Re: [patch, libgomp] Enable OpenACC GCN testing

[Andrew Stubbs <andrew_stubbs at mentor dot com>]

On 15/11/2019 12:43, Jakub Jelinek wrote:
> > APUs, such as Carizzo are shared memory. DGPUs, such as Fiji and Vega, have
> > their own memory. A DGPU can access host memory, provided that it has been
> > set up just so, but that is very slow, and I don't know of a way to do that
> > without still having to copy the program data into that special region.
>
> Ah, that is going to be interesting e.g. for
> #pragma omp requires unified_shared_memory
> My plan was that for this the TU constructor would call some libgomp
> library function that in the end would when loading plugins ensure that
> only shared memory supporting plugins are loaded.  If the gcn plugin
> will support shared memory only conditionally, we'll need some interfaces to
> ensure that.

This is a problem indeed. Right now the capability has to be declared up 
front, before the device has even been initialized. In theory, there 
could be both devices in a single machine, and that would be even worse, 
but then we don't support machines with heterogeneous DGPUs well either 
(it works fine for user programs that just use the GPU they're given).

We've been basically ignoring it because a) we don't have any APUs to 
test with, and b) we're being paid to produce a toolchain for 
supercomputers that will never use APUs (nor mismatched DGPUs).

The APU can (pretend to) not use shared memory, of course, so it might 
Just Work, but I can't be sure.

There's also the problem of the XNACK support, which we haven't 
implemented in the backend. That's specific to APUs, so we'll never need 
it. You can get some distance without -- there was no mention of it in 
Honza's original port that targeted Carizzo -- but I don't really know 
how far.

Really, if nobody with an APU wants to fix it, we should just remove the 
Carizzo support that is there.

Andrew