Chris@0: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
Chris@0: 
Chris@0: #include "Resampler.h"
Chris@0: 
Chris@0: #include "qm-dsp/maths/MathUtilities.h"
Chris@0: #include "qm-dsp/base/KaiserWindow.h"
Chris@0: #include "qm-dsp/base/SincWindow.h"
Chris@9: #include "qm-dsp/thread/Thread.h"
Chris@0: 
Chris@0: #include <iostream>
Chris@1: #include <vector>
Chris@8: #include <map>
Chris@10: #include <cassert>
Chris@1: 
Chris@1: using std::vector;
Chris@8: using std::map;
Chris@0: 
Chris@4: //#define DEBUG_RESAMPLER 1
Chris@4: 
Chris@0: Resampler::Resampler(int sourceRate, int targetRate) :
Chris@0:     m_sourceRate(sourceRate),
Chris@0:     m_targetRate(targetRate)
Chris@0: {
Chris@16:     initialise(100, 0.02);
Chris@16: }
Chris@16: 
Chris@16: Resampler::Resampler(int sourceRate, int targetRate, 
Chris@16:                      double snr, double bandwidth) :
Chris@16:     m_sourceRate(sourceRate),
Chris@16:     m_targetRate(targetRate)
Chris@16: {
Chris@16:     initialise(snr, bandwidth);
Chris@0: }
Chris@0: 
Chris@0: Resampler::~Resampler()
Chris@0: {
Chris@0:     delete[] m_phaseData;
Chris@0: }
Chris@0: 
Chris@9: // peakToPole -> length -> beta -> window
Chris@9: static map<int, map<int, map<double, vector<double> > > >
Chris@9: knownFilters;
Chris@9: 
Chris@9: static Mutex
Chris@9: knownFilterMutex;
Chris@9: 
Chris@0: void
Chris@16: Resampler::initialise(double snr, double bandwidth)
Chris@0: {
Chris@0:     int higher = std::max(m_sourceRate, m_targetRate);
Chris@0:     int lower = std::min(m_sourceRate, m_targetRate);
Chris@0: 
Chris@0:     m_gcd = MathUtilities::gcd(lower, higher);
Chris@0: 
Chris@0:     int peakToPole = higher / m_gcd;
Chris@0: 
Chris@0:     KaiserWindow::Parameters params =
Chris@16: 	KaiserWindow::parametersForBandwidth(snr, bandwidth, peakToPole);
Chris@0: 
Chris@0:     params.length =
Chris@0: 	(params.length % 2 == 0 ? params.length + 1 : params.length);
Chris@0:     
Chris@10:     params.length =
Chris@10:         (params.length > 200001 ? 200001 : params.length);
Chris@10: 
Chris@0:     m_filterLength = params.length;
Chris@8: 
Chris@9:     vector<double> filter;
Chris@9:     knownFilterMutex.lock();
Chris@0: 
Chris@9:     if (knownFilters[peakToPole][m_filterLength].find(params.beta) ==
Chris@9: 	knownFilters[peakToPole][m_filterLength].end()) {
Chris@9: 
Chris@9: 	KaiserWindow kw(params);
Chris@9: 	SincWindow sw(m_filterLength, peakToPole * 2);
Chris@9: 
Chris@9: 	filter = vector<double>(m_filterLength, 0.0);
Chris@9: 	for (int i = 0; i < m_filterLength; ++i) filter[i] = 1.0;
Chris@9: 	sw.cut(filter.data());
Chris@9: 	kw.cut(filter.data());
Chris@9: 
Chris@9: 	knownFilters[peakToPole][m_filterLength][params.beta] = filter;
Chris@9:     }
Chris@9: 
Chris@9:     filter = knownFilters[peakToPole][m_filterLength][params.beta];
Chris@9:     knownFilterMutex.unlock();
Chris@0: 
Chris@0:     int inputSpacing = m_targetRate / m_gcd;
Chris@0:     int outputSpacing = m_sourceRate / m_gcd;
Chris@0: 
Chris@4: #ifdef DEBUG_RESAMPLER
Chris@4:     std::cerr << "resample " << m_sourceRate << " -> " << m_targetRate
Chris@4: 	      << ": inputSpacing " << inputSpacing << ", outputSpacing "
Chris@4: 	      << outputSpacing << ": filter length " << m_filterLength
Chris@4: 	      << std::endl;
Chris@4: #endif
Chris@0: 
Chris@10:     // Now we have a filter of (odd) length flen in which the lower
Chris@10:     // sample rate corresponds to every n'th point and the higher rate
Chris@10:     // to every m'th where n and m are higher and lower rates divided
Chris@10:     // by their gcd respectively. So if x coordinates are on the same
Chris@10:     // scale as our filter resolution, then source sample i is at i *
Chris@10:     // (targetRate / gcd) and target sample j is at j * (sourceRate /
Chris@10:     // gcd).
