Confusing remark on documentation of cuMemcpyDtoH()

Accelerated Computing CUDA CUDA Programming and Performance
1

I find this puzzling.

In the documentation for cuMemcpyDtoH() it is stated:

This function exhibits synchronous behavior for most use cases.

So… how do I know what my use case falls under? It is not further specified what is meant by that!

2

The word “synchronous” in that part of the documentation is a link. Have you clicked on it? The enumeration at the link seems exhaustive.

I would argue that this sentence should not be part of the cuMemcpyDtoH() documentation, because there are no asynchronous cases for device->host transfers. It does, however, apply to host->device transfers.

I will let the NVIDIA folks decide whether the documentation needs updating, or whether there are in fact asynchronous cases for device->host transfers that are not enumerated.