Clarification of MultiTransformBlock usage

[telegraphic]

Hi all, I'm experimenting with a pipeline that joins two input streams (blocks) together, and was hoping for some clarification on how to implement it. Firstly, what should happens on the on_sequence event if there are N input streams, and M output streams? Is it as simple as parsing the N input headers, and returning a list of M output headers? Similarly, the on_data method should return a list with M entries, each specifying the number of frames per stream?

Here is my attempt, using numpy to concantenate two inputs. Is there a suggested way to implement this on the GPU, perhaps using bf.map?

class StitchBlock(bf.pipeline.MultiTransformBlock):
    """ Concatenate two rings along axis """
    def __init__(self, irings, axis=-1, *args, **kwargs):
        super(StitchBlock, self).__init__(irings, *args, **kwargs)
        self.axis = axis

    def on_sequence(self, iseqs):
        ohdr = deepcopy(iseqs[0].header)
        if isinstance(self.axis, str):
            self.axis = ohdr["_tensor"]["labels"].index(self.axis)
        ohdr["_tensor"]["shape"][self.axis] *= 2
        return [ohdr,]

    def on_data(self, ispans, ospans):
        out_nframe = ispans[0].nframe
        d0 = ispans[0].data
        d1 = ispans[1].data
        d  = np.concatenate((d0, d1), axis=self.axis)

        odata = ospans[0].data
        odata[...] = d
        return [out_nframe,]

Cheers,
Danny

[jaycedowell]

Out of curiosity, what is the use case you are thinking about for this block?

[telegraphic]

The plan is to recombine F-engine data that arrives on two different UDP ports. So input tensors are ['time', 'ant', 'channel', 'pol'], and concatenating on channel axis.

Our capture code is actually PSRDADA-based; I'm using the bifrost bindings to attach to existing ring buffer capture code. It's working! (I'll add some details on how to do this to the bifrost documentation at some stage)

[jaycedowell]

That's cool. Were you also able to get this running on the GPU with bf.map() or are you still on the CPU with numpy.concatenate()?

[telegraphic]

It's working with numpy.concatenate; haven't made a gpu version (yet)