arsenm wrote:

> Tested on Intel with rusticl. Though atm we don't support denorms with fp32 
> and I'm wondering if that's related?

I kind of doubt it, this is two near maximum normal values 

> 
> But could also be just some hardware inaccuracy or something else.

There aren't any special case instructions in this function. The main hardware 
variance would be fmuladd handling, but if I force that to the separate mul and 
add, it's still correct. Can you try removing  -ffp-contract=fast-honor-pragmas 
from libclc/CMakeLists.txt?

https://github.com/llvm/llvm-project/pull/188706
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to