Open gigaio-staylor opened 6 months ago
I have found this issue when compiling mpich 4.2.1 with ROCm 6.1.0. The suggested change has solved the compilation error.
On further testing, it looks like the enum hipMemoryType
values, and errors returned by hipPointerGetAttributes
changed.
hipPointerGetAttributes
return success when an 'unregistered' buffer is passed in, and there's a new hipMemoryTypeUnregistered
enum value.
ROCM-6 hipMemoryType
:
https://github.com/ROCm/HIP/blob/900a5e9828f55ae004de99314c5f992e6a0ebc9b/include/hip/hip_runtime_api.h#L245
Also, this breaks the logic of MPI_gpu_query_pointer_attr so it will need to be rewritten...
Looks like MPICH code needs to change from using
memoryType
totype
.Per ROCm 6.0 Changelog: Deprecated enum hipMemoryType memoryType; is removed from HIP struct hipPointerAttribute_t union.
Affected files appear to be:
./modules/yaksa/src/backend/hip/pup/yaksuri_hipi_get_ptr_attr.c ./src/pmi/mpl/src/gpu/mpl_gpu_hip.c ./src/mpi/romio/mpl/src/gpu/mpl_gpu_hip.c ./src/mpl/src/gpu/mpl_gpu_hip.c ./src/pm/hydra/modules/pmi/mpl/src/gpu/mpl_gpu_hip.c ./src/pm/hydra/modules/mpl/src/gpu/mpl_gpu_hip.c