Stream sync before calling cufile #437
Conversation
madsbk
added
bug
madsbk
force-pushed
the
stream_sync_when_calling_cufile
branch
from
822a40f to
63f750f
Compare
madsbk
added
improvement
cpp/include/kvikio/file_handle.hpp
Outdated
| @@ -314,16 +314,26 @@ class FileHandle { | |||
| * @param file_offset Offset in the file to read from. | |||
| * @param devPtr_offset Offset relative to the `devPtr_base` pointer to read into. | |||
| * This parameter should be used only with registered buffers. | |||
| * @param sync_null_stream Synchronize the CUDA null stream prior to calling cuFile. Contrary to | |||
There was a problem hiding this comment.
Do we mean "default stream" instead of "null stream" everywhere here? https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html
There was a problem hiding this comment.
Right, I have renamed to default stream. cuFile calls it the null stream, but I agree default stream is less confusing.
…o stream_sync_when_calling_cufile
There was a problem hiding this comment.
This looks good!
Would an equivalent alternative be to use the async API and pass in the default stream? Is that possible with the current design? If so, I would explain in the docs how this new option behaves compared to that alternative (or simply direct users to an async API rather than add a new parameter if it is equivalent).
Thanks @bdice, the async API comes with its own quirks so it isn't really an equivalent alternative. |
|
/merge |
Contrary to most of the non-async CUDA API, cuFile does not have the semantic of being ordered with respect to other work in the null stream. Since this is surprising to most users, we now synchronize on the null stream before calling cuFIle read and write by default.
See cuFile IO API Functional Specification and CUDA Concurrent Execution between Host and Device.