Dist1 coloring PPS improvements #578
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
instead of magic numbers 0,1,2
PPS worklist construction (VB and EB) now happens in a single scan, without using a temporary array to store indices. Faster and saves memory.
srajama1
reviewed
Jan 27, 2020
| Kokkos::parallel_for ("KokkosGraph::GraphColoring::CreateNewWorkArray", | ||
| my_exec_space(0, current_vertexListLength_), | ||
| create_new_work_array<nnz_lno_temp_work_view_t>(current_vertexList_, next_iteration_recolorList_, pps_work_view)); | ||
| ppsWorklistFunctorVB<nnz_lno_temp_work_view_t>(this->nv, current_vertexList_, next_iteration_recolorList_)); |
There was a problem hiding this comment.
Does the values in recolocList need to be reinitialized ?
There was a problem hiding this comment.
@srajama1 It doesn't need to be initialized because the exact size of the next worklist is produced by a parallel_reduce of functorFindConflicts.
There was a problem hiding this comment.
I meant reinitialized, from second iteration to third iteration
srajama1
approved these changes
Jan 29, 2020
There was a problem hiding this comment.
@brian-kelley gave me a lesson on why we don't need the parallel prefix array.
Thanks @brian-kelley
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Distance-1 coloring with parallel prefix sum worklist used to require 2 passes: a scan to get new worklist index view, and a for to populate the worklist. The original paper about it even says that the 2-pass approach is a downside of PPS compared to atomics.
This makes PPS use a single scan pass that scatters directly to the new worklist without an index view. This saves some time, one kernel launch per iteration and a significant amount of memory:
(num_edges / 2) * sizeof(nnz_lno_t)for EB, andnum_rows * sizeof(nnz_lno_t)for VB.Also, renamed
_conflictlistmember to_conflictlist_schemefor clarity - it's not a list, it's an enum representing how the worklist is built. I also use those actual enum constants (COLORING_NOCONFLICT,COLORING_ATOMIC,COLORING_PPS) instead of the raw integer values (0,1,2) throughout the code.Bowman checks:
#######################################################
PASSED TESTS
#######################################################
intel-16.4.258-Pthread-release build_time=724 run_time=1027
intel-16.4.258-Pthread_Serial-release build_time=1034 run_time=1998
intel-16.4.258-Serial-release build_time=706 run_time=963
intel-17.2.174-OpenMP-release build_time=869 run_time=566
intel-17.2.174-OpenMP_Serial-release build_time=1225 run_time=1444
intel-17.2.174-Pthread-release build_time=814 run_time=899
intel-17.2.174-Pthread_Serial-release build_time=1131 run_time=1803
intel-17.2.174-Serial-release build_time=778 run_time=891
RIDE checks:
#######################################################
PASSED TESTS
#######################################################
cuda-9.2.88-Cuda_OpenMP-release build_time=517 run_time=418
cuda-9.2.88-Cuda_Serial-release build_time=507 run_time=527
gcc-6.4.0-OpenMP_Serial-release build_time=194 run_time=396
gcc-7.2.0-OpenMP-release build_time=120 run_time=128
gcc-7.2.0-OpenMP_Serial-release build_time=214 run_time=357
gcc-7.2.0-Serial-release build_time=112 run_time=226