Device id is now properly set when using MPI and there are multiple visible GPUs per node
This commit is contained in:
@@ -1242,7 +1242,9 @@ acDeviceRunMPITest(void)
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
Device device;
|
||||
acDeviceCreate(0, submesh_info, &device);
|
||||
int devices_per_node = -1;
|
||||
cudaGetDeviceCount(&devices_per_node);
|
||||
acDeviceCreate(pid % devices_per_node, submesh_info, &device);
|
||||
acDeviceLoadMesh(device, STREAM_DEFAULT, submesh);
|
||||
|
||||
// Benchmark
|
||||
|
Reference in New Issue
Block a user