-#define WIDTH 1280
-#define HEIGHT 720
-#define EXTRAHEIGHT 30
-
#undef Success
#include "mixer.h"
#include <movit/flat_input.h>
#include <movit/image_format.h>
#include <movit/resource_pool.h>
+#include <movit/util.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <time.h>
-#include <util.h>
#include <algorithm>
#include <cmath>
#include <condition_variable>
}
}
+void insert_new_frame(RefCountedFrame frame, unsigned field_num, bool interlaced, unsigned card_index, InputState *input_state)
+{
+ if (interlaced) {
+ for (unsigned frame_num = FRAME_HISTORY_LENGTH; frame_num --> 1; ) { // :-)
+ input_state->buffered_frames[card_index][frame_num] =
+ input_state->buffered_frames[card_index][frame_num - 1];
+ }
+ input_state->buffered_frames[card_index][0] = { frame, field_num };
+ } else {
+ for (unsigned frame_num = 0; frame_num < FRAME_HISTORY_LENGTH; ++frame_num) {
+ input_state->buffered_frames[card_index][frame_num] = { frame, field_num };
+ }
+ }
+}
+
} // namespace
Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
- : httpd("test.ts", WIDTH, HEIGHT),
+ : httpd(LOCAL_DUMP_FILE_NAME, WIDTH, HEIGHT),
num_cards(num_cards),
mixer_surface(create_surface(format)),
h264_encoder_surface(create_surface(format)),
- level_compressor(OUTPUT_FREQUENCY)
+ level_compressor(OUTPUT_FREQUENCY),
+ limiter(OUTPUT_FREQUENCY),
+ compressor(OUTPUT_FREQUENCY)
{
httpd.start(9095);
CaptureCard *card = &cards[card_index];
card->usb = new BMUSBCapture(card_index);
card->usb->set_frame_callback(bind(&Mixer::bm_frame, this, card_index, _1, _2, _3, _4, _5, _6, _7));
- card->frame_allocator.reset(new PBOFrameAllocator(WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44, WIDTH, HEIGHT));
+ card->frame_allocator.reset(new PBOFrameAllocator(8 << 20, WIDTH, HEIGHT)); // 8 MB.
card->usb->set_video_frame_allocator(card->frame_allocator.get());
card->surface = create_surface(format);
card->usb->set_dequeue_thread_callbacks(
[this]{
resource_pool->clean_context();
});
- card->resampler.reset(new Resampler(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
+ card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
card->usb->configure_card();
}
cards[card_index].usb->start_bm_capture();
}
- //chain->enable_phase_timing(true);
-
// Set up stuff for NV12 conversion.
// Cb/Cr shader.
"void main() { \n"
" gl_FragColor = texture2D(cbcr_tex, tc0); \n"
"} \n";
- cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader);
+ vector<string> frag_shader_outputs;
+ cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
r128.init(2, OUTPUT_FREQUENCY);
r128.integr_start();
+
+ locut.init(FILTER_HPF, 2);
+
+ // hlen=16 is pretty low quality, but we use quite a bit of CPU otherwise,
+ // and there's a limit to how important the peak meter is.
+ peak_resampler.setup(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY * 4, /*num_channels=*/2, /*hlen=*/16);
+
+ alsa.reset(new ALSAOutput(OUTPUT_FREQUENCY, /*num_channels=*/2));
}
Mixer::~Mixer()
}
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;
{
CaptureCard *card = &cards[card_index];
- if (audio_frame.len - audio_offset > 30000) {
+ unsigned width, height, second_field_start, frame_rate_nom, frame_rate_den, extra_lines_top, extra_lines_bottom;
+ bool interlaced;
+
+ decode_video_format(video_format, &width, &height, &second_field_start, &extra_lines_top, &extra_lines_bottom,
+ &frame_rate_nom, &frame_rate_den, &interlaced); // Ignore return value for now.
