An (almost*) header only downsampling library in plain C++.
It can downsample multiplexed data of any type (that implements
operator*) after filtering it with a FIR filter of your choice.
After constructing the downsampler, no further allocations and almost no copies of your data are made. Most compilers should be smart enough to use SSE/AVX intrinsics (tested with icc 19, msvc v19.22, GCC 9.2, clang 9.0).
int data1[] = {
1, 13, 24,
1, 13, 26,
2, 14, 28,
2, 14, 30,
3, 15, 32,
3, 15, 34
};
int coef1[] = {1, 1, 0};
auto downrate = 2u, nchan = 3u, ncoef = 3u;
auto r(firdn::makeDownsampler<int>(downrate, nchan, coef1, ncoef));
std::vector<int> out;
// Downsample first half, one sample is left over
r(data, nsample / downrate, out);
// Downsample second half, continueing with the left over sample and appending
// to the `out` vector
r(data + nsample * nchan / downrate, nsample / downrate, out2);No C++ code to construct a lowpass FIR filter is included, but the included Python script can create and export a filter of your choice. You can then compile the filter coefficients into an object file so you can use them as any other array.
Downsampling is only possible in integer ratios, e.g. 100->50Hz, but not 100->75Hz. PRs to implement upsampling are welcome.
Currently, the input buffer needs to have at least as many samples as filter has coefficients. This requirement will be removed in a later release.
This code is based on the upfirdn polyphase FIR resampling package, (c) 2009, Motorola, Inc.
Everything else is licensed under the LGPL 3.0.
In practice, this means that if you use the library you have to mention it (e.g. in the about dialog or somewhere buried in a list of projects you're using) and you can sell your program and keep its source code private.
Also, you have provide only the firdn.hpp header file to anyone you
distributed your compiled program to if they're asking politely for it.
For more information, see the eigen licensing FAQ.