Hi! On Thu, 17 Dec 2015 20:26:56 +0300 (MSK), Alexander Monakov <amonakov@xxxxxxxxx> wrote: > On Thu, 17 Dec 2015, Esteban Hernández wrote: > > libgomp: cuModuleLoadData error: device kernel image is invalid > > Are you perhaps using new Maxwell-generation GPU for testing (Geforce 750, > 950+)? There's odd code in libgomp plugin that insists on producing binary > code for Kepler-class hardware only. > > Nathan, is that a mistake, or was there some reason to fix JIT target? We are > removing that code on gomp-nvptx branch: > https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=1099ad42d145faf428053018d02b26678c47adce Bernd, Jim, do you remember the history why in [GCC]/libgomp/plugin/plugin-nvptx.c:link_ptx we force CU_JIT_TARGET:CU_TARGET_COMPUTE_30 -- apparently for no good reason? (And then, similar code exists in [nvptx-tools]/nvptx-run.c.) Unless there is any reason, would you, Alexander, please commit your patch to trunk? Grüße Thomas
Attachment:
signature.asc
Description: PGP signature
