Hi,
On 10/9/25 22:53, Tobias Burnus wrote:
Hi Joseph,
Josef Melcr wrote:
Thank you for the detailed explanation :) It should be fairly easy to
implement if ipa-cp acts before the tables are dumped. Otherwise it
would probably require a bit of work, but if we implement some
reorganizing logic for the tables, then we might be able to drop the
noclone attribute altogether (if I understand it correctly). I will
tinker around with it in the future, unfortunately I can't get the
output_offload_tables to run, so I will have to figure that out. But
the application sounds really exciting.
I think you only get this when you build with offloading
support – either for Nvidia GPUs (nvptx) or AMD GPUs (gcn)
or for both.
https://gcc.gnu.org/wiki/Offloading#Building_and_Obtaining_GCC
describes how you do this.
That was indeed it, I finally got it running, thank you. It seems that
the decision stage of ipa-cp is executed after outputting the tables, so
the tables would have to be updated. I am not sure how difficult that's
going to be, but having the ability to clone the kernels would be
useful, so it's definitely worth looking into. Not sure when I'll have
the time to do so, but it's on my radar :)
Tobias
Best regards,
Josef
