pytorch_scatter
pytorch_scatter copied to clipboard
Fix idx overflows in scatter_cuda kernel to enable benchmark/scatter_segment.py with long rows
This PR exists to show the "state of the world" for the "before" benchmark (rendered benchmarking results cuda)
- fixes CUDA illegal memory accesses run into by benchmarking script mentioned on #301 that are run into when
src.numel() > std::numeric_limits<int>::max()
on reduce options other thansum
andmean
- Comments out arg kernel for max/min since we do not support argmax/argmin in core
- Some small edits to
benchmark/scatter_segment.py
to reproduce errors on master prior to this PR run python benchmark/scatter_segment.py --reduce max
, illegal memory accesses start occuring on the Janna dataset for size=128
Codecov Report
Merging #303 (4b34a87) into master (847995d) will not change coverage. The diff coverage is
n/a
.
@@ Coverage Diff @@
## master #303 +/- ##
=======================================
Coverage 97.54% 97.54%
=======================================
Files 9 9
Lines 204 204
=======================================
Hits 199 199
Misses 5 5
:mega: Codecov can now indicate which changes are the most critical in Pull Requests. Learn more
This pull request had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity.