1 /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*- vi:set ts=8 sts=4 sw=4: */
5 Centre for Digital Music, Queen Mary, University of London.
6 This file by Chris Cannam.
8 This program is free software; you can redistribute it and/or
9 modify it under the terms of the GNU General Public License as
10 published by the Free Software Foundation; either version 2 of the
11 License, or (at your option) any later version. See the file
12 COPYING included with this distribution for more information.
15 #include "Resampler.h"
17 #include "maths/MathUtilities.h"
18 #include "base/KaiserWindow.h"
19 #include "base/SincWindow.h"
20 #include "thread/Thread.h"
32 //#define DEBUG_RESAMPLER 1
33 //#define DEBUG_RESAMPLER_VERBOSE 1
35 Resampler::Resampler(int sourceRate, int targetRate) :
36 m_sourceRate(sourceRate),
37 m_targetRate(targetRate)
39 initialise(100, 0.02);
42 Resampler::Resampler(int sourceRate, int targetRate,
43 double snr, double bandwidth) :
44 m_sourceRate(sourceRate),
45 m_targetRate(targetRate)
47 initialise(snr, bandwidth);
50 Resampler::~Resampler()
55 // peakToPole -> length -> beta -> window
56 static map<double, map<int, map<double, vector<double> > > >
63 Resampler::initialise(double snr, double bandwidth)
65 int higher = std::max(m_sourceRate, m_targetRate);
66 int lower = std::min(m_sourceRate, m_targetRate);
68 m_gcd = MathUtilities::gcd(lower, higher);
69 m_peakToPole = higher / m_gcd;
71 if (m_targetRate < m_sourceRate) {
72 // antialiasing filter, should be slightly below nyquist
73 m_peakToPole = m_peakToPole / (1.0 - bandwidth/2.0);
76 KaiserWindow::Parameters params =
77 KaiserWindow::parametersForBandwidth(snr, bandwidth, higher / m_gcd);
80 (params.length % 2 == 0 ? params.length + 1 : params.length);
83 (params.length > 200001 ? 200001 : params.length);
85 m_filterLength = params.length;
87 vector<double> filter;
88 knownFilterMutex.lock();
90 if (knownFilters[m_peakToPole][m_filterLength].find(params.beta) ==
91 knownFilters[m_peakToPole][m_filterLength].end()) {
93 KaiserWindow kw(params);
94 SincWindow sw(m_filterLength, m_peakToPole * 2);
96 filter = vector<double>(m_filterLength, 0.0);
97 for (int i = 0; i < m_filterLength; ++i) filter[i] = 1.0;
98 sw.cut(filter.data());
99 kw.cut(filter.data());
101 knownFilters[m_peakToPole][m_filterLength][params.beta] = filter;
104 filter = knownFilters[m_peakToPole][m_filterLength][params.beta];
105 knownFilterMutex.unlock();
107 int inputSpacing = m_targetRate / m_gcd;
108 int outputSpacing = m_sourceRate / m_gcd;
110 #ifdef DEBUG_RESAMPLER
111 cerr << "resample " << m_sourceRate << " -> " << m_targetRate
112 << ": inputSpacing " << inputSpacing << ", outputSpacing "
113 << outputSpacing << ": filter length " << m_filterLength
117 // Now we have a filter of (odd) length flen in which the lower
118 // sample rate corresponds to every n'th point and the higher rate
119 // to every m'th where n and m are higher and lower rates divided
120 // by their gcd respectively. So if x coordinates are on the same
121 // scale as our filter resolution, then source sample i is at i *
122 // (targetRate / gcd) and target sample j is at j * (sourceRate /
125 // To reconstruct a single target sample, we want a buffer (real
126 // or virtual) of flen values formed of source samples spaced at
127 // intervals of (targetRate / gcd), in our example case 3. This
128 // is initially formed with the first sample at the filter peak.
