Alpaka inhomogeneous magnetic field #1202
Conversation
…nt Alpaka flavours
Expand on the number of Kalman fit status codes and track fit outcomes to make a more fine-grained distinction between different error types in the Kalman fitter possible. Takes up the improvements in acts-project#1177. Edit: The wire chamber Kalman fitter test had to be adapted, because more error states are recognized Edit2: Also fixed the condition of is_complete for the backward fit to point one position below the first state
If the navigation in the forward KF missed a surface, advance the track state iterator correctly (for example if the propagation config between CKF and KF is different or we are running truth fitting). This is only needed for navigators that are not "guided" or "direct". It also adds an option to the KF to do the full hole count, which is only possible if the propagation starts at the IP and continues through the whole detector, otherwise the first and last holes are missed and the hole count remains an underestimation. Since this will not help in increasing the fit accuracy, but eat more compute resources it will remain optional.
This commit switched the measurement IO code to optionally get the dimensionality from the detector description rather than from the measurement input data.
This commit reworks the covariance estimation logic to match that used in ACTS. It greatly increases the covariances of tracks, hopefully increasing the tracking efficiency.
The CUDA CI jobs are currently running out of disk space. This commit builds ACTS separately, as a release build, in order to save space.
This PR is a revival of acts-project#652, removing the obsolete Kokkos and Futhark code bases.
Did not bother with the host algorithm, because: - It's not actively used in any current workflow; - It works correctly without this optimization as well.
|
Think this is ready to merge. |
krasznaa
left a comment
krasznaa
left a comment
There was a problem hiding this comment.
The feature addition in general is great of course. Let's just arrange the code in a better way. 😉
| magnetic_field make_magnetic_field(const magnetic_field& bfield, | ||
| [[maybe_unused]] const queue& queue) { | ||
| #if defined(ALPAKA_ACC_GPU_CUDA_ENABLED) | ||
| return traccc::cuda::make_magnetic_field( |
There was a problem hiding this comment.
I don't like this. 😦 I don't want to make traccc::alpaka depend on traccc::cuda and traccc::sycl.
I would very much prefer duplicating the code from those instead. If you prefer, we could set up some CUDA, SYCL and HIP "utility code" outside of the main libraries. Let's say traccc::cuda_utils or similar. Which could hold utility code like this.
But I definitely don't want to also slow down the CI even further, by the Alpaka build getting even heavier. 🤔
There was a problem hiding this comment.
OK, I like putting these in sycl_utils and cuda_utils. I will fix that when I come back to this PR.
|
4 participants
Implementation of alpaka on-device magnetic field. This either calls the CUDA or SYCL versions, or a HIP implementation written here. In principle this could be factorised out as well.
Seems to give reasonable results, although not identical track parameters.
1)AMD HIP:
2)NVidia CUDA:
3)After some delay, Intel (with FP64 emulation):