#include "resampling_queue.h"
#include <assert.h>
-#include <math.h>
-#include <stddef.h>
#include <stdio.h>
+#include <stdlib.h>
#include <string.h>
#include <zita-resampler/vresampler.h>
+#include <algorithm>
+#include <cmath>
-ResamplingQueue::ResamplingQueue(unsigned card_num, unsigned freq_in, unsigned freq_out, unsigned num_channels)
- : card_num(card_num), freq_in(freq_in), freq_out(freq_out), num_channels(num_channels),
- ratio(double(freq_out) / double(freq_in))
+using namespace std;
+using namespace std::chrono;
+
+ResamplingQueue::ResamplingQueue(DeviceSpec device_spec, unsigned freq_in, unsigned freq_out, unsigned num_channels, double expected_delay_seconds)
+ : device_spec(device_spec), freq_in(freq_in), freq_out(freq_out), num_channels(num_channels),
+ current_estimated_freq_in(freq_in),
+ ratio(double(freq_out) / double(freq_in)), expected_delay(expected_delay_seconds * OUTPUT_FREQUENCY)
{
vresampler.setup(ratio, num_channels, /*hlen=*/32);
vresampler.process ();
}
-void ResamplingQueue::add_input_samples(double pts, const float *samples, ssize_t num_samples)
+void ResamplingQueue::add_input_samples(steady_clock::time_point ts, const float *samples, ssize_t num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
{
if (num_samples == 0) {
return;
}
- if (first_input) {
- // Synthesize a fake length.
- last_input_len = double(num_samples) / freq_in;
- first_input = false;
- } else {
- last_input_len = pts - last_input_pts;
- }
- last_input_pts = pts;
+ assert(duration<double>(ts.time_since_epoch()).count() >= 0.0);
- k_a0 = k_a1;
- k_a1 += num_samples;
+ bool good_sample = (rate_adjustment_policy == ADJUST_RATE);
+ if (good_sample && a1.good_sample) {
+ a0 = a1;
+ }
+ a1.ts = ts;
+ a1.input_samples_received += num_samples;
+ a1.good_sample = good_sample;
+ if (a0.good_sample && a1.good_sample) {
+ current_estimated_freq_in = (a1.input_samples_received - a0.input_samples_received) / duration<double>(a1.ts - a0.ts).count();
+ if (!(current_estimated_freq_in >= 0.0)) {
+ fprintf(stderr, "%s: PANIC: Input audio clock going backwards, ignoring.\n",
+ spec_to_string(device_spec).c_str());
+ current_estimated_freq_in = freq_in;
+ }
- for (ssize_t i = 0; i < num_samples * num_channels; ++i) {
- buffer.push_back(samples[i]);
+ // Bound the frequency, so that a single wild result won't throw the filter off guard.
+ current_estimated_freq_in = min(current_estimated_freq_in, 1.2 * freq_in);
+ current_estimated_freq_in = max(current_estimated_freq_in, 0.8 * freq_in);
}
+
+ buffer.insert(buffer.end(), samples, samples + num_samples * num_channels);
}
-bool ResamplingQueue::get_output_samples(double pts, float *samples, ssize_t num_samples)
+bool ResamplingQueue::get_output_samples(steady_clock::time_point ts, float *samples, ssize_t num_samples, ResamplingQueue::RateAdjustmentPolicy rate_adjustment_policy)
{
- if (first_input) {
+ assert(num_samples > 0);
+ if (a1.input_samples_received == 0) {
// No data yet, just return zeros.
- memset(samples, 0, num_samples * 2 * sizeof(float));
+ memset(samples, 0, num_samples * num_channels * sizeof(float));
return true;
}
- double last_output_len;
- if (first_output) {
- // Synthesize a fake length.
- last_output_len = double(num_samples) / freq_out;
- } else {
- last_output_len = pts - last_output_pts;
+ // This can happen when we get dropped frames on the master card.
