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;
|
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);
|
acDeviceLoadMesh(device, STREAM_DEFAULT, submesh);
|
||||||
|
|
||||||
// Benchmark
|
// Benchmark
|
||||||
|
Reference in New Issue
Block a user