Test cuda graph capture #1112
Test cuda graph capture #1112
Conversation
gevtushenko
requested review from
alliepiper,
miscco and
elstehle
and removed request for
a team
cub/test/CMakeLists.txt
Outdated
| ## cub_cdp_by_lid | ||
| # | ||
| # If given launcher id corresponds to a CDP launcher, set `out_var` to 1. | ||
| function(cub_cdp_by_lid out_var launcher_id) |
There was a problem hiding this comment.
Suggestion: rename this function to _cub_launcher_requires_rdc.
There was a problem hiding this comment.
Could you please elaborate on this suggestion so that I can understand intention behind it? Is underscore used to add semantics of a "private" function? If so, why cub_get_test_params and cub_add_test do not have leading underscore in their names?
There was a problem hiding this comment.
The intention of the underscore was to distinguish between "heavy" functions that might be used repeatedly across multiple files as a sort of API vs. helper functions that just simplify code flow in our implementations (like these one-off functions that don't do anything interesting except map one value into another to make life a little easier).
This was more important for thrust, which exposes some functions in the cmake package (thrust_create_target) that are actual user APIs. My goal was to have the more useful functions appear in IDE autocompletion while the helpers wouldn't clutter up the list of completions. It doesn't have any actual impact on correctness, though, it was just a convention.
In this suggestion, the new name was more important with the underscore. launcher_requires_rdc makes it clear that this is used solely to map launcher values to RDC flags.
There was a problem hiding this comment.
Thank you for elaborating on this! I've changed the function name per recommendation. The convention doesn't seem intuitive. If I'm not missing anything, it's not documented. Maybe it's worth adding cmake developer overview, describing it.
Description
closes #321
This PR generalizes CDP wrapper to support multiple launch backends.
It also introduces graph capture backend and adds coverage for graph capture to basic variants of device-scope tests.
Checklist