https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81177
Bug ID: 81177
Summary: Dump cubin module to file before offloading in nvptx
libgomp plugin
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
CC: jakub at gcc dot gnu.org
Target Milestone: ---
The nvptx libgomp plugin adds a number of ptx sources to a CUlinkState object
using cuLinkAddData, and then links them together using the cuLinkComplete
call:
...
static bool
link_ptx (CUmodule *module, const struct targ_ptx_obj *ptx_objs,
unsigned num_objs)
{
...
GOMP_PLUGIN_debug (0, "Linking\n");
r = CUDA_CALL_NOCHECK (cuLinkComplete, linkstate, &linkout, &linkoutsize);
GOMP_PLUGIN_debug (0, "Link complete: %fms\n", elapsed);
GOMP_PLUGIN_debug (0, "Link log %s\n", &ilog[0]);
if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_error ("cuLinkComplete error: %s", cuda_error (r));
return false;
}
CUDA_CALL (cuModuleLoadData, module, linkout);
CUDA_CALL (cuLinkDestroy, linkstate);
return true;
}
...
After that, the module is loaded using cuModuleLoadData.
For debugging purposes, it would be good to be able to inspect the module.