Chris@10: 
Chris@10:     // To reconstruct a single target sample, we want a buffer (real
Chris@10:     // or virtual) of flen values formed of source samples spaced at
Chris@10:     // intervals of (targetRate / gcd), in our example case 3.  This
Chris@10:     // is initially formed with the first sample at the filter peak.
Chris@10:     //
Chris@10:     // 0  0  0  0  a  0  0  b  0
Chris@10:     //
Chris@10:     // and of course we have our filter
Chris@10:     //
Chris@10:     // f1 f2 f3 f4 f5 f6 f7 f8 f9
Chris@10:     //
Chris@10:     // We take the sum of products of non-zero values from this buffer
Chris@10:     // with corresponding values in the filter
Chris@10:     //
Chris@10:     // a * f5 + b * f8
Chris@10:     //
Chris@10:     // Then we drop (sourceRate / gcd) values, in our example case 4,
Chris@10:     // from the start of the buffer and fill until it has flen values
Chris@10:     // again
Chris@10:     //
Chris@10:     // a  0  0  b  0  0  c  0  0
Chris@10:     //
Chris@10:     // repeat to reconstruct the next target sample
Chris@10:     //
Chris@10:     // a * f1 + b * f4 + c * f7
Chris@10:     //
Chris@10:     // and so on.
Chris@10:     //
Chris@10:     // Above I said the buffer could be "real or virtual" -- ours is
Chris@10:     // virtual. We don't actually store all the zero spacing values,
Chris@10:     // except for padding at the start; normally we store only the
Chris@10:     // values that actually came from the source stream, along with a
Chris@10:     // phase value that tells us how many virtual zeroes there are at
Chris@10:     // the start of the virtual buffer.  So the two examples above are
Chris@10:     //
Chris@10:     // 0 a b  [ with phase 1 ]
Chris@10:     // a b c  [ with phase 0 ]
Chris@10:     //
Chris@10:     // Having thus broken down the buffer so that only the elements we
Chris@10:     // need to multiply are present, we can also unzip the filter into
Chris@10:     // every-nth-element subsets at each phase, allowing us to do the
Chris@10:     // filter multiplication as a simply vector multiply. That is, rather
Chris@10:     // than store
Chris@10:     //
Chris@10:     // f1 f2 f3 f4 f5 f6 f7 f8 f9
Chris@10:     // 
Chris@10:     // we store separately
Chris@10:     //
Chris@10:     // f1 f4 f7
Chris@10:     // f2 f5 f8
Chris@10:     // f3 f6 f9
Chris@10:     //
Chris@10:     // Each time we complete a multiply-and-sum, we need to work out
Chris@10:     // how many (real) samples to drop from the start of our buffer,
Chris@10:     // and how many to add at the end of it for the next multiply.  We
Chris@10:     // know we want to drop enough real samples to move along by one
Chris@10:     // computed output sample, which is our outputSpacing number of
Chris@10:     // virtual buffer samples. Depending on the relationship between
Chris@10:     // input and output spacings, this may mean dropping several real
Chris@10:     // samples, one real sample, or none at all (and simply moving to
Chris@10:     // a different "phase").
Chris@10: 
Chris@0:     m_phaseData = new Phase[inputSpacing];
Chris@0: 
Chris@0:     for (int phase = 0; phase < inputSpacing; ++phase) {
Chris@0: 
Chris@0: 	Phase p;
Chris@0: 
Chris@0: 	p.nextPhase = phase - outputSpacing;
Chris@0: 	while (p.nextPhase < 0) p.nextPhase += inputSpacing;
Chris@0: 	p.nextPhase %= inputSpacing;
Chris@0: 	
Chris@4: 	p.drop = int(ceil(std::max(0.0, double(outputSpacing - phase))
Chris@4: 			  / inputSpacing));
Chris@0: 
Chris@4: 	int filtZipLength = int(ceil(double(m_filterLength - phase)
Chris@4: 				     / inputSpacing));
Chris@10: 
Chris@0: 	for (int i = 0; i < filtZipLength; ++i) {
Chris@0: 	    p.filter.push_back(filter[i * inputSpacing + phase]);
Chris@0: 	}
Chris@0: 
Chris@0: 	m_phaseData[phase] = p;
Chris@0:     }
Chris@0: 
Chris@0:     // The May implementation of this uses a pull model -- we ask the
Chris@0:     // resampler for a certain number of output samples, and it asks
Chris@0:     // its source stream for as many as it needs to calculate
Chris@0:     // those. This means (among other things) that the source stream
Chris@0:     // can be asked for enough samples up-front to fill the buffer
Chris@0:     // before the first output sample is generated.
