[microNPU][ETHOSU] Add fixed point for matmul

[Aleksei-grovety]

Added support for calculation matmul with 16 bits fixed point. Added legalization of non-quantized dense operation with quantization by fixed point multiplication.
cc @lhutton1, @ekalda, @leandron

[lhutton1]

Thanks for the PR @Aleksei-grovety, I had a couple of questions about the implementation

[lhutton1]

Thanks for the responses @Aleksei-grovety. My main concern is that we need to expose an option to the user, rather than encapsulating this information as part of a relay operator. Are you able to share any more context about where this type of operation comes from?

[Aleksei-grovety]

Are you able to share any more context about where this type of operation comes from?

One of the earliest uses of fixed-point arithmetic was in digital signal processing (DSP) applications, where it was used to perform high-speed and high-precision calculations on audio and video signals. Fixed-point arithmetic was well-suited for these applications because it could provide the necessary precision and speed while minimizing the resources required to perform the calculations.
Here is article Introduction to Fixed Point Number Representation.

[lhutton1]

Thanks for the cleanup @Aleksei-grovety. I think overall I'm still a bit unsure why fixed_point_fraction_size needs to be exposed from the microNPU compiler attrs, rather than the data being encoded in the Relay graph. For example, is it possible to insert relay.fixed_point_multiply operators on the inputs/outputs of the graph in between the matmul? Apologies if I'm still missing something here

[Aleksei-grovety]

For example, is it possible to insert relay.fixed_point_multiply operators on the inputs/outputs of the graph in between the matmul

Yes, it's possible, I've updated the code.

[lhutton1]

Thanks @Aleksei-grovety, LGTM!

[lhutton1]

Thanks @Aleksei-grovety!