Allow Runtime Number of Hits for Alpaka Pixel Reconstruction#44773
Allow Runtime Number of Hits for Alpaka Pixel Reconstruction#44773cmsbuild merged 1 commit intocms-sw:masterfrom
Conversation
|
cms-bot internal usage |
|
tagging @malbouis (ORM) |
|
urgent |
|
type hlt-int |
|
enable gpu |
|
I think as soon as the checks finish we can merge in the interest of time. |
|
test parameters:
|
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-44773/40004
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
- update for HitToTuple map in CA
d1d19b5 to
29d1688
Compare
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-44773/40005
|
|
A new Pull Request was created by @AdrianoDee for master. It involves the following packages:
@civanch, @Dr15Jones, @jfernan2, @makortel, @mdhildreth, @mandrenguyen, @cmsbuild, @bsunanda can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
|
please test |
|
assign heterogeneous |
|
+1 Summary: https://cmssdt.cern.ch/SDT/jenkins-artifacts/pull-request-integration/PR-bfffbf/38935/summary.html Comparison SummarySummary:
GPU Comparison SummarySummary:
|
|
+1 |
|
merge |
|
+1 |
|
+heterogeneous |
This PR proposes a solution for #44769. The culprit there is
cmssw/RecoLocalTracker/SiPixelClusterizer/plugins/alpaka/SiPixelRawToClusterKernel.dev.cc
Lines 509 to 513 in 6e82cc8
where we cut the number of hits to "only":
cmssw/Geometry/CommonTopologies/interface/SimplePixelTopology.h
Line 420 in 6e82cc8
And the event of the crash has:
The poor man solution would be to rise the max to something safer such as
static constexpr uint32_t maxNumberOfHits = 96 * 1024;. But seems to me a waste also because this procedure is anyway a remnant of that intermediate phase of the Alpaka port when we had no runtime sized histograms (before #43064).So this PR involves a slightly bigger code refactoring in order to drop the fixed number of hits and use the runtime sized histograms where needed (namely for the hit to tuple map).
PR validation:
The setup in #44769 with this PR (on top of
masteror14_0_X) doesn't crash anymore.Backported to 14_0_X in #44774.
solves #44769.