Chris@0:     // 
Chris@0:     // In this implementation we're using a push model in which a
Chris@0:     // certain number of source samples is provided and we're asked
Chris@0:     // for as many output samples as that makes available. But we
Chris@0:     // can't return any samples from the beginning until half the
Chris@0:     // filter length has been provided as input. This means we must
Chris@0:     // either return a very variable number of samples (none at all
Chris@0:     // until the filter fills, then half the filter length at once) or
Chris@0:     // else have a lengthy declared latency on the output. We do the
Chris@0:     // latter. (What do other implementations do?)
Chris@11:     //
Chris@10:     // We want to make sure the first "real" sample will eventually be
Chris@10:     // aligned with the centre sample in the filter (it's tidier, and
Chris@10:     // easier to do diagnostic calculations that way). So we need to
Chris@10:     // pick the initial phase and buffer fill accordingly.
Chris@10:     // 
Chris@10:     // Example: if the inputSpacing is 2, outputSpacing is 3, and
Chris@10:     // filter length is 7,
Chris@10:     // 
Chris@10:     //    x     x     x     x     a     b     c ... input samples
Chris@10:     // 0  1  2  3  4  5  6  7  8  9 10 11 12 13 ... 
Chris@10:     //          i        j        k        l    ... output samples
Chris@10:     // [--------|--------] <- filter with centre mark
Chris@10:     //
Chris@10:     // Let h be the index of the centre mark, here 3 (generally
Chris@10:     // int(filterLength/2) for odd-length filters).
Chris@10:     //
Chris@10:     // The smallest n such that h + n * outputSpacing > filterLength
Chris@10:     // is 2 (that is, ceil((filterLength - h) / outputSpacing)), and
Chris@10:     // (h + 2 * outputSpacing) % inputSpacing == 1, so the initial
Chris@10:     // phase is 1.
Chris@10:     //
Chris@10:     // To achieve our n, we need to pre-fill the "virtual" buffer with
Chris@10:     // 4 zero samples: the x's above. This is int((h + n *
Chris@10:     // outputSpacing) / inputSpacing). It's the phase that makes this
Chris@10:     // buffer get dealt with in such a way as to give us an effective
Chris@10:     // index for sample a of 9 rather than 8 or 10 or whatever.
Chris@10:     //
Chris@10:     // This gives us output latency of 2 (== n), i.e. output samples i
Chris@10:     // and j will appear before the one in which input sample a is at
Chris@10:     // the centre of the filter.
Chris@10: 
Chris@10:     int h = int(m_filterLength / 2);
Chris@10:     int n = ceil(double(m_filterLength - h) / outputSpacing);
Chris@4:     
Chris@10:     m_phase = (h + n * outputSpacing) % inputSpacing;
Chris@10: 
Chris@10:     int fill = (h + n * outputSpacing) / inputSpacing;
Chris@10:     
Chris@10:     m_latency = n;
Chris@10: 
Chris@10:     m_buffer = vector<double>(fill, 0);
Chris@8:     m_bufferOrigin = 0;
Chris@4: 
Chris@4: #ifdef DEBUG_RESAMPLER
Chris@4:     std::cerr << "initial phase " << m_phase << " (as " << (m_filterLength/2) << " % " << inputSpacing << ")"
Chris@4: 	      << ", latency " << m_latency << std::endl;
Chris@4: #endif
Chris@0: }
Chris@0: 
Chris@0: double
Chris@4: Resampler::reconstructOne()
Chris@0: {
Chris@0:     Phase &pd = m_phaseData[m_phase];
Chris@4:     double v = 0.0;
Chris@0:     int n = pd.filter.size();
Chris@10: 
Chris@11:     assert(n + m_bufferOrigin <= (int)m_buffer.size());
Chris@10: 
Chris@8:     const double *const __restrict__ buf = m_buffer.data() + m_bufferOrigin;
Chris@8:     const double *const __restrict__ filt = pd.filter.data();
Chris@10: 
Chris@10: //    std::cerr << "phase = " << m_phase << ", drop = " << pd.drop << ", buffer for reconstruction starts...";
Chris@10: //    for (int i = 0; i < 20; ++i) {
Chris@10: //        if (i % 5 == 0) std::cerr << "\n" << i << " ";
Chris@10: //        std::cerr << buf[i] << " ";
Chris@10: //    }
Chris@10: //    std::cerr << std::endl;
Chris@10: 
