Follow up PR for PR #37695 ECAL Phase 2 weights method amplitude reconstruction on GPU

[ChrisSandever]

PR #37695 ECAL Phase 2 weights method amplitude reconstruction on GPU, added a producer to reconstruct uncalibrated RecHits a converter and modified certain other converters. It completed the reconstruction on a GPU for Phase-2.

This PR will need to be followed up by a separate PR to modify the added modules.
The planned modifications are as follows:
-To change the modules (EcalPhase2DigiToGPUProducer, EcalUncalibRecHitPhase2WeightsProducerGPU) such that they do not use the external_work template.
-To modify the kernel to be able to run with any number of threads and blocks, by allowing the kernel to reconstruct as many channels per thread as is required.
-To use the single buffer data format for use on the GPU device.

[cmsbuild]

A new Issue was created by @ChrisSandever Christopher Rhys Sandever.

@Dr15Jones, @perrotta, @dpiparo, @makortel, @smuzaffar, @qliphy can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[fwyzard]

assign ecal-dpg

[fwyzard]

assign heterogeneous

[cmsbuild]

New categories assigned: heterogeneous,ecal-dpg

@fwyzard,@simonepigazzini,@makortel,@jainshilpi,@thomreis you have been requested to review this Pull request/Issue and eventually sign? Thanks

[thomreis]

Hi @fwyzard , working on this issue at the moment. The first two tasks are done. For the single buffer data format I remember you mentioning that a generic format is being worked on. What is the status of the development there? I found #37952 but it is not merged yet.

[fwyzard]

hi @thomreis

For the single buffer data format I remember you mentioning that a generic format is being worked on.

We have it implemented for the Alpaka port, see for example my slides from the GPU meeting last Monday or the Core Software meeting last Tuesday (https://indico.cern.ch/event/1192819/contributions/5014619/attachments/2496301/4287337/Alpaka%20GPU%20framework.pdf) (yes, we need better documentation).

If you would like to give it a try already for the CUDA version of the modules, we can make that version available as well, currently it's only used in a private branch.

[thomreis]

Thanks @fwyzard , in that case I will integrate the first two tasks already. We will probably move the code to Alpaka anyway so the single buffer can be done then.

[thomreis]

#39195 takes care of the first two tasks.

• To change the modules (EcalPhase2DigiToGPUProducer, EcalUncalibRecHitPhase2WeightsProducerGPU) such that they do not use the external_work template.
• To modify the kernel to be able to run with any number of threads and blocks, by allowing the kernel to reconstruct as many channels per thread as is required.
• To use the single buffer data format for use on the GPU device.