Add Device Memory Support to GRIN

GRIN provides graph retrieval APIs for graph computing engines to access various storages in a uniform way. As GRIN also supports GNN engines similar to PyG, some storages in our workload may store graphs in device memory such as GPUs. Therefore, it is necessary to add device memory support to GRIN. This feature will enable GNN engines to access device memory storages in a uniform way. In this issue, we will propose the device memory specific APIs of GRIN, and PoC the API design using the storage of PyG.

To add device memory support to GRIN, we propose to implement custom APIs for device memory access that are tailored to the needs of GRIN. The PyTorch library can be used to provide efficient tensor operations for GPUs and other accelerators. PyTorch also provides a unified API for CPU and GPU tensors, which can simplify the implementation of device memory support in GRIN.

We propose to define a new set of GRIN APIs that provides device memory specific functions. The APIs may include functions for allocating and deallocating device memory, copying data between host and device memory, and performing tensor operations on device memory.

Thank you.