132 // and of course we have our filter
134 // f1 f2 f3 f4 f5 f6 f7 f8 f9
136 // We take the sum of products of non-zero values from this buffer
137 // with corresponding values in the filter
141 // Then we drop (sourceRate / gcd) values, in our example case 4,
142 // from the start of the buffer and fill until it has flen values
147 // repeat to reconstruct the next target sample
149 // a * f1 + b * f4 + c * f7
153 // Above I said the buffer could be "real or virtual" -- ours is
154 // virtual. We don't actually store all the zero spacing values,
155 // except for padding at the start; normally we store only the
156 // values that actually came from the source stream, along with a
157 // phase value that tells us how many virtual zeroes there are at
158 // the start of the virtual buffer. So the two examples above are
160 // 0 a b [ with phase 1 ]
161 // a b c [ with phase 0 ]
163 // Having thus broken down the buffer so that only the elements we
164 // need to multiply are present, we can also unzip the filter into
165 // every-nth-element subsets at each phase, allowing us to do the
166 // filter multiplication as a simply vector multiply. That is, rather
169 // f1 f2 f3 f4 f5 f6 f7 f8 f9
171 // we store separately
177 // Each time we complete a multiply-and-sum, we need to work out
178 // how many (real) samples to drop from the start of our buffer,
179 // and how many to add at the end of it for the next multiply. We
180 // know we want to drop enough real samples to move along by one
181 // computed output sample, which is our outputSpacing number of
182 // virtual buffer samples. Depending on the relationship between
183 // input and output spacings, this may mean dropping several real
184 // samples, one real sample, or none at all (and simply moving to
185 // a different "phase").
187 m_phaseData = new Phase[inputSpacing];
189 for (int phase = 0; phase < inputSpacing; ++phase) {
193 p.nextPhase = phase - outputSpacing;
194 while (p.nextPhase < 0) p.nextPhase += inputSpacing;
195 p.nextPhase %= inputSpacing;
197 p.drop = int(ceil(std::max(0.0, double(outputSpacing - phase))
200 int filtZipLength = int(ceil(double(m_filterLength - phase)
203 for (int i = 0; i < filtZipLength; ++i) {
204 p.filter.push_back(filter[i * inputSpacing + phase]);
207 m_phaseData[phase] = p;
210 #ifdef DEBUG_RESAMPLER
213 for (int i = 0; i < inputSpacing; ++i) {
214 cerr << "phase = " << cp << ", drop = " << m_phaseData[cp].drop
215 << ", filter length = " << m_phaseData[cp].filter.size()
216 << ", next phase = " << m_phaseData[cp].nextPhase << endl;
217 totDrop += m_phaseData[cp].drop;
218 cp = m_phaseData[cp].nextPhase;
220 cerr << "total drop = " << totDrop << endl;
223 // The May implementation of this uses a pull model -- we ask the
224 // resampler for a certain number of output samples, and it asks
225 // its source stream for as many as it needs to calculate
226 // those. This means (among other things) that the source stream
227 // can be asked for enough samples up-front to fill the buffer
228 // before the first output sample is generated.
230 // In this implementation we're using a push model in which a
231 // certain number of source samples is provided and we're asked
232 // for as many output samples as that makes available. But we
233 // can't return any samples from the beginning until half the
234 // filter length has been provided as input. This means we must
235 // either return a very variable number of samples (none at all
236 // until the filter fills, then half the filter length at once) or
237 // else have a lengthy declared latency on the output. We do the
238 // latter. (What do other implementations do?)
240 // We want to make sure the first "real" sample will eventually be
241 // aligned with the centre sample in the filter (it's tidier, and
242 // easier to do diagnostic calculations that way). So we need to
243 // pick the initial phase and buffer fill accordingly.
245 // Example: if the inputSpacing is 2, outputSpacing is 3, and
246 // filter length is 7,
248 // x x x x a b c ... input samples
249 // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 ...
250 // i j k l ... output samples
251 // [--------|--------] <- filter with centre mark
253 // Let h be the index of the centre mark, here 3 (generally
254 // int(filterLength/2) for odd-length filters).
256 // The smallest n such that h + n * outputSpacing > filterLength
257 // is 2 (that is, ceil((filterLength - h) / outputSpacing)), and
258 // (h + 2 * outputSpacing) % inputSpacing == 1, so the initial
261 // To achieve our n, we need to pre-fill the "virtual" buffer with
262 // 4 zero samples: the x's above. This is int((h + n *
263 // outputSpacing) / inputSpacing). It's the phase that makes this
264 // buffer get dealt with in such a way as to give us an effective
265 // index for sample a of 9 rather than 8 or 10 or whatever.
267 // This gives us output latency of 2 (== n), i.e. output samples i
268 // and j will appear before the one in which input sample a is at
269 // the centre of the filter.
