Allow extension of built-in tuners with additional tuning axes #3961
Conversation
… into feature/cross-tune
|
Any chance that you get rid of the massive indentation changes in |
|
I have no idea what set of clang-formats, IDE's, and other lovely scripts caused that. I'll try to get to it Monday |
| MultidimensionalSparseTuningProblem(StoredProblemSpace space, | ||
| const std::vector<std::string>& names) |
There was a problem hiding this comment.
Since StoredProblemSpace is private, I guess this constructor can also be private.
There was a problem hiding this comment.
This constructor needs to be public. We construct tuners using this directly when we extend them
core/src/Kokkos_Tuners.hpp
Outdated
| @@ -65,19 +65,29 @@ namespace Tools { | |||
| namespace Experimental { | |||
|
|
|||
| // forward declarations | |||
| SetOrRange make_candidate_set(size_t size, int64_t* data); | |||
| SetOrRange make_candidate_set(size_t size, int64_t *data); | |||
|
|
|||
| bool have_tuning_tool(); | |||
There was a problem hiding this comment.
Are all these whitespace changes intentional?
There was a problem hiding this comment.
So, I'm applying apply-clang-format. I don't want to worry about this kind of thing without a good reason.
Co-authored-by: Daniel Arndt <arndtd@ornl.gov>
masterleinad
dismissed
their stale review
CMakeLists.txt isn't changed anymore.
|
The indentation is still off (due to my unindented suggestions). 🙂 |
|
@masterleinad : feel free, though I think it's an iffy use of a person's time |
Implements part 1 of #3960 . Right now, a user can't say "I want to tune what a team size does and my own thing" without making nested tuning problems, which get tricky very quickly. This allows that functionality, a user can simply call "combine" on an instance of one of our tuners, pass in a vector of items, and get back a tuner that tunes their axis as well as what the original tuner tuned. I also add tests for this. The intended use case is in Kokkos Kernels, in which to tune their SPMV, they need to tune the team size and vector length (TeamSizeTuner), but also need to tune the number of rows of a matrix assigned to each thread. Since the number of rows is orthogonal to the TeamPolicy configuration, they can simply combine that axis into a TeamSizeTuner and be good to go
Note that as of now the tests refer to things in namespace
Impl. I'm considering moving those things out of the Impl namespace, but don't want to bloat this PR.