+ int64_t frame_length = TIMEBASE * frame_rate_den / frame_rate_nom;
+
+ size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
+ if (num_samples > OUTPUT_FREQUENCY / 10) {
printf("Card %d: Dropping frame with implausible audio length (len=%d, offset=%d) [timecode=0x%04x video_len=%d video_offset=%d video_format=%x)\n",
card_index, int(audio_frame.len), int(audio_offset),
timecode, int(video_frame.len), int(video_offset), video_format);
return;
}
- int unwrapped_timecode = timecode;
+ int64_t local_pts = card->next_local_pts;
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;
+ dropped_frames = unwrap_timecode(timecode, card->last_timecode) - card->last_timecode - 1;
}
- card->last_timecode = unwrapped_timecode;
// Convert the audio to stereo fp32 and add it.
- size_t num_samples = (audio_frame.len >= audio_offset) ? (audio_frame.len - audio_offset) / 8 / 3 : 0;
vector<float> audio;
audio.resize(num_samples * 2);
convert_fixed24_to_fp32(&audio[0], 2, audio_frame.data + audio_offset, 8, num_samples);
{
unique_lock<mutex> lock(card->audio_mutex);
- int unwrapped_timecode = timecode;
- if (dropped_frames > 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(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
+ // Number of samples per frame if we need to insert silence.
+ // (Could be nonintegral, but resampling will save us then.)
+ int silence_samples = OUTPUT_FREQUENCY * frame_rate_den / frame_rate_nom;
+
+ if (dropped_frames > MAX_FPS * 2) {
+ fprintf(stderr, "Card %d lost more than two seconds (or time code jumping around; from 0x%04x to 0x%04x), resetting resampler\n",
+ card_index, card->last_timecode, timecode);
+ card->resampling_queue.reset(new ResamplingQueue(OUTPUT_FREQUENCY, OUTPUT_FREQUENCY, 2));
+ dropped_frames = 0;
} else if (dropped_frames > 0) {
// Insert silence as needed.
fprintf(stderr, "Card %d dropped %d frame(s) (before timecode 0x%04x), inserting silence.\n",
card_index, dropped_frames, timecode);
- vector<float> silence;
- silence.resize((OUTPUT_FREQUENCY / FPS) * 2);
+ vector<float> silence(silence_samples * 2, 0.0f);
for (int i = 0; i < dropped_frames; ++i) {
- card->resampler->add_input_samples((unwrapped_timecode - dropped_frames + i) / double(FPS), silence.data(), (OUTPUT_FREQUENCY / FPS));
+ card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), silence.data(), silence_samples);
+ // Note that if the format changed in the meantime, we have
+ // no way of detecting that; we just have to assume the frame length
+ // is always the same.
+ local_pts += frame_length;
}
}
- card->resampler->add_input_samples(unwrapped_timecode / double(FPS), audio.data(), num_samples);
+ if (num_samples == 0) {
+ audio.resize(silence_samples * 2);
+ num_samples = silence_samples;
+ }
+ card->resampling_queue->add_input_samples(local_pts / double(TIMEBASE), audio.data(), num_samples);
+ card->next_local_pts = local_pts + frame_length;
}
+ card->last_timecode = timecode;
+
// Done with the audio, so release it.
if (audio_frame.owner) {
audio_frame.owner->release_frame(audio_frame);
if (card->should_quit) return;
}
- if (video_frame.len - video_offset != WIDTH * (HEIGHT+EXTRAHEIGHT) * 2) {
+ if (video_frame.len - video_offset == 0 ||
+ video_frame.len - video_offset != size_t(width * (height + extra_lines_top + extra_lines_bottom) * 2)) {
if (video_frame.len != 0) {
printf("Card %d: Dropping video frame with wrong length (%ld)\n",
card_index, video_frame.len - video_offset);
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_frame_interlaced = false;
card->new_data_ready_fence = nullptr;
card->dropped_frames = dropped_frames;
card->new_data_ready_changed.notify_all();
return;
}
- const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)video_frame.userdata;
- GLuint pbo = userdata->pbo;
- check_error();
- glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
- check_error();
- glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, 0, video_frame.size);
- check_error();
- //glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
- //check_error();
+ PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)video_frame.userdata;
+
+ unsigned num_fields = interlaced ? 2 : 1;
+ timespec frame_upload_start;
+ if (interlaced) {
+ // NOTE: This isn't deinterlacing. This is just sending the two fields along
+ // as separate frames without considering anything like the half-field offset.
