Fix crashing PixelOnlyGPU workflows (add missing CUDAServices) #31333
Conversation
|
The code-checks are being triggered in jenkins. |
|
assign heterogeneous |
|
please test workflows 136.885502,136.888502,10824.502,10842.502,11634.502,11650.502 |
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-31333/18093
|
|
The tests are being triggered in jenkins.
|
|
A new Pull Request was created by @silviodonato (Silvio Donato) for master. It involves the following packages: RecoVertex/BeamSpotProducer @perrotta, @makortel, @slava77, @christopheralanwest, @tocheng, @cmsbuild, @tlampen, @jpata, @fwyzard, @pohsun can you please review it and eventually sign? Thanks. cms-bot commands are listed here |
| def _addCUDAServices(theProcess): | ||
| theProcess.load("HeterogeneousCore.CUDAServices.CUDAService_cfi") | ||
|
|
||
| modifyRecoVertexBeamSpotProducerBeamSpotAddCUDAService_ = gpu.makeProcessModifier( _addCUDAServices ) |
There was a problem hiding this comment.
It was sort of agreed in #28575 to place this in Configuration/StandardSequences/Services_cff. Unless we want to place this in every future cfi/cff that introduces a CUDA module.
There was a problem hiding this comment.
Indeed a central place would be better.
By the way, how do we deal with this for the GCC 10 builds ?
The CUDAService does not build there (though maybe in principle it could), so there is no CUDAService_cfi.
There was a problem hiding this comment.
By the way, how do we deal with this for the GCC 10 builds ?
TheCUDAServicedoes not build there (though maybe in principle it could), so there is noCUDAService_cfi.
It will indeed continue to fail in GCC 10 on workflows enabling gpu modifier (better than all workflows but not fully satisfactory). Once we agree with the immediate-term solution in #31261, the same pattern could be applied to CUDAService as well.
On the other hand, I'd think the gpu workflows would not make much sense on GCC 10 anyway, so should we look into disabling them in GCC 10 IBs? (in which case we could leave CUDAService untouched for now)
There was a problem hiding this comment.
I think a better solution for the CUDAService could be to try and build it also on GCC 10.
With
diff --git a/HeterogeneousCore/CUDAServices/BuildFile.xml b/HeterogeneousCore/CUDAServices/BuildFile.xml
index e063a0dd4e7..945d272ba85 100644
--- a/HeterogeneousCore/CUDAServices/BuildFile.xml
+++ b/HeterogeneousCore/CUDAServices/BuildFile.xml
@@ -1,4 +1,4 @@
-<iftool name="cuda-gcc-support">
+<iftool name="cuda">
<use name="FWCore/Framework"/>
<use name="FWCore/ServiceRegistry"/>
<use name="FWCore/ParameterSet"/>
diff --git a/HeterogeneousCore/CUDAServices/plugins/BuildFile.xml b/HeterogeneousCore/CUDAServices/plugins/BuildFile.xml
index 73a760fa117..0928247e400 100644
--- a/HeterogeneousCore/CUDAServices/plugins/BuildFile.xml
+++ b/HeterogeneousCore/CUDAServices/plugins/BuildFile.xml
@@ -1,4 +1,4 @@
-<iftool name="cuda-gcc-support">
+<iftool name="cuda">
<use name="cuda"/>
<use name="DataFormats/Common"/>
<use name="DataFormats/Provenance"/>
@@ -11,6 +11,7 @@
<use name="FWCore/ServiceRegistry"/>
<use name="FWCore/Utilities"/>
<use name="HeterogeneousCore/CUDAServices"/>
+
<library file="*.cc" name="HeterogeneousCoreCUDAServicesPlugins">
<flags EDM_PLUGIN="1"/>
</library>
diff --git a/HeterogeneousCore/CUDAUtilities/BuildFile.xml b/HeterogeneousCore/CUDAUtilities/BuildFile.xml
index fc2752ff845..e7193e1ccfd 100644
--- a/HeterogeneousCore/CUDAUtilities/BuildFile.xml
+++ b/HeterogeneousCore/CUDAUtilities/BuildFile.xml
@@ -1,4 +1,4 @@
-<iftool name="cuda-gcc-support">
+<iftool name="cuda">
<use name="cuda"/>
<use name="eigen"/>
<use name="FWCore/Utilities"/>the CUDAService can be built and loaded.
There was a problem hiding this comment.
The idea is that <iftool name="cuda"> tells if we have the CUDA libraries, while <iftool name="cuda-gcc-support"> tells if we can compile the device code.
|
+1 |
|
Comparison job queued. |
move modifier to Configuration/StandardSequences/python/Services_cff.py
silviodonato
force-pushed
the
addCUDAservices
branch
from
c44a78a to
b282750
Compare
|
please test workflows 136.885502,136.888502,10824.502,10842.502,11634.502,11650.502 |
|
The tests are being triggered in jenkins.
|
|
Comparison is ready @slava77 comparisons for the following workflows were not done due to missing matrix map:
Comparison Summary:
|
|
+1 |
|
Comparison job queued. |
cmsbuild
added
comparison-pending
tests-approved
and removed
comparison-available
tests-started
labels
|
Comparison is ready @slava77 comparisons for the following workflows were not done due to missing matrix map:
Comparison Summary:
|
|
+operations |
|
+1 |
|
This pull request is fully signed and it will be integrated in one of the next master IBs (tests are also fine). This pull request will now be reviewed by the release team before it's merged. @silviodonato, @dpiparo, @qliphy (and backports should be raised in the release meeting by the corresponding L2) |
cmsbuild
added
fully-signed
heterogeneous-approved
and removed
heterogeneous-pending
pending-signatures
labels
|
+1 |
PR description:
This PR fixes the crash of workflows 136.885502,136.888502,10824.502,10842.502,11634.502,11650.502 as reported in #31130 (comment) .
This fix loads
HeterogeneousCore.CUDAServices.CUDAService_cfiinRecoVertex/BeamSpotProducer/python/BeamSpot_cff.pythrough thegpuModifier.PR validation:
runTheMatrix.py -l 136.885502works