FVP (CMSIS-NN Integ) : Add support to Quantized Arithmetic Operations: cortex_m::quantized_sub.out

[psiddh]

cortex_m::quantized_sub.out — Element-wise subtraction

[digantdesai]

@AdrianLundell do you want to take a stab at this now that qadd @psiddh landed?

[AdrianLundell]

Sure, I should be able to start this next week!

[AdrianLundell]

@digantdesai Do you have any plan for the testing framework? I could start trying to integrate with the general tester in backends/test for example?

[digantdesai]

Not yet, for now we have been overloading the arm baremetal e2e test scripts

General tester is good. We do need FVP support or other sim for other MCUs, in the future.

[AdrianLundell]

Okay sounds good, I will start with the tester part then before doing the sub op.

[AdrianLundell]

@digantdesai Hi, does it really make sense to add subtraction right now since CMSIS-NN does not have an optimized kernel for it? We could decompose it as (x + (-1)*y), but the question is how much quicker that is compared to the regular portable add, and we do not have a good way to benchmark that right now.

Since it's the matmul ops (conv + depthwise conv + linear) optimizations which are responsible for >80% of the uplifts in CMSIS-NN , it makes sense to me to chase that first. What do you think?