Split the CUDA CKF into different TUs #742
Conversation
stephenswat
added
refactor
stephenswat
force-pushed
the
refactor/ckf_tus
branch
2 times, most recently
from
bbeb4f7
to
30aebf1
Compare
|
My plan to finally get rid of those .ipp files has failed. 🫡 Anyway, updated. |
stephenswat
force-pushed
the
refactor/ckf_tus
branch
4 times, most recently
from
24a9c9e
to
4ec0290
Compare
There was a problem hiding this comment.
Okay, let's go with this general setup at the moment. As we discussed in person, I have some further ideas for how to tweak the code later on even further, but these changes do not make those future changes any more difficult. (The general direction of my idea is very similar.)
| typename detector_t::view_type det_data; | ||
| const int n_params; | ||
| bound_track_parameters_collection_types::view params_view; | ||
| vecmem::data::vector_view<const unsigned int> params_liveness_view; |
There was a problem hiding this comment.
Some Doxygen documentation should be put on the variables, and the struct as well. Since now that documentation is not on the main function anymore.
stephenswat
force-pushed
the
refactor/ckf_tus
branch
from
4ec0290
to
54e7489
Compare
|
Great! I've rebased and incorporated the change requests |
This commit splits the monstrously large CUDA track finding translation unit up into smaller ones, one for each of the kernels. This should speed up compilation times and decrease memory usage. Also groups the payloads for each of the functions into convenient structs, so we don't need to pass 20+ arguments for some of the kernel calls. Does not change the functionality of the code.
|
| measurement_collection_types::const_view measurements_view, | ||
| bound_track_parameters_collection_types::const_view seeds_view, | ||
| vecmem::data::jagged_vector_view<const candidate_link> links_view, | ||
| vecmem::data::vector_view<const typename candidate_link::link_index_type> | ||
| tips_view, | ||
| track_candidate_container_types::view track_candidates_view, | ||
| vecmem::data::vector_view<unsigned int> valid_indices_view, | ||
| unsigned int* n_valid_tracks); |
There was a problem hiding this comment.
Forgot to use traccc::device::build_tracks_payload here.
| unsigned int* shared_num_candidates; | ||
| std::pair<unsigned int, unsigned int>* shared_candidates; | ||
| unsigned int& shared_candidates_size; |
There was a problem hiding this comment.
As long as I complain anyway, these variables should also get some description.
This commit splits the monstrously large CUDA track finding (and some CCL) translation unit up into smaller ones, one for each of the kernels. This should speed up compilation times and decrease memory usage.
Also groups the payloads for each of the functions into convenient structs, so we don't need to pass 20+ arguments for some of the kernel calls.
Does not change the functionality of the code.