Bug#1021660: gcc-12-offload-nvptx: offloading to nvidia via nvptx fails with cuda version 11 (default in sid)
On Thu, 13 Oct 2022, Thomas Schwinge wrote:
It does work, but only for the code that GCC/nvptx generates in that
'gcc-12' invocation, but not for the support libraries that it linkes in,
which are built for sm_30.
Does this mean then that the support libraries of gcc-11-offload-nvptx
include both support for sm_30 and sm_35? Is it possible to compile such
support libraries so that they do support more than one cuda arch level,
instead of having, as in the case of GCC 12 support libraries, _only_ sm_30
as available option (if I understood you correctly)?
However, that doesn't really help you as a user of GCC, as long as the
distributions don't (have an easy way to) build more variants for several
sm_[...]. More work is necessary in GCC/nvptx upstream to make that
feasible.
well, debian in itself does support this kind of setup, doesn't it? With
alternatives, provides in dpkg... Of course, I gather that putting together
the machinery to build a number of versions of the same package would be
somewhat of a pain to set up and maintain.
But anyway, given all you said, can this issue be solved at all, even acting
on nvptx-tools? If the issue lies in the support libraries, that problem
would still remain regardless of what you do on nvptx-tools, wouldn't it?
Thanks, bye
Giacomo
--
_________________________________________________________________
Giacomo Mulas <giacomo.mulas@inaf.it>
_________________________________________________________________
INAF - Osservatorio Astronomico di Cagliari
via della scienza 5 - 09047 Selargius (CA)
tel. +39 070 71180255
mob. : +39 329 6603810
_________________________________________________________________
"It's just a shadow of the man you should be
Like a garden in the forest that the world will never see
You have no thought of answers only questions to be filled"
(Big Country)
_________________________________________________________________
Reply to: