c@362: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
c@362: 
c@362: #include "Resampler.h"
c@362: 
c@362: #include "qm-dsp/maths/MathUtilities.h"
c@362: #include "qm-dsp/base/KaiserWindow.h"
c@362: #include "qm-dsp/base/SincWindow.h"
c@362: 
c@362: #include <iostream>
c@363: #include <vector>
c@370: #include <map>
c@363: 
c@363: using std::vector;
c@370: using std::map;
c@362: 
c@366: //#define DEBUG_RESAMPLER 1
c@366: 
c@362: Resampler::Resampler(int sourceRate, int targetRate) :
c@362:     m_sourceRate(sourceRate),
c@362:     m_targetRate(targetRate)
c@362: {
c@362:     initialise();
c@362: }
c@362: 
c@362: Resampler::~Resampler()
c@362: {
c@362:     delete[] m_phaseData;
c@362: }
c@362: 
c@362: void
c@362: Resampler::initialise()
c@362: {
c@362:     int higher = std::max(m_sourceRate, m_targetRate);
c@362:     int lower = std::min(m_sourceRate, m_targetRate);
c@362: 
c@362:     m_gcd = MathUtilities::gcd(lower, higher);
c@362: 
c@362:     int peakToPole = higher / m_gcd;
c@362: 
c@362:     KaiserWindow::Parameters params =
c@362: 	KaiserWindow::parametersForBandwidth(100, 0.02, peakToPole);
c@362: 
c@362:     params.length =
c@362: 	(params.length % 2 == 0 ? params.length + 1 : params.length);
c@362:     
c@362:     m_filterLength = params.length;
c@370: 
c@362:     KaiserWindow kw(params);
c@362:     SincWindow sw(m_filterLength, peakToPole * 2);
c@362: 
c@370:     vector<double> filter(m_filterLength, 0.0);
c@362:     for (int i = 0; i < m_filterLength; ++i) filter[i] = 1.0;
c@370:     sw.cut(filter.data());
c@370:     kw.cut(filter.data());
c@362: 
c@362:     int inputSpacing = m_targetRate / m_gcd;
c@362:     int outputSpacing = m_sourceRate / m_gcd;
c@362: 
c@366: #ifdef DEBUG_RESAMPLER
c@366:     std::cerr << "resample " << m_sourceRate << " -> " << m_targetRate
c@366: 	      << ": inputSpacing " << inputSpacing << ", outputSpacing "
c@366: 	      << outputSpacing << ": filter length " << m_filterLength
c@366: 	      << std::endl;
c@366: #endif
c@362: 
c@362:     m_phaseData = new Phase[inputSpacing];
c@362: 
c@362:     for (int phase = 0; phase < inputSpacing; ++phase) {
c@362: 
c@362: 	Phase p;
c@362: 
c@362: 	p.nextPhase = phase - outputSpacing;
c@362: 	while (p.nextPhase < 0) p.nextPhase += inputSpacing;
c@362: 	p.nextPhase %= inputSpacing;
c@362: 	
c@366: 	p.drop = int(ceil(std::max(0.0, double(outputSpacing - phase))
c@366: 			  / inputSpacing));
c@362: 
c@366: 	int filtZipLength = int(ceil(double(m_filterLength - phase)
c@366: 				     / inputSpacing));
c@362: 	for (int i = 0; i < filtZipLength; ++i) {
c@362: 	    p.filter.push_back(filter[i * inputSpacing + phase]);
c@362: 	}
c@362: 
c@362: 	m_phaseData[phase] = p;
c@362:     }
c@362: 
c@362:     // The May implementation of this uses a pull model -- we ask the
c@362:     // resampler for a certain number of output samples, and it asks
c@362:     // its source stream for as many as it needs to calculate
c@362:     // those. This means (among other things) that the source stream
c@362:     // can be asked for enough samples up-front to fill the buffer
c@362:     // before the first output sample is generated.
c@362:     // 
c@362:     // In this implementation we're using a push model in which a
c@362:     // certain number of source samples is provided and we're asked
c@362:     // for as many output samples as that makes available. But we
c@362:     // can't return any samples from the beginning until half the
c@362:     // filter length has been provided as input. This means we must
c@362:     // either return a very variable number of samples (none at all
c@362:     // until the filter fills, then half the filter length at once) or
c@362:     // else have a lengthy declared latency on the output. We do the
c@362:     // latter. (What do other implementations do?)
