Clean up AnalyzePolicy #3564
Clean up AnalyzePolicy #3564
Conversation
|
Retest this please |
| template <class... Traits> | ||
| struct AnalyzePolicy<void, Impl::IsGraphKernelTag, Traits...> | ||
| : AnalyzePolicy<void, Traits...> { | ||
| using is_graph_kernel = std::true_type; |
There was a problem hiding this comment.
Are multiple GraphKernelTags allowed? Don't you need to make sure that none of the base class template parameters is Impl::IsGraphKernelTag?
| typename PolicyBase::launch_bounds, | ||
| typename PolicyBase::work_item_property, std::true_type>; | ||
| //------------------------------------------------------------------------------ | ||
| // Work item propoerty case |
There was a problem hiding this comment.
| // Work item propoerty case | |
| // Work item property case |
| // A tag class for dependent defaults that must be handled by the | ||
| // PolicyTraitsWithDefaults wrapper, since their defaults depend on other traits | ||
| struct dependent_policy_trait_default; |
|
Apparently
so apparently |
Right now, the size of the code for AnalyzePolicy scales quadratically with the number of traits we analyze. With this change, the scaling is linear and adding a new trait doesn't have to touch any other code.
The details of the runtime storage that may be associated with a policy trait is now implemented only as part of the trait-specific specializations. Traits are now fully disjoint with the exception of defaults in AnalyzePolicy<void> and PolicyTraitsWithDefaults
dhollman
force-pushed
the
analyze-policy-cleanup
branch
from
6f572a5
to
200440f
Compare
|
More extensive disentangling of Policy Traits (Replaces #3564)
Right now, the size of the code for
AnalyzePolicyscales quadraticly with the number of traits we analyze. With this change, the scaling is linear and adding a new trait doesn't have to touch any other code. There's nothing fancy going on here; we're just using partial specialization to detect the various traits. The one trick that was missing before is that we have to extract the work tag using an unspecialized template that defers to a base with a different name (which is partially specialized) before going back to the analysis of the rest of the traits.