+ if (duration<double>(ts.time_since_epoch()).count() <= 0.0) {
+ rate_adjustment_policy = DO_NOT_ADJUST_RATE;
}
- last_output_pts = pts;
-
- // Using the time point since just before the last call to add_input_samples() as a base,
- // estimate actual delay based on activity since then, measured in number of input samples:
- double actual_delay = 0.0;
- assert(last_input_len != 0);
- actual_delay += (k_a1 - k_a0) * last_output_len / last_input_len; // Inserted samples since k_a0, rescaled for the different time periods.
- actual_delay += k_a0 - total_consumed_samples; // Samples inserted before k_a0 but not consumed yet.
- actual_delay += vresampler.inpdist(); // Delay in the resampler itself.
- double err = actual_delay - expected_delay;
- if (first_output && err < 0.0) {
- // Before the very first block, insert artificial delay based on our initial estimate,
- // so that we don't need a long period to stabilize at the beginning.
- int delay_samples_to_add = lrintf(-err);
- for (ssize_t i = 0; i < delay_samples_to_add * num_channels; ++i) {
- buffer.push_front(0.0f);
+
+ if (rate_adjustment_policy == ADJUST_RATE && (a0.good_sample || a1.good_sample)) {
+ // Estimate the current number of input samples produced at
+ // this instant in time, by extrapolating from the last known
+ // good point. Note that we could be extrapolating backward or
+ // forward, depending on the timing of the calls.
+ const InputPoint &base_point = a1.good_sample ? a1 : a0;
+ assert(duration<double>(base_point.ts.time_since_epoch()).count() >= 0.0);
+
+ // NOTE: Due to extrapolation, input_samples_received can
+ // actually go negative here the few first calls (ie., we asked
+ // about a timestamp where we hadn't actually started producing
+ // samples yet), but that is harmless.
+ const double input_samples_received = base_point.input_samples_received +
+ current_estimated_freq_in * duration<double>(ts - base_point.ts).count();
+
+ // Estimate the number of input samples _consumed_ after we've run the resampler.
+ const double input_samples_consumed = total_consumed_samples +
+ num_samples / (ratio * rcorr);
+
+ double actual_delay = input_samples_received - input_samples_consumed;
+ actual_delay += vresampler.inpdist(); // Delay in the resampler itself.
+ double err = actual_delay - expected_delay;
+ if (first_output) {
+ // Before the very first block, insert artificial delay based on our initial estimate,
+ // so that we don't need a long period to stabilize at the beginning.
+ if (err < 0.0) {
+ int delay_samples_to_add = lrintf(-err);
+ for (ssize_t i = 0; i < delay_samples_to_add * num_channels; ++i) {
+ buffer.push_front(0.0f);
+ }
+ total_consumed_samples -= delay_samples_to_add; // Equivalent to increasing input_samples_received on a0 and a1.
+ err += delay_samples_to_add;
+ } else if (err > 0.0) {
+ int delay_samples_to_remove = min<int>(lrintf(err), buffer.size() / num_channels);
+ buffer.erase(buffer.begin(), buffer.begin() + delay_samples_to_remove * num_channels);
+ total_consumed_samples += delay_samples_to_remove;
+ err -= delay_samples_to_remove;
+ }
}
- total_consumed_samples -= delay_samples_to_add; // Equivalent to increasing k_a0 and k_a1.
- err += delay_samples_to_add;
+ first_output = false;
+
+ // Compute loop filter coefficients for the two filters. We need to compute them
+ // every time, since they depend on the number of samples the user asked for.
+ //
+ // The loop bandwidth is at 0.02 Hz; our jitter is pretty large
+ // since none of the threads involved run at real-time priority.
+ // However, the first four seconds, we use a larger loop bandwidth (2 Hz),
+ // because there's a lot going on during startup, and thus the
+ // initial estimate might be tainted by jitter during that phase,
+ // and we want to converge faster.
