Closed shwina closed 1 week ago
I see that this logic is being overhauled in https://github.com/apache/arrow-nanoarrow/pull/509/. If the maintainers want to close this PR out, that's fine by me!
It's probably worth including the changes to the tests though.
According to the CUDA docs for cuMemAlloc:
We end up calling
cuMemAlloc()
with0
bytesize when allocating device buffers with no null mask. Thus, the following change was causingnanoarrow_device_cuda_test
to fail:In this PR, I've fixed this by simply skipping the call to
cuMemAlloc
. The resulting buffer will havenullptr
as itsdata
member and0
as itssize_bytes
, which I believe is the desired outcome.I also modified the test above to include cases with no nulls.