Hi all,
is there any documentation about the alignment of the memory returned by cuMemAllocHost / cuMemHostAlloc ?
The documentation from CUDA 10.2 doesn’t say anything about it (while for cuMemAlloc / cuMemAllocManaged it does specify that “the allocated memory is suitably aligned for any kind of variable”.
From a quick test (Linux, GTX 960M, CUDA 10.2) it looks like the memory is aligned to1024 bytes - but it would be useful to know if this is just and accident or something guaranteed by the CUDA driver.
Thank you,
.Andrea