In nvidia.rs, we use cuFuncGetAttribute to workaround, this is because we use cuda driver API to register the module, so we can only use cuda driver api to retrive function information. However, I just realize we use cuda_runtime_sys::cudaFuncAttributes, but this structure might be out-of-date, so some information might be missing for current implementation. E.g. clusterSchedulingPolicyPreference.
So maybe better to define our own structure based on cuda header file, and not use this structure cuda_runtime_sys::cudaFuncAttributes
In nvidia.rs, we use cuFuncGetAttribute to workaround, this is because we use cuda driver API to register the module, so we can only use cuda driver api to retrive function information. However, I just realize we use
cuda_runtime_sys::cudaFuncAttributes
, but this structure might be out-of-date, so some information might be missing for current implementation. E.g.clusterSchedulingPolicyPreference
. So maybe better to define our own structure based on cuda header file, and not use this structurecuda_runtime_sys::cudaFuncAttributes