Should OpenCL C have a memcpy?

[karolherbst]

At the moment the OpenCL C spec states:

Of these, only memcpy is strictly correct in C99. Since OpenCL does not provide memcpy, other methods are needed.

Should we provide a type generic (for all OpenCL C scalars and vector types?) memcpy in OpenCL C? In SPIR-V we have OpCopyMemory and OpCopyMemorySized?

I think I'd prefer to provide a higher level function to developers rather than them trying to implement memcpy with loops or something that are fundamentally more difficult to optimize and more difficult to reason about pointer alignments.

Should we have a memcpy extension for this?

Should we have a memcpy_aligned version where developers can specify explicit pointer alignment?

[Oblomov]

I think I'd prefer to provide a higher level function to developers rather than them trying to implement memcpy with loops or something that are fundamentally more difficult to optimize and more difficult to reason about pointer alignments.

I've had bad experience in the past with implementations of async_work_group_copy & friends, with trivially coded manual copies being (sometimes even significantly) faster. A built-in memcpy without any guarantees that its performance will be “no less efficient” than at least trivially coded manual copy will mean that developers will still have to try manual coding, and potentially use separate paths on different vendors anyway.