Global (currently TPC only) refit on GPU using either GPU or TrackParCov track model

[davidrohr]

@shahor02 : For reference. The TrackParCov and propagator now work on the GPU (using sergeys polynomial version of the magnetic field, the MagFieldFast is not compatible). I did a quick benchmark, and it is ~2x slower than Sergey's model on the GPU. On the CPU Sergey's model is still ~30% faster.

[shahor02]

Thanks! Was your benchmark with mat.LUT attached? If I understand correctly, with GPU track model a pre-calculated material correction is applied, while with TrackParCov the material is queried per step.

[davidrohr]

@shahor02 : Good point. I actually tried to run the benchmark with the LUT in both cases, but after your question I double-checked and found a bug in the refit code which passed a nullptr in case of the GPU track model... :) I just reran with LUT in all cases:

• NVIDIA 3090 GPU Track Model: 66.341 us
• NVIDIA 3090 TrackParCov: 96.374 us
• AMD Rome 64 core 4 GHz GPU Track Model: 196.900 us
• AMD Rome 64 core 4 GHz TrackParCov: 221.634

[shahor02]

Still on mat.correction: on the CPU this will use by default MatCorrType matCorr = MatCorrType::USEMatCorrTGeo, i.e. slow TGeo query (although, in case of TPC volume it should be faster than average TGeo query).
To avoid this, I can change the default matCorr to USEMatCorrLUT, with fall-back to using TGeo in

AliceO2/Detectors/Base/src/Propagator.cxx

Line 516 in 4c65d47

GPUd() MatBudget Propagator::getMatBudget(Propagator::MatCorrType corrType, const math_utils::Point3D<float>& p0, const math_utils::Point3D<float>& p1) const

in case the mMatLUT was not set.

[davidrohr]

@shahor02 : I am quite sure I was not using TGeo but the matLut, because my benchmark is not even aware of TGeo. I have added the forced switch here: https://github.com/AliceO2Group/AliceO2/blob/22cdc82f69322f0710d549944740667d0d3130dd/Detectors/Base/src/Propagator.cxx#L521
So for GPU and for my standalone tool it will always use the matLut.
Perhaps I should throw an exception if the setting points to TGeo, or do you think it is OK as is?
In general, I think it makes sense to use matLut by default.

[shahor02]

OK, if for the benchmark you have used GPUCA_... also for CPU / TrackParCov case, then you have used matLUT.
I will change default to USEMatCorrLUT and add a fall-back to TGeo in getMatBudget (respecting your ifdef).