Repository navigation
GPUCommonMath: double functions, 64-bit bit counts, atomic load/store… - #15986
Merged
Merged
Conversation
davidrohr
reviewed
Oct 20, 2026
davidrohr
left a comment
davidrohr
left a comment
Collaborator
There was a problem hiding this comment.
Why did you change all the GPUd() to GPUhd(). GPUhd() means it is executed by the host code compiled with CUDA, so that should be only the code that drives the devices?
Does that code really need the GPUCommonMath functions? IMHO in that case one can use std:: directly.
Named double functions for GPUCommonMath::Sqrt,Sin,Cos,Tan,ASin,ACos,ATan,ATan2,Log,Copysign. Clz64, Ctz64, Popcount64: 64-bit counterparts of Clz, Ctz and Popcount on the same builtins and CUDA/HIP intrinsics. Named rather than uint64_t overloads, which would change existing calls with 64-bit arguments and make calls with int arguments ambiguous. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Felix Schlepper <felix.schlepper@cern.ch>
Collaborator
Author
|
Hi @davidrohr, you right of course. I had declared functors with GPUhd() in code linked to O2 using CAMath. My compilation then failed without declaring the functions also GPUhd(), so I had 'bug' in my code... sorry. I think I never understood the difference between GPUhd() and GPUd(). Also I had the question about the deterministic mode, currently it forces some computations to be made in double precision and then downcasts to float, but this trick does not work if the computation is supposed to be in double see for example |
Collaborator
|
Hm, deterministic mode in double is a good question. IEEE double precision does not specify log, trigonimetric functions etc., i.e. they do not need to be 100% precise. |
matthias-kleiner
added a commit
to matthias-kleiner/AliceO2
that referenced
this pull request
Oct 20, 2026
Follow-up to AliceO2Group#15986. TrackParametrization, TrackParametrizationWithError and TrackUtils