+ //
+ // NOTE: The above logic might only hold during Nageru startup
+ // (we start ResamplingQueues also when we e.g. switch sound sources),
+ // but in general, a little bit of increased timing jitter is acceptable
+ // right after a setup change like this.
+ double loop_bandwidth_hz = (total_consumed_samples < 4 * freq_in) ? 0.2 : 0.02;
+
+ // Set filters. The first filter much wider than the first one (20x as wide).
+ double w = (2.0 * M_PI) * loop_bandwidth_hz * num_samples / freq_out;
+ double w0 = 1.0 - exp(-20.0 * w);
+ double w1 = w * 1.5 / num_samples / ratio;
+ double w2 = w / 1.5;
+
+ // Filter <err> through the loop filter to find the correction ratio.
+ z1 += w0 * (w1 * err - z1);
+ z2 += w0 * (z1 - z2);
+ z3 += w2 * z2;
+ rcorr = 1.0 - z2 - z3;
+ if (rcorr > 1.05) rcorr = 1.05;
+ if (rcorr < 0.95) rcorr = 0.95;
+ assert(!isnan(rcorr));
+ vresampler.set_rratio(rcorr);
}
- first_output = false;
-
- // Compute loop filter coefficients for the two filters. We need to compute them
- // every time, since they depend on the number of samples the user asked for.
- //
- // The loop bandwidth is at 0.02 Hz; we trust the initial estimate quite well,
- // and our jitter is pretty large since none of the threads involved run at
- // real-time priority.
- double loop_bandwidth_hz = 0.02;
-
- // Set filters. The first filter much wider than the first one (20x as wide).
- double w = (2.0 * M_PI) * loop_bandwidth_hz * num_samples / freq_out;
- double w0 = 1.0 - exp(-20.0 * w);
- double w1 = w * 1.5 / num_samples / ratio;
- double w2 = w / 1.5;
-
- // Filter <err> through the loop filter to find the correction ratio.
- z1 += w0 * (w1 * err - z1);
- z2 += w0 * (z1 - z2);
- z3 += w2 * z2;
- double rcorr = 1.0 - z2 - z3;
- if (rcorr > 1.05) rcorr = 1.05;
- if (rcorr < 0.95) rcorr = 0.95;
- assert(!isnan(rcorr));
- vresampler.set_rratio(rcorr);
-
- // Finally actually resample, consuming exactly <num_samples> output samples.
+
+ // Finally actually resample, producing exactly <num_samples> output samples.
vresampler.out_data = samples;
vresampler.out_count = num_samples;
while (vresampler.out_count > 0) {
if (buffer.empty()) {
// This should never happen unless delay is set way too low,
// or we're dropping a lot of data.
- fprintf(stderr, "Card %u: PANIC: Out of input samples to resample, still need %d output samples! (correction factor is %f)\n",
- card_num, int(vresampler.out_count), rcorr);
- memset(vresampler.out_data, 0, vresampler.out_count * 2 * sizeof(float));
+ fprintf(stderr, "%s: PANIC: Out of input samples to resample, still need %d output samples! (correction factor is %f)\n",
+ spec_to_string(device_spec).c_str(), int(vresampler.out_count), rcorr);
+ memset(vresampler.out_data, 0, vresampler.out_count * num_channels * sizeof(float));
+
+ // Reset the loop filter.
+ z1 = z2 = z3 = 0.0;
+
return false;
}
if (num_input_samples * num_channels > buffer.size()) {
num_input_samples = buffer.size() / num_channels;
}
- for (size_t i = 0; i < num_input_samples * num_channels; ++i) {
- inbuf[i] = buffer[i];
- }
+ copy(buffer.begin(), buffer.begin() + num_input_samples * num_channels, inbuf);
vresampler.inp_count = num_input_samples;
vresampler.inp_data = inbuf;