Polyphase _resampler or channelize_poly: how to choose (change) the prototype filter

[neon5d]

I am looking at the source code of polyphase _resampler or channelize_poly and their documents. However, I can't figure out how to determine the filter coefficient from the source code. Also, how i can change the filter coefficient. There is only option for the number of the filter taps.
Can you help to clarify them?

[tbensonatl]

Hi @neon5d - the polyphase resampler and channelizer both take the filter as an operator or tensor, so the user provides the filter coefficients. There is no default filter as there is in scipy and similar frameworks. This is partly because MatX does not currently have much support for filter design -- we have some windowing functions, but no firwin(), for example. If you would like to see how scipy generates its default window for resample_poly, then you can check here:

https://github.com/scipy/scipy/blob/8c75ae75176236f233824e9a0483c26a69e6dfec/scipy/signal/_signaltools.py#L4022

One option if you already have window coefficients is to write those coefficients to a file, read those coefficients into your MatX application, and copy them to a filter tensor to pass to the resampler/channelizer.

[neon5d]

Thank you for your quick response.
From the matx example, channelize_poly_bench.cu, there is no code to initiate filter coefficient except creating a tensor for the filter. However, in the channelize_poly.cuh, shmem_h[t] =filter.operator()(t).
It shows some values (not garage).
Can you explain what the operator () operation and where i can find the corresponding code?

[tbensonatl]

For the channelize_poly_bench.cu example, the filter values are uninitialized, so you are seeing whatever values happened to exist in memory. We should fix that to initialize with random values -- we would not want to benchmark performance with all zeros or NaNs in the input, which is possible with uninitialized memory. If you check the unit tests, you can see a randomly initialized filter here:

MatX/test/test_vectors/generators/00_transforms.py

Line 186 in ccc805f

'filter_random': matx_common.randn_ndarray((filter_len,), dtype=dtype),

which is copied to the filter tensor here (and several other places):

MatX/test/00_transform/ChannelizePoly.cu

Line 167 in ccc805f

this->pb->NumpyToTensorView(f, "filter_random");

For correctness, we are testing against a Python implementation of a channelizer in the unit test code:

MatX/test/test_vectors/generators/00_transforms.py

Line 194 in ccc805f

def channelize(self) -> Dict[str, np.ndarray]:

The intent is that you will build the filter coefficients offline using a filter design package (MATLAB, scipy.signal, etc.) and then load those filter coefficients from file at runtime. There are currently no MatX examples that use real-world coefficients for the polyphase channelizer. For the polyphase resampler unit tests, we use scipy to build default coefficients (

MatX/test/test_vectors/generators/00_transforms.py

Line 145 in ccc805f

# Create a filter compatible with scipy's resample_poly

) and we test with those.

[neon5d]

Thank you for the clarification.
This is really helpful. Thank you.