+ // We'll need to add a proper deinterlacer on the receiving side to get this right.
+ assert(height % 2 == 0);
+ height /= 2;
+ assert(frame_length % 2 == 0);
+ frame_length /= 2;
+ num_fields = 2;
+ clock_gettime(CLOCK_MONOTONIC, &frame_upload_start);
+ }
+ RefCountedFrame new_frame(video_frame);
// Upload the textures.
- glBindTexture(GL_TEXTURE_2D, userdata->tex_y);
- check_error();
- glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH, HEIGHT, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET((WIDTH * (HEIGHT+EXTRAHEIGHT) * 2 + 44) / 2 + WIDTH * 25 + 22));
- check_error();
- glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr);
- check_error();
- glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, WIDTH/2, HEIGHT, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(WIDTH * 25 + 22));
- check_error();
- glBindTexture(GL_TEXTURE_2D, 0);
- check_error();
- GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
- check_error();
- assert(fence != nullptr);
+ size_t cbcr_width = width / 2;
+ size_t cbcr_offset = video_offset / 2;
+ size_t y_offset = video_frame.size / 2 + video_offset / 2;
+
+ for (unsigned field = 0; field < num_fields; ++field) {
+ unsigned field_start_line = (field == 1) ? second_field_start : extra_lines_top + field * (height + 22);
+
+ if (userdata->tex_y[field] == 0 ||
+ userdata->tex_cbcr[field] == 0 ||
+ width != userdata->last_width[field] ||
+ height != userdata->last_height[field]) {
+ // We changed resolution since last use of this texture, so we need to create
+ // a new object. Note that this each card has its own PBOFrameAllocator,
+ // we don't need to worry about these flip-flopping between resolutions.
+ glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
+ check_error();
+ glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, cbcr_width, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
+ check_error();
+ glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
+ check_error();
+ glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
+ check_error();
+ userdata->last_width[field] = width;
+ userdata->last_height[field] = height;
+ }
- {
- unique_lock<mutex> lock(bmusb_mutex);
- card->new_data_ready = true;
- card->new_frame = RefCountedFrame(video_frame);
- card->new_data_ready_fence = fence;
- card->dropped_frames = dropped_frames;
- card->new_data_ready_changed.notify_all();
+ GLuint pbo = userdata->pbo;
+ check_error();
+ glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbo);
+ check_error();
+ glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
+ check_error();
+
+ glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
+ check_error();
+ glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, cbcr_width, height, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset + cbcr_width * field_start_line * sizeof(uint16_t)));
+ check_error();
+ glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
+ check_error();
+ glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(y_offset + width * field_start_line));
+ check_error();
+ glBindTexture(GL_TEXTURE_2D, 0);
+ check_error();
+ GLsync fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
+ check_error();
+ assert(fence != nullptr);
+
+ if (field == 1) {
+ // Don't upload the second field as fast as we can; wait until
+ // the field time has approximately passed. (Otherwise, we could
+ // get timing jitter against the other sources, and possibly also
+ // against the video display, although the latter is not as critical.)
+ // This requires our system clock to be reasonably close to the
+ // video clock, but that's not an unreasonable assumption.
+ timespec second_field_start;
+ second_field_start.tv_nsec = frame_upload_start.tv_nsec +
+ frame_length * 1000000000 / TIMEBASE;
+ second_field_start.tv_sec = frame_upload_start.tv_sec +
+ second_field_start.tv_nsec / 1000000000;
+ second_field_start.tv_nsec %= 1000000000;
+
+ while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME,
+ &second_field_start, nullptr) == -1 &&
+ errno == EINTR) ;
+ }
+
+ {
+ unique_lock<mutex> lock(bmusb_mutex);
+ card->new_data_ready = true;
+ card->new_frame = new_frame;
+ card->new_frame_length = frame_length;
+ card->new_frame_field = field;
+ card->new_frame_interlaced = interlaced;
+ card->new_data_ready_fence = fence;
+ card->dropped_frames = dropped_frames;
+ card->new_data_ready_changed.notify_all();
+
+ if (field != num_fields - 1) {
+ // Wait until the previous frame was consumed.
