VariantXYZ

Ah, one other quick thing: for fp16, it looks like the dummy fmla file for pragma testing should specify fullfp16 for fp16 arithmetic support (fp16 on its own apparently only...

> Did you find an example that is not working? Just building with the latest nightly with llvm/clang 15 for cortex-a55 seems to trigger this warning. I haven't tried with...

Ah, sorry for misleading you. No I haven't actually gotten that far, I was actually getting stuck trying to figure out how to properly dispatch it. It's actually a bit...

(For now I'll just force it to call into an external C function for testing)

I'm going with the first option for testing, but I think the main issue is actually going to be with offsets, as I don't believe Rust offers any FFI guarantees...

@kali , I ended up just going with a bit of a lazy solution for now. https://github.com/sonos/tract/compare/main...VariantXYZ:move_dot_prod I did verify the output file will correctly generate half-width floating point instructions...

Some good news though: I commented the C dot product function out and the allocations, and instead just substituted the multiply and add with a quick madd_f16 function written in...

> My intuition says that the sizeof() thing is the bug. It was, I think I'm in a holiday mindset already. > I don't foresee any big issue in replacing...

@kali , Happy New Year! I'm back but honestly I don't know what the next action item is here. The PR 'works' for the platforms I was testing, but seems...

Hi @kali , I started digging back into this, and found a few interesting things: * Swapping values_offsets to koffsets and ioffsets separately is very useful (and avoids allocations in...