Chris@0:     for (int i = 0; i < n; ++i) {
Chris@8: 	// NB gcc can only vectorize this with -ffast-math
Chris@8: 	v += buf[i] * filt[i];
Chris@0:     }
Chris@16: 
Chris@8:     m_bufferOrigin += pd.drop;
Chris@4:     m_phase = pd.nextPhase;
Chris@0:     return v;
Chris@0: }
Chris@0: 
Chris@0: int
Chris@4: Resampler::process(const double *src, double *dst, int n)
Chris@0: {
Chris@4:     for (int i = 0; i < n; ++i) {
Chris@4: 	m_buffer.push_back(src[i]);
Chris@0:     }
Chris@0: 
Chris@4:     int maxout = int(ceil(double(n) * m_targetRate / m_sourceRate));
Chris@4:     int outidx = 0;
Chris@2: 
Chris@4: #ifdef DEBUG_RESAMPLER
Chris@4:     std::cerr << "process: buf siz " << m_buffer.size() << " filt siz for phase " << m_phase << " " << m_phaseData[m_phase].filter.size() << std::endl;
Chris@4: #endif
Chris@4: 
Chris@5:     double scaleFactor = 1.0;
Chris@5:     if (m_targetRate < m_sourceRate) {
Chris@5: 	scaleFactor = double(m_targetRate) / double(m_sourceRate);
Chris@5:     }
Chris@5: 
Chris@4:     while (outidx < maxout &&
Chris@8: 	   m_buffer.size() >= m_phaseData[m_phase].filter.size() + m_bufferOrigin) {
Chris@5: 	dst[outidx] = scaleFactor * reconstructOne();
Chris@4: 	outidx++;
Chris@2:     }
Chris@8: 
Chris@8:     m_buffer = vector<double>(m_buffer.begin() + m_bufferOrigin, m_buffer.end());
Chris@8:     m_bufferOrigin = 0;
Chris@4:     
Chris@4:     return outidx;
Chris@0: }
Chris@4:     
Chris@1: std::vector<double>
Chris@1: Resampler::resample(int sourceRate, int targetRate, const double *data, int n)
Chris@1: {
Chris@1:     Resampler r(sourceRate, targetRate);
Chris@1: 
Chris@1:     int latency = r.getLatency();
Chris@1: 
Chris@6:     // latency is the output latency. We need to provide enough
Chris@6:     // padding input samples at the end of input to guarantee at
Chris@6:     // *least* the latency's worth of output samples. that is,
Chris@6: 
Chris@11:     int inputPad = int(ceil((double(latency) * sourceRate) / targetRate));
Chris@6: 
Chris@6:     // that means we are providing this much input in total:
Chris@6: 
Chris@6:     int n1 = n + inputPad;
Chris@6: 
Chris@6:     // and obtaining this much output in total:
Chris@6: 
Chris@11:     int m1 = int(ceil((double(n1) * targetRate) / sourceRate));
Chris@6: 
Chris@6:     // in order to return this much output to the user:
Chris@6: 
Chris@11:     int m = int(ceil((double(n) * targetRate) / sourceRate));
Chris@6:     
Chris@11: //    std::cerr << "n = " << n << ", sourceRate = " << sourceRate << ", targetRate = " << targetRate << ", m = " << m << ", latency = " << latency << ", inputPad = " << inputPad << ", m1 = " << m1 << ", n1 = " << n1 << ", n1 - n = " << n1 - n << std::endl;
Chris@1: 
Chris@1:     vector<double> pad(n1 - n, 0.0);
Chris@6:     vector<double> out(m1 + 1, 0.0);
Chris@1: 
Chris@1:     int got = r.process(data, out.data(), n);
Chris@1:     got += r.process(pad.data(), out.data() + got, pad.size());
Chris@1: 
Chris@4: #ifdef DEBUG_RESAMPLER
Chris@4:     std::cerr << "resample: " << n << " in, " << got << " out" << std::endl;
Chris@10:     std::cerr << "first 10 in:" << std::endl;
Chris@10:     for (int i = 0; i < 10; ++i) {
Chris@10:         std::cerr << data[i] << " ";
Chris@10:         if (i == 5) std::cerr << std::endl;
Chris@4:     }
Chris@10:     std::cerr << std::endl;
Chris@4: #endif
Chris@4: 
Chris@6:     int toReturn = got - latency;
Chris@6:     if (toReturn > m) toReturn = m;
Chris@6: 
Chris@10:     vector<double> sliced(out.begin() + latency, 
Chris@6: 			  out.begin() + latency + toReturn);
Chris@10: 
Chris@10: #ifdef DEBUG_RESAMPLER
Chris@10:     std::cerr << "all out (after latency compensation), length " << sliced.size() << ":";
Chris@10:     for (int i = 0; i < sliced.size(); ++i) {
Chris@10: 	if (i % 5 == 0) std::cerr << std::endl << i << "... ";
Chris@10: 	std::cerr << sliced[i] << " ";
Chris@10:     }
Chris@10:     std::cerr << std::endl;
Chris@10: #endif
Chris@10: 
Chris@10:     return sliced;
Chris@1: }
Chris@1: