On Tue, 1 Apr 2025, Richard Biener via Gcc wrote:

> In C++ there could be runtime initializers for a const qualified
> object.  I think all
> you need to do is make sure the logic that places the object in .const
> vs. .global
> is consistent with the logic deciding how to access it.

I think it is misguided to implicitly use PTX .const memory space for all
readonly data. CUDA constant memory is a specialized memory region, limited
in size (64KB) with small dedicated caches (8 KB). It is intended for data
that is accessed in a uniform manner by threads. It would be more appropriate
to treat it like .shared memory (exposed in nvptx via an attribute).

Alexander

Reply via email to