Upgrade CMSSW to 14_1_0_pre0

SegmentLinking / TrackLooper

Apache License 2.0

5 stars 14 forks source link

Upgrade CMSSW to 14_1_0_pre0 #373

Closed ariostas closed 8 months ago

ariostas commented 8 months ago

Following @slava77's suggestion in #372, this PR upgrades CMSSW to 14_1_0_pre0. There were some minor changes that needed to be made to account for the newer version of Alpaka.

Since it now uses CUDA 12.2, I checked using the Warnings flag to see if the printf bug #354 was fixed. It wasn't fixed, but I did notice that there were some errors due to the warnings being treated as errors that I forgot to fix in #367, so I fixed them.

I marked this PR as draft since it shouldn't be merged yet for a few reasons:

cgpu1 currently has a driver version that is not compatible with CUDA 12
There needs to be a corresponding branch in the cmssw repo (I'll work on that)
We could use pre1 if it happens to release soon

ariostas commented 8 months ago

/run cmssw CMSSW_14_1_0_pre0_LST_X

github-actions[bot] commented 8 months ago

The PR was built and ran successfully with CMSSW. Here are some plots.

OOTB All Tracks

Efficiency and fake rate vs pT, eta, and phi

The full set of validation and comparison plots can be found here.

ariostas commented 8 months ago

I added a bug fix that @slava77 found out since it's somewhat related to the CUDA upgrade and it's small enough that it doesn't really need its own PR.

nvcc was only generating SASS code for sm_70 and sm_89. This meant that the A30 on cgpu-1 couldn't be used since it's sm_80. Now it also generates PTX code for compute_70, which can be JIT-compiled to sm_80.

slava77 commented 7 months ago

There is an effort to add cuda 12.4, perhaps that will be a better opportunity than catching pre1.

CUDA 12.4 is available since 14_1_0_pre2 (released Mar 28)

ariostas commented 7 months ago

CUDA 12.4 is available since 14_1_0_pre2 (released Mar 28)

I think we should wait for pre_3 since the ROCm fix didn't make it to pre_2