} \n";
cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
- r128.init(2, 48000);
+ r128.init(2, OUTPUT_FREQUENCY);
r128.integr_start();
+
+ locut.init(FILTER_HPF, 2);
+
+ // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
+ // and there's a limit to how important the peak meter is.
+ peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
+
+ alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
}
Mixer::~Mixer()
resource_pool->release_glsl_program(cbcr_program_num);
BMUSBCapture::stop_bm_thread();
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ 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].usb->stop_dequeue_thread();
}
+
+ h264_encoder.reset(nullptr);
}
namespace {
}
}
-float find_peak(const vector<float> &samples)
+float find_peak(const float *samples, size_t num_samples)
{
float m = fabs(samples[0]);
- for (size_t i = 1; i < samples.size(); ++i) {
+ for (size_t i = 1; i < num_samples; ++i) {
m = std::max(m, fabs(samples[i]));
}
return m;
}
}
+// Returns length of a frame with the given format, in TIMEBASE units.
+int64_t find_frame_length(uint16_t video_format)
+{
+ if (video_format == 0x0800) {
+ // No video signal. These green pseudo-frames seem to come at about 30.13 Hz.
+ // It's a strange thing, but what can you do.
+ return TIMEBASE * 100 / 3013;
+ }
+ if ((video_format & 0xe800) != 0xe800) {
+ printf("Video format 0x%04x does not appear to be a video format. Assuming 60 Hz.\n",
+ video_format);
+ return TIMEBASE / 60;
+ }
+
+ // 0x8 seems to be a flag about availability of deep color on the input,
+ // except when it's not (e.g. it's the only difference between NTSC 23.98
+ // and PAL). Rather confusing. But we clear it here nevertheless, because
+ // usually it doesn't mean anything.
+ //
+ // We don't really handle interlaced formats at all yet.
+ uint16_t normalized_video_format = video_format & ~0xe808;
+ if (normalized_video_format == 0x0143) { // 720p50.
+ return TIMEBASE / 50;
+ } else if (normalized_video_format == 0x0103) { // 720p60.
+ return TIMEBASE / 60;
+ } else if (normalized_video_format == 0x0121) { // 720p59.94.
+ return TIMEBASE * 1001 / 60000;
+ } else if (normalized_video_format == 0x01c3 || // 1080p30.
+ normalized_video_format == 0x0003) { // 1080i60.
+ return TIMEBASE / 30;
+ } else if (normalized_video_format == 0x01e1 || // 1080p29.97.
+ normalized_video_format == 0x0021 || // 1080i59.94.
+ video_format == 0xe901 || // NTSC (480i59.94, I suppose).
+ video_format == 0xe9c1 || // Ditto.
+ video_format == 0xe801) { // Ditto.
+ return TIMEBASE * 1001 / 30000;
+ } else if (normalized_video_format == 0x0063 || // 1080p25.
+ normalized_video_format == 0x0043 || // 1080i50.
+ video_format == 0xe909) { // PAL (576i50, I suppose).
+ return TIMEBASE / 25;
+ } else if (normalized_video_format == 0x008e) { // 1080p24.
+ return TIMEBASE / 24;
+ } else if (normalized_video_format == 0x00a1) { // 1080p23.98.
+ return TIMEBASE * 1001 / 24000;
+ return TIMEBASE / 25;
+ } else {
+ printf("Unknown video format 0x%04x. Assuming 60 Hz.\n", video_format);
+ return TIMEBASE / 60;
+ }
+}
+
} // namespace
-void Mixer::bm_frame(int card_index, uint16_t timecode,
+void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
FrameAllocator::Frame video_frame, size_t video_offset, uint16_t video_format,
FrameAllocator::Frame audio_frame, size_t audio_offset, uint16_t audio_format)
{
CaptureCard *card = &cards[card_index];
+ int64_t frame_length = find_frame_length(video_format);
+
if (audio_frame.len - audio_offset > 30000) {
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),
return;
}
+ int64_t local_pts = card->next_local_pts;
+ int dropped_frames = 0;
+ if (card->last_timecode != -1) {
+ dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
+ }
+ card->last_timecode = timecode;
+
// Convert the audio to stereo fp32 and add it.
- size_t num_samples = (audio_frame.len - audio_offset) / 8 / 3;
+ size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
vector<float> audio;
audio.resize(num_samples * 2);
convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
- int unwrapped_timecode = timecode;
- int dropped_frames = 0;
- if (card->last_timecode != -1) {
- unwrapped_timecode = unwrap_timecode(unwrapped_timecode, card->last_timecode);
- dropped_frames = unwrapped_timecode - card->last_timecode - 1;
- }
- card->last_timecode = unwrapped_timecode;
-
// Add the audio.
{
unique_lock<mutex> lock(card->audio_mutex);
- int unwrapped_timecode = timecode;
- if (dropped_frames > 60 * 2) {
+ if (dropped_frames > MAX_FPS * 2) {
fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around), resetting resampler\n",
card_index);
- card->resampler.reset(new Resampler(48000.0, 48000.0, 2));
+ card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
} else if (dropped_frames > 0) {
- // Insert silence as needed.
+ // Insert silence as needed. (The number of samples could be nonintegral,
+ // but resampling will save us then.)
fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
card_index, dropped_frames, timecode);
vector<float> silence;
- silence.resize((48000 / 60) * 2);
+ silence.resize((OUTPUT_FREQUENCY * frame_length / TIMEBASE) * 2);
for (int i = 0; i < dropped_frames; ++i) {
- card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / 60.0, silence.data(), (48000 / 60));
+ card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence.size() / 2);
+ // Note that if the format changed in the meantime, we have
+ // no way of detecting that; we just have to assume the frame length
+ // is always the same.
+ local_pts += frame_length;
}
}
- card->resampler->add_input_samples(unwrapped_timecode / 60.0, audio.data(), num_samples);
+ card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
+ card->next_local_pts = local_pts + frame_length;
}
// Done with the audio, so release it.
unique_lock<mutex> lock(bmusb_mutex);
card->new_data_ready = true;
card->new_frame = RefCountedFrame(FrameAllocator::Frame());
+ card->new_frame_length = frame_length;
card->new_data_ready_fence = nullptr;
card->dropped_frames = dropped_frames;
card->new_data_ready_changed.notify_all();
unique_lock<mutex> lock(bmusb_mutex);
card->new_data_ready = true;
card->new_frame = RefCountedFrame(video_frame);
+ card->new_frame_length = frame_length;
card->new_data_ready_fence = fence;
card->dropped_frames = dropped_frames;
card->new_data_ready_changed.notify_all();
void Mixer::thread_func()
{
eglBindAPI(EGL_OPENGL_API);
- QOpenGLContext *context = create_context();
+ QOpenGLContext *context = create_context(mixer_surface);
if (!make_current(context, mixer_surface)) {
printf("oops\n");
exit(1);
int dropped_frames = 0;
while (!should_quit) {
- CaptureCard card_copy[NUM_CARDS];
+ CaptureCard card_copy[MAX_CARDS];
+ int num_samples[MAX_CARDS];
{
unique_lock<mutex> lock(bmusb_mutex);
// TODO: Make configurable, and with a timeout.
cards[0].new_data_ready_changed.wait(lock, [this]{ return cards[0].new_data_ready; });
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
CaptureCard *card = &cards[card_index];
card_copy[card_index].usb = card->usb;
card_copy[card_index].new_data_ready = card->new_data_ready;
card_copy[card_index].new_frame = card->new_frame;
+ card_copy[card_index].new_frame_length = card->new_frame_length;
card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
- card_copy[card_index].new_frame_audio = move(card->new_frame_audio);
card_copy[card_index].dropped_frames = card->dropped_frames;
card->new_data_ready = false;
card->new_data_ready_changed.notify_all();
+
+ int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
+ num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
+ card->fractional_samples = num_samples_times_timebase % TIMEBASE;
}
}
// Resample the audio as needed, including from previously dropped frames.
- vector<float> samples_out;
- // TODO: Allow using audio from the other card(s) as well.
for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
- for (unsigned card_index = 0; card_index < NUM_CARDS; ++card_index) {
- samples_out.resize((48000 / 60) * 2);
- {
- unique_lock<mutex> lock(cards[card_index].audio_mutex);
- if (!cards[card_index].resampler->get_output_samples(pts(), &samples_out[0], 48000 / 60)) {
- printf("Card %d reported previous underrun.\n", card_index);
- }
- }
- if (card_index == 0) {
- vector<float> left, right;
- peak = std::max(peak, find_peak(samples_out));
- deinterleave_samples(samples_out, &left, &right);
- float *ptrs[] = { left.data(), right.data() };
- r128.process(left.size(), ptrs);
- h264_encoder->add_audio(pts_int, move(samples_out));
- }
+ {
+ // Signal to the audio thread to process this frame.
