ldc/dcompute and shared (programmer managed cache) access

[Johan Engelen]

On Friday, 5 March 2021 at 04:49:56 UTC, Bruce Carneal wrote:
>
> The code generation through standalone (uninitialized) 
> __shared__ pointers, which we *can* declare, already emits the 
> correct PTX instructions (ld.shared, st.shared) but we have no 
> way, at least that I've found, to declare __shared__ address 
> space storage that we can point to.  (clang cuda can do this, 
> nvvc can obviously do this, and we can do this currently only 
> if we manually modify the .PTX file)

I'm not very familiar with DCompute, but sounds to me that all 
that is needed is a proposed syntax for what you want. The 
compiler support is probably pretty easy to add.

-Johan