H100 cluster group compilation error
I use RunPod Desktop on Secure H100 (both SXM5 and PCI3).
CUDA Driver Version / Runtime Version 12.1 / 11.8
CUDA Capability Major/Minor version number: 9.0
I am trying to use cluster groups but having trouble in compiling files.
I first tested a cuda example file simpleCooperativeGroups.cu (https://github.com/NVIDIA/cuda-samples.git). This file has a line “#include <cooperative_groups.h>” and “using namespace cooperative_groups;” at the beginning. In the global function the function has “thread_block” and “this_thread_block()”, which are members of cooperative_groups. The file can be compiled and run without a problem.
I next added lines “cluster_group cluster = this_cluster();” in the the global function. Now I get error messages like:
error: namespace "cooperative_groups" has no member "cluster_group"
error: namespace "cooperative_groups" has no member "this_cluster."
I get similar errors even when I exchange the line with “cooperative_groups::cluster_group cluster = cooperative_groups::this_cluster();” or “cg::cluster_group cluster = cg::this_cluster();”.
So, it seems that the complier and H100 hardware are only partially compatible with cooperative_groups.h and cannot handle thread block clusters. Can I have any help?
1 Reply
PCI3 -> PCIe