c@362: 
c@366:     m_phase = (m_filterLength/2) % inputSpacing;
c@366:     
c@366:     m_buffer = vector<double>(m_phaseData[0].filter.size(), 0);
c@370:     m_bufferOrigin = 0;
c@366: 
c@366:     m_latency =
c@366: 	((m_buffer.size() * inputSpacing) - (m_filterLength/2)) / outputSpacing
c@366: 	+ m_phase;
c@366: 
c@366: #ifdef DEBUG_RESAMPLER
c@366:     std::cerr << "initial phase " << m_phase << " (as " << (m_filterLength/2) << " % " << inputSpacing << ")"
c@366: 	      << ", latency " << m_latency << std::endl;
c@366: #endif
c@362: }
c@362: 
c@362: double
c@366: Resampler::reconstructOne()
c@362: {
c@362:     Phase &pd = m_phaseData[m_phase];
c@366:     double v = 0.0;
c@362:     int n = pd.filter.size();
c@370:     const double *const __restrict__ buf = m_buffer.data() + m_bufferOrigin;
c@370:     const double *const __restrict__ filt = pd.filter.data();
c@362:     for (int i = 0; i < n; ++i) {
c@370: 	// NB gcc can only vectorize this with -ffast-math
c@370: 	v += buf[i] * filt[i];
c@362:     }
c@370:     m_bufferOrigin += pd.drop;
c@366:     m_phase = pd.nextPhase;
c@362:     return v;
c@362: }
c@362: 
c@362: int
c@366: Resampler::process(const double *src, double *dst, int n)
c@362: {
c@366:     for (int i = 0; i < n; ++i) {
c@366: 	m_buffer.push_back(src[i]);
c@362:     }
c@362: 
c@366:     int maxout = int(ceil(double(n) * m_targetRate / m_sourceRate));
c@366:     int outidx = 0;
c@364: 
c@366: #ifdef DEBUG_RESAMPLER
c@366:     std::cerr << "process: buf siz " << m_buffer.size() << " filt siz for phase " << m_phase << " " << m_phaseData[m_phase].filter.size() << std::endl;
c@366: #endif
c@366: 
c@367:     double scaleFactor = 1.0;
c@367:     if (m_targetRate < m_sourceRate) {
c@367: 	scaleFactor = double(m_targetRate) / double(m_sourceRate);
c@367:     }
c@367: 
c@366:     while (outidx < maxout &&
c@370: 	   m_buffer.size() >= m_phaseData[m_phase].filter.size() + m_bufferOrigin) {
c@367: 	dst[outidx] = scaleFactor * reconstructOne();
c@366: 	outidx++;
c@364:     }
c@370: 
c@370:     m_buffer = vector<double>(m_buffer.begin() + m_bufferOrigin, m_buffer.end());
c@370:     m_bufferOrigin = 0;
c@366:     
c@366:     return outidx;
c@362: }
c@366:     
c@363: std::vector<double>
c@363: Resampler::resample(int sourceRate, int targetRate, const double *data, int n)
c@363: {
c@363:     Resampler r(sourceRate, targetRate);
c@363: 
c@363:     int latency = r.getLatency();
c@363: 
c@368:     // latency is the output latency. We need to provide enough
c@368:     // padding input samples at the end of input to guarantee at
c@368:     // *least* the latency's worth of output samples. that is,
c@368: 
c@368:     int inputPad = int(ceil(double(latency * sourceRate) / targetRate));
c@368: 
c@368:     // that means we are providing this much input in total:
c@368: 
c@368:     int n1 = n + inputPad;
c@368: 
c@368:     // and obtaining this much output in total:
c@368: 
c@368:     int m1 = int(ceil(double(n1 * targetRate) / sourceRate));
c@368: 
c@368:     // in order to return this much output to the user:
c@368: 
c@366:     int m = int(ceil(double(n * targetRate) / sourceRate));
c@368:     
c@370: //    std::cerr << "n = " << n << ", sourceRate = " << sourceRate << ", targetRate = " << targetRate << ", m = " << m << ", latency = " << latency << ", m1 = " << m1 << ", n1 = " << n1 << ", n1 - n = " << n1 - n << std::endl;
c@363: 
c@363:     vector<double> pad(n1 - n, 0.0);
c@368:     vector<double> out(m1 + 1, 0.0);
c@363: 
c@363:     int got = r.process(data, out.data(), n);
c@363:     got += r.process(pad.data(), out.data() + got, pad.size());
c@363: 
c@366: #ifdef DEBUG_RESAMPLER
c@366:     std::cerr << "resample: " << n << " in, " << got << " out" << std::endl;
c@366:     for (int i = 0; i < got; ++i) {
c@366: 	if (i % 5 == 0) std::cout << std::endl << i << "... ";
c@366: 	std::cout << (float) out[i] << " ";
c@366:     }
c@366:     std::cout << std::endl;
c@366: #endif
c@366: 
c@368:     int toReturn = got - latency;
c@368:     if (toReturn > m) toReturn = m;
c@368: 
c@368:     return vector<double>(out.begin() + latency, 
c@368: 			  out.begin() + latency + toReturn);
c@363: }
c@363: