Optimise fragment offset calculation for GPU batches

Also add an application that benchmarks fragment offset calculation

The speed of the mep_offsets function improves from 14.3 Hz to 648 Hz (45x speedup). 4 threads are now more than enough to calculate the offsets; 2 would probably still be fine. It seems that EB::get_padding( s, 1 << align ) is very slow.

requires Allen!1127 (merged)

FYI @gligorov @kaaricha

mentioned in merge request Allen!1127 (merged)

changed the description

mentioned in issue Allen#393 (closed)

added RTA label

Hi Roel, I pulled this branch and the one line change to Allen for the throughput tests and I am seeing an intermittent crash. I tried running gdb on a CPU dbg build, but didn't see it that time (unclear if this means it won't happen with that build or just didn't happen when I tried). I can observe the crash and then immediately run the same command and not have it. crash.txt

/ci-test --merge Allen!1127 (merged)

added ci-test-triggered label

• [2023-02-23 14:06] Validation started with lhcb-master-mr#7120
• [2023-02-28 19:01] Validation started with lhcb-master-mr#7183

added 1 commit

• ddbe96fe - Fix build warning in opt builds

Compare with previous version

added 1 commit

• ea82a329 - Fix build warning in opt builds

Compare with previous version

mentioned in issue Moore#530 (closed)

/ci-test Allen!1127 (merged)

assigned to @rmatev

resolved all threads

merged

mentioned in commit 20b1df16