271 int h = int(m_filterLength / 2);
272 int n = ceil(double(m_filterLength - h) / outputSpacing);
274 m_phase = (h + n * outputSpacing) % inputSpacing;
276 int fill = (h + n * outputSpacing) / inputSpacing;
280 m_buffer = vector<double>(fill, 0);
283 #ifdef DEBUG_RESAMPLER
284 cerr << "initial phase " << m_phase << " (as " << (m_filterLength/2) << " % " << inputSpacing << ")"
285 << ", latency " << m_latency << endl;
290 Resampler::reconstructOne()
292 Phase &pd = m_phaseData[m_phase];
294 int n = pd.filter.size();
296 assert(n + m_bufferOrigin <= (int)m_buffer.size());
298 const double *const __restrict__ buf = m_buffer.data() + m_bufferOrigin;
299 const double *const __restrict__ filt = pd.filter.data();
301 for (int i = 0; i < n; ++i) {
302 // NB gcc can only vectorize this with -ffast-math
303 v += buf[i] * filt[i];
306 m_bufferOrigin += pd.drop;
307 m_phase = pd.nextPhase;
312 Resampler::process(const double *src, double *dst, int n)
314 for (int i = 0; i < n; ++i) {
315 m_buffer.push_back(src[i]);
318 int maxout = int(ceil(double(n) * m_targetRate / m_sourceRate));
321 #ifdef DEBUG_RESAMPLER
322 cerr << "process: buf siz " << m_buffer.size() << " filt siz for phase " << m_phase << " " << m_phaseData[m_phase].filter.size() << endl;
325 double scaleFactor = (double(m_targetRate) / m_gcd) / m_peakToPole;
327 while (outidx < maxout &&
328 m_buffer.size() >= m_phaseData[m_phase].filter.size() + m_bufferOrigin) {
329 dst[outidx] = scaleFactor * reconstructOne();
333 m_buffer = vector<double>(m_buffer.begin() + m_bufferOrigin, m_buffer.end());
340 Resampler::process(const double *src, int n)
342 int maxout = int(ceil(double(n) * m_targetRate / m_sourceRate));
343 vector<double> out(maxout, 0.0);
344 int got = process(src, out.data(), n);
345 assert(got <= maxout);
346 if (got < maxout) out.resize(got);
351 Resampler::resample(int sourceRate, int targetRate, const double *data, int n)
353 Resampler r(sourceRate, targetRate);
355 int latency = r.getLatency();
357 // latency is the output latency. We need to provide enough
358 // padding input samples at the end of input to guarantee at
359 // *least* the latency's worth of output samples. that is,
361 int inputPad = int(ceil((double(latency) * sourceRate) / targetRate));
363 // that means we are providing this much input in total:
365 int n1 = n + inputPad;
367 // and obtaining this much output in total:
369 int m1 = int(ceil((double(n1) * targetRate) / sourceRate));
371 // in order to return this much output to the user:
373 int m = int(ceil((double(n) * targetRate) / sourceRate));
375 #ifdef DEBUG_RESAMPLER
376 cerr << "n = " << n << ", sourceRate = " << sourceRate << ", targetRate = " << targetRate << ", m = " << m << ", latency = " << latency << ", inputPad = " << inputPad << ", m1 = " << m1 << ", n1 = " << n1 << ", n1 - n = " << n1 - n << endl;
379 vector<double> pad(n1 - n, 0.0);
380 vector<double> out(m1 + 1, 0.0);
382 int gotData = r.process(data, out.data(), n);
383 int gotPad = r.process(pad.data(), out.data() + gotData, pad.size());
384 int got = gotData + gotPad;
386 #ifdef DEBUG_RESAMPLER
387 cerr << "resample: " << n << " in, " << pad.size() << " padding, " << got << " out (" << gotData << " data, " << gotPad << " padding, latency = " << latency << ")" << endl;
389 #ifdef DEBUG_RESAMPLER_VERBOSE
391 cerr << "first " << printN << " in:" << endl;
392 for (int i = 0; i < printN && i < n; ++i) {
393 if (i % 5 == 0) cerr << endl << i << "... ";
394 cerr << data[i] << " ";
399 int toReturn = got - latency;
400 if (toReturn > m) toReturn = m;
402 vector<double> sliced(out.begin() + latency,
403 out.begin() + latency + toReturn);
405 #ifdef DEBUG_RESAMPLER_VERBOSE
406 cerr << "first " << printN << " out (after latency compensation), length " << sliced.size() << ":";
407 for (int i = 0; i < printN && i < sliced.size(); ++i) {
408 if (i % 5 == 0) cerr << endl << i << "... ";
409 cerr << sliced[i] << " ";