+ unique_lock<mutex> lock(audio_mutex);
+ audio_task_queue.push(AudioTask{pts_int, num_samples[0]});
+ audio_task_queue_changed.notify_one();
}
if (frame_num != card_copy[0].dropped_frames) {
- // For dropped frames, increase the pts.
+ // For dropped frames, increase the pts. Note that if the format changed
+ // in the meantime, we have no way of detecting that; we just have to
+ // assume the frame length is always the same.
++dropped_frames;
- pts_int += TIMEBASE / 60;
+ pts_int += card_copy[0].new_frame_length;
}
}
double loudness_range_high = r128.range_max();
audio_level_callback(loudness_s, 20.0 * log10(peak),
- loudness_i, loudness_range_low, loudness_range_high);
+ loudness_i, loudness_range_low, loudness_range_high,
+ last_gain_staging_db);
+ }
+
+ for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
+ if (card_copy[card_index].new_data_ready && card_copy[card_index].new_frame->len == 0) {
+ ++card_copy[card_index].dropped_frames;
+ }
+ if (card_copy[card_index].dropped_frames > 0) {
+ printf("Card %u dropped %d frames before this\n",
+ card_index, int(card_copy[card_index].dropped_frames));
+ }
}
// If the first card is reporting a corrupted or otherwise dropped frame,
// just increase the pts (skipping over this frame) and don't try to compute anything new.
if (card_copy[0].new_frame->len == 0) {
++dropped_frames;
- pts_int += TIMEBASE / 60;
+ pts_int += card_copy[0].new_frame_length;
continue;
}
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
CaptureCard *card = &card_copy[card_index];
if (!card->new_data_ready || card->new_frame->len == 0)
continue;
// input frames needed, so that they are not released back
// until the rendering is done.
vector<RefCountedFrame> input_frames;
- for (int card_index = 0; card_index < NUM_CARDS; ++card_index) {
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
input_frames.push_back(bmusb_current_rendering_frame[card_index]);
}
- const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
+ const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
++frame;
- pts_int += TIMEBASE / 60;
+ pts_int += card_copy[0].new_frame_length;
// The live frame just shows the RGBA texture we just rendered.
// It owns rgba_tex now.
display_frame.chain = chain.first;
display_frame.setup_chain = chain.second;
display_frame.ready_fence = fence;
- display_frame.input_frames = { bmusb_current_rendering_frame[0], bmusb_current_rendering_frame[1] }; // FIXME: possible to do better?
+
+ // FIXME: possible to do better?
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ display_frame.input_frames.push_back(bmusb_current_rendering_frame[card_index]);
+ }
display_frame.temp_textures = {};
output_channel[i].output_frame(display_frame);
}
resource_pool->clean_context();
}
+void Mixer::audio_thread_func()
+{
+ while (!should_quit) {
+ AudioTask task;
+
+ {
+ unique_lock<mutex> lock(audio_mutex);
+ audio_task_queue_changed.wait(lock, [this]{ return !audio_task_queue.empty(); });
+ task = audio_task_queue.front();
+ audio_task_queue.pop();
+ }
+
+ process_audio_one_frame(task.pts_int, task.num_samples);
+ }
+}
+
+void Mixer::process_audio_one_frame(int64_t frame_pts_int, int num_samples)
+{
+ vector<float> samples_card;
+ vector<float> samples_out;
+ for (unsigned card_index = 0; card_index < num_cards; ++card_index) {
+ samples_card.resize(num_samples * 2);
+ {
+ unique_lock<mutex> lock(cards[card_index].audio_mutex);
+ if (!cards[card_index].resampling_queue->get_output_samples(double(frame_pts_int) / TIMEBASE, &samples_card[0], num_samples)) {
+ printf("Card %d reported previous underrun.\n", card_index);
+ }
+ }
+ // TODO: Allow using audio from the other card(s) as well.
+ if (card_index == 0) {
+ samples_out = move(samples_card);
+ }
+ }
+
+ // Cut away everything under 120 Hz (or whatever the cutoff is);
+ // we don't need it for voice, and it will reduce headroom
+ // and confuse the compressor. (In particular, any hums at 50 or 60 Hz
+ // should be dampened.)
+ locut.render(samples_out.data(), samples_out.size() / 2, locut_cutoff_hz * 2.0 * M_PI / OUTPUT_FREQUENCY, 0.5f);
+
+ // Apply a level compressor to get the general level right.
+ // Basically, if it's over about -40 dBFS, we squeeze it down to that level
+ // (or more precisely, near it, since we don't use infinite ratio),
+ // then apply a makeup gain to get it to -14 dBFS. -14 dBFS is, of course,
+ // entirely arbitrary, but from practical tests with speech, it seems to
+ // put ut around -23 LUFS, so it's a reasonable starting point for later use.
+ float ref_level_dbfs = -14.0f;
+ {
+ float threshold = 0.01f; // -40 dBFS.
+ float ratio = 20.0f;
+ float attack_time = 0.5f;
+ float release_time = 20.0f;
+ float makeup_gain = pow(10.0f, (ref_level_dbfs - (-40.0f)) / 20.0f); // +26 dB.
+ level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+ last_gain_staging_db = 20.0 * log10(level_compressor.get_attenuation() * makeup_gain);
+ }
+
+#if 0
+ printf("level=%f (%+5.2f dBFS) attenuation=%f (%+5.2f dB) end_result=%+5.2f dB\n",
+ level_compressor.get_level(), 20.0 * log10(level_compressor.get_level()),
+ level_compressor.get_attenuation(), 20.0 * log10(level_compressor.get_attenuation()),
+ 20.0 * log10(level_compressor.get_level() * level_compressor.get_attenuation() * makeup_gain));
+#endif
+
+// float limiter_att, compressor_att;
+
+ // The real compressor.
+ if (compressor_enabled) {
+ float threshold = pow(10.0f, compressor_threshold_dbfs / 20.0f);
+ float ratio = 20.0f;
+ float attack_time = 0.005f;
+ float release_time = 0.040f;
+ float makeup_gain = 2.0f; // +6 dB.
+ compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+// compressor_att = compressor.get_attenuation();
+ }
+
+ // Finally a limiter at -4 dB (so, -10 dBFS) to take out the worst peaks only.
+ // Note that since ratio is not infinite, we could go slightly higher than this.
+ if (limiter_enabled) {
+ float threshold = pow(10.0f, limiter_threshold_dbfs / 20.0f);
+ float ratio = 30.0f;
+ float attack_time = 0.0f; // Instant.
+ float release_time = 0.020f;
+ float makeup_gain = 1.0f; // 0 dB.
+ limiter.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+// limiter_att = limiter.get_attenuation();
+ }
+
+// printf("limiter=%+5.1f compressor=%+5.1f\n", 20.0*log10(limiter_att), 20.0*log10(compressor_att));
+
+ // Upsample 4x to find interpolated peak.
+ peak_resampler.inp_data = samples_out.data();
+ peak_resampler.inp_count = samples_out.size() / 2;
+
+ vector<float> interpolated_samples_out;
+ interpolated_samples_out.resize(samples_out.size());
+ while (peak_resampler.inp_count > 0) { // About four iterations.
+ peak_resampler.out_data = &interpolated_samples_out[0];
+ peak_resampler.out_count = interpolated_samples_out.size() / 2;
+ peak_resampler.process();
+ size_t out_stereo_samples = interpolated_samples_out.size() / 2 - peak_resampler.out_count;
+ peak = max<float>(peak, find_peak(interpolated_samples_out.data(), out_stereo_samples * 2));
+ }
+
+ // Find R128 levels.
+ vector<float> left, right;
+ deinterleave_samples(samples_out, &left, &right);
+ float *ptrs[] = { left.data(), right.data() };
+ r128.process(left.size(), ptrs);
+
+ // Send the samples to the sound card.
+ if (alsa) {
+ alsa->write(samples_out);
+ }
+
+ // And finally add them to the output.
+ h264_encoder->add_audio(frame_pts_int, move(samples_out));
+}
+
void Mixer::subsample_chroma(GLuint src_tex, GLuint dst_tex)
{
GLuint vao;
void Mixer::start()
{
mixer_thread = thread(&Mixer::thread_func, this);
+ audio_thread = thread(&Mixer::audio_thread_func, this);
}
void Mixer::quit()
{
should_quit = true;
mixer_thread.join();
+ audio_thread.join();
}
void Mixer::transition_clicked(int transition_num)
theme->channel_clicked(preview_num);
}
+void Mixer::reset_meters()
+{
+ peak_resampler.reset();
+ peak = 0.0f;
+ r128.reset();
+ r128.integr_start();
+}
+
Mixer::OutputChannel::~OutputChannel()
{
if (has_current_frame) {
projects / nageru / blobdiff