[SYCL][COMPAT] Memset API updated to support 2-byte and 4-byte memsets #11340
Conversation
Alcpz
commented
Sep 28, 2023
Alcpz
commented
- The memory interface has changed:
- fill has been removed
- memset now offers a templated version (which works as fill did)
- memset_d8, memset_d16, and memset_d32 were added to allow memsetting 1, 2 and 4 byte-sized values.
|
Working through this PR, I see that we have previously been somewhat inconsistent about |
joeatodd
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
joeatodd
temporarily deployed
to
WindowsCILock
GitHub Actions
Inactive
|
@konradkusiak97 is this a duplication of your fill work? Maybe there should just be a single interface for this functionality? |
After #12702 is merged, the fill command will use backend-specific calls like |
Yeah that is what I was thinking. That would mean that sycl spec already supports this functionality in theory and you have now implemented it in practice. Therefore this compat PR doesn't seem to be a good idea? Why not just translate cuda code to the intended part of the sycl specification? |
|
In this case, SYCLcompat wraps over the fill functionality @konradkusiak97 has been working on (in the specific case of the CUDA backend). While I agree that we should translate that way directly, this serves both as an example of how to translate, and is needed for applications that have been already been translated. |
|
Closed. Replaced by #13409 |
…#13409) This PR replaces #11340 This PR extends the memory header to include 2 byte and 4 byte memsets. - memset remains unchanged. - 2D / 3D memsets are templated and wrap `sycl::fill`. Functionality remains unchanged as it is exposed through `detail::memset<unsigned char>`, equivalent to what we had before. - memset_d16 and memset_d32 calls are added wrapped around `sycl::fill` using 2-byte and 4-byte datatypes Added tests for memset_d16 and memset_d32.
