Subtract Streams Support

[SpontaneousDuck]

Subtract Streams Support

[maltanar]

A quick note if you are using the finn compiler to produce the finn-hlslib description: we have a ConvertSubToAdd transformation that converts subtraction ops to add ops, which would handle this without any extra implementation: https://github.com/Xilinx/finn-base/blob/dev/src/finn/transformation/general.py#L233

(If you are manually putting together finn-hlslib layers then what I wrote above does not apply)

[SpontaneousDuck]

Thanks for the response! Unfortunately, I am subtracting two unique streams. I assume I would need a block to invert the sign of one fo the streams then? Which is the better solution?

[giuliogamba]

Hi, changing the API of AddStreams could have implications in the finn compiler, so I would advise against it.
I would suggest trying using the Thresholding_Batch function (here) with the activation function std::negate, that should allow you to perform sign inversion on a per-channel basis on one of the inputs allowing you to do a subtraction in 2 steps.

[SpontaneousDuck]

Sounds like a plan! Is there a method for doing that in FINN? A matching CustomOp?

[maltanar]

This approach best discussed on the finn gitter channel as it's not an hlslib issue per se. Also feel free to open an issue on the finn compiler repo with a minimal example and we can try to help formulate a way of making this work. But in summary:

If you express the subtraction of streams explicitly as a negate-and-add I believe this should be doable with existing transformations and CustomOps, namely InferChannelwiseLinearLayer to convert standalone Mul by -1 into the CustomOp ChannelWiseOp_Batch which is implemented with the Thresholding_Batch function under the hood (https://github.com/Xilinx/finn/blob/master/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py). Followed by appropriate transformations to infer the AddStreams.

[SpontaneousDuck]

Thanks for the help!

[SpontaneousDuck]

Thanks for the help!