aarongreig

Thanks for the feedback! > For example on Intel and Nvidia, the GLSL fma implementation seems to be compliant. > I manage to have them working on Intel by forcing...

Ah I didn't realize fma was getting inlined everywhere, I'm looking to see if there's some hint or attribute I could patch in to prevent that - I don't see...

This should go some way to prevent the unwanted inlining https://reviews.llvm.org/D132362

I've opened a revision to add generic implementations for builtins that failed full bruteforce on the devices I'm testing with https://reviews.llvm.org/D134887

Turns out we already have the mapping for [rsqrt = inversesqrt](https://github.com/google/clspv/blob/main/lib/SPIRVProducerPass.cpp#L5758), I tried it out and simply removing rsqrt from clspv's libclc SOURCES resulted in clspv generating a call to...

You're right, the ideal solution would be to include that check and then call the glsl InverseSqrt instruction. I'll need to think about how we could do that

with this https://github.com/google/clspv/pull/917 we should be good to just delete sqrt and rsqrt from the builtins library

Yeah it's similar in that many are defined differently (usually vulkan defines the accuracy in terms of other functions, and occasionally it uses absolute error within certain ranges), and of...

opened a revision to add fp16 implementations https://reviews.llvm.org/D135268 incidentally, does anyone know who to contact about getting reviewed libclc revisions committed?

We need soft implementations for a lot of the conversion modes to be conformant there, turns out adding them is pretty easy https://reviews.llvm.org/D136772. Note this does make linking a bit...