migration of cudaMemcpyToSymbol

[zjin-lcf]

Could you please advise how to map cudaMemcpyToSymbol() in CUDA to SYCL USM ? Thanks.

[JackAKirk]

This question is relevant for the current work on using constant symbols. Basically I think that this functionality isn't fully supported yet but it is being worked on. We may have to consider whether we want something that maps directly to cudaMemcpyToSymbol() or just cudaGetSymbolAddress I guess. CC @jchlanda

[zjin-lcf]

The USM allocation functions with properties support (#6346) may be relevant.

I assume that users first allocate a memory by calling the function malloc_device(). The function accepts an attribute for the constant memory space.
Then, users copy data from a host to this allocated memory with q.memcpy().
Finally, the memory is accessed like other global memory in a SYCL kernel.

In a CUDA program, a CUDA kernel could access a constant memory without passing it as a kernel argument. In some CUDA applications, there are many constant memories, so this improves productivity at least.