+ card->new_data_ready_changed.wait(lock, [card]{ return !card->new_data_ready || card->should_quit; });
+ if (card->should_quit) return;
+ }
+ }
}
}
clock_gettime(CLOCK_MONOTONIC, &start);
int frame = 0;
- int dropped_frames = 0;
+ int stats_dropped_frames = 0;
while (!should_quit) {
CaptureCard card_copy[MAX_CARDS];
+ int num_samples[MAX_CARDS];
{
unique_lock<mutex> lock(bmusb_mutex);
card_copy[card_index].usb = card->usb;
card_copy[card_index].new_data_ready = card->new_data_ready;
card_copy[card_index].new_frame = card->new_frame;
+ card_copy[card_index].new_frame_length = card->new_frame_length;
+ card_copy[card_index].new_frame_field = card->new_frame_field;
+ card_copy[card_index].new_frame_interlaced = card->new_frame_interlaced;
card_copy[card_index].new_data_ready_fence = card->new_data_ready_fence;
card_copy[card_index].dropped_frames = card->dropped_frames;
card->new_data_ready = false;
card->new_data_ready_changed.notify_all();
+
+ int num_samples_times_timebase = OUTPUT_FREQUENCY * card->new_frame_length + card->fractional_samples;
+ num_samples[card_index] = num_samples_times_timebase / TIMEBASE;
+ card->fractional_samples = num_samples_times_timebase % TIMEBASE;
+ assert(num_samples[card_index] >= 0);
}
}
// Resample the audio as needed, including from previously dropped frames.
for (unsigned frame_num = 0; frame_num < card_copy[0].dropped_frames + 1; ++frame_num) {
- process_audio_one_frame();
+ {
+ // 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.
- ++dropped_frames;
- pts_int += TIMEBASE / FPS;
+ // For dropped frames, increase the pts. Note that if the format changed
+ // in the meantime, we have no way of detecting that; we just have to
+ // assume the frame length is always the same.
+ ++stats_dropped_frames;
+ pts_int += card_copy[0].new_frame_length;
}
}
if (audio_level_callback != nullptr) {
+ unique_lock<mutex> lock(r128_mutex);
double loudness_s = r128.loudness_S();
double loudness_i = r128.integrated();
double loudness_range_low = r128.range_min();
double loudness_range_high = r128.range_max();
audio_level_callback(loudness_s, 20.0 * log10(peak),
- loudness_i, loudness_range_low, loudness_range_high);
+ loudness_i, loudness_range_low, loudness_range_high,
+ last_gain_staging_db);
}
for (unsigned card_index = 1; card_index < num_cards; ++card_index) {
// If the first card is reporting a corrupted or otherwise dropped frame,
// just increase the pts (skipping over this frame) and don't try to compute anything new.
if (card_copy[0].new_frame->len == 0) {
- ++dropped_frames;
- pts_int += TIMEBASE / FPS;
+ ++stats_dropped_frames;
+ pts_int += card_copy[0].new_frame_length;
continue;
}
continue;
assert(card->new_frame != nullptr);
- bmusb_current_rendering_frame[card_index] = card->new_frame;
+ insert_new_frame(card->new_frame, card->new_frame_field, card->new_frame_interlaced, card_index, &input_state);
check_error();
// The new texture might still be uploaded,
glDeleteSync(card->new_data_ready_fence);
check_error();
}
- const PBOFrameAllocator::Userdata *userdata = (const PBOFrameAllocator::Userdata *)card->new_frame->userdata;
- theme->set_input_textures(card_index, userdata->tex_y, userdata->tex_cbcr);
}
// Get the main chain from the theme, and set its state immediately.
- pair<EffectChain *, function<void()>> theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT);
- EffectChain *chain = theme_main_chain.first;
- theme_main_chain.second();
+ Theme::Chain theme_main_chain = theme->get_chain(0, pts(), WIDTH, HEIGHT, input_state);
+ EffectChain *chain = theme_main_chain.chain;
+ theme_main_chain.setup_chain();
+ //theme_main_chain.chain->enable_phase_timing(true);
GLuint y_tex, cbcr_tex;
bool got_frame = h264_encoder->begin_frame(&y_tex, &cbcr_tex);
RefCountedGLsync fence(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
check_error();
- // Make sure the H.264 gets a reference to all the
- // input frames needed, so that they are not released back
- // until the rendering is done.
- vector<RefCountedFrame> input_frames;
- 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.
- h264_encoder->end_frame(fence, pts_int + av_delay, input_frames);
+ const int64_t av_delay = TIMEBASE / 10; // Corresponds to the fixed delay in resampling_queue.h. TODO: Make less hard-coded.
+ h264_encoder->end_frame(fence, pts_int + av_delay, theme_main_chain.input_frames);
++frame;
- pts_int += TIMEBASE / FPS;
+ pts_int += card_copy[0].new_frame_length;
// The live frame just shows the RGBA texture we just rendered.
// It owns rgba_tex now.
// Set up preview and any additional channels.
for (int i = 1; i < theme->get_num_channels() + 2; ++i) {
DisplayFrame display_frame;
- pair<EffectChain *, function<void()>> chain = theme->get_chain(i, pts(), WIDTH, HEIGHT); // FIXME: dimensions
- display_frame.chain = chain.first;
- display_frame.setup_chain = chain.second;
+ 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;
-
- // 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.input_frames = chain.input_frames;
display_frame.temp_textures = {};
output_channel[i].output_frame(display_frame);
}
1e-9 * (now.tv_nsec - start.tv_nsec);
if (frame % 100 == 0) {
printf("%d frames (%d dropped) in %.3f seconds = %.1f fps (%.1f ms/frame)\n",
- frame, dropped_frames, elapsed, frame / elapsed,
+ frame, stats_dropped_frames, elapsed, frame / elapsed,
1e3 * elapsed / frame);
// chain->print_phase_timing();
}
resource_pool->clean_context();
}
-void Mixer::process_audio_one_frame()
+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((OUTPUT_FREQUENCY / FPS) * 2);
+ samples_card.resize(num_samples * 2);
{
unique_lock<mutex> lock(cards[card_index].audio_mutex);
- if (!cards[card_index].resampler->get_output_samples(pts(), &samples_card[0], OUTPUT_FREQUENCY / FPS)) {
+ 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);
}
}
}
}
+ // 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 -12 dBFS. -12 dBFS is, of course,
+ // 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.
- //
- // TODO: Hook this up to a UI, so we can see the effects, and/or turn it off
- // to control the gain manually instead. For now, there's only the #if-ed out
- // code below.
- //
- // TODO: Add the actual compressors/limiters (for taking care of transients)
- // later in the chain.
- float threshold = 0.01f; // -40 dBFS.
- float ratio = 20.0f;
- float attack_time = 0.1f;
- float release_time = 10.0f;
- float makeup_gain = pow(10.0f, 28.0f / 20.0f); // +28 dB takes us to -12 dBFS.
- level_compressor.process(samples_out.data(), samples_out.size() / 2, threshold, ratio, attack_time, release_time, makeup_gain);
+ 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",
- compressor.get_level(), 20.0 * log10(compressor.get_level()),
- compressor.get_attenuation(), 20.0 * log10(compressor.get_attenuation()),
- 20.0 * log10(compressor.get_level() * compressor.get_attenuation() * makeup_gain));
+ 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
- // Find peak and R128 levels.
- peak = std::max(peak, find_peak(samples_out));
+// 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);
+ {
+ unique_lock<mutex> lock(r128_mutex);
+ r128.process(left.size(), ptrs);
+ }
- // Actually add the samples to the output.
- h264_encoder->add_audio(pts_int, move(samples_out));
+ // 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)
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) {