Mariya Podchishchaeva
Mariya Podchishchaeva
Could you please take a look at failing test?
> @Fznamznon, can we consider passing the whole SYCL kernel (lambda object) via single kernel parameter? This should allow reduce overhead on setting multiple kernel parameters. I think it will...
FYI: @elizabethandrews , @gmlueck , results are good so far. I'm waiting for summary run.
> If such classes are still decomposed (due to some other special type in them), do we want the pointers to remain wrapped? I would say yes. This wrapping was...
Okay, I see a problem here. Due to decomposition all pointer struct members were transformed to kernel arguments with global address space. When we stop decomposing structs that contain pointers...
> > * Try to pass structs with global pointer inside, but convert them to structs with generic pointers inside to avoid further problems with IR emission (that will be...
> All you're doing here is building in an assumption to the CFE that all pointers in kernel arguments are to the global address space. Not quite. The assumption is...
> @Fznamznon what are the possible explicit address spaces? Looking at the code, it looks like only sycl_global_device and sycl_global_host address spaces are retained Technically can be any of address...
> I guess a more precision definition of the assumption is that any undecorated pointer in a kernel argument can be assumed to be a global address space pointer. Yes,...