News from GPU parameter tuning

• Tested Radeon VII on dev machine
• Used current default parameters of MI50 which are optimal for lhc24ar_raw dataset
• Noticeably slower than MI50 32GBs for the first three events
• The reported times have been taken with the following command:
  ROCR_VISIBLE_DEVICES=0 numactl --membind 0 --cpunodebind 0 ./ca -e lhc24ar_raw --gpuType HIP --memSize 15000000000 --inputMemory 6000000000 --outputMemory 10000000000
--sync -s [event_index] -n [event_index + 1] --runs 10 --PROCdoublePipeline --RTCenable --RTCcacheOutput
  
  

• Bug present in current O2 when an RDNA gpu or a gpu equal or newer than a MI210 tries to compile. Since specific parameters are not present yet, static_assert(GPUCA_PAR_AMD_EUS_PER_CU > 0); will fail since the parameter will be defaulted to 0. I will tune the MI210 and open a PR with the parameters for both archs which will fix this. Quickest temporary f ix is to fallback to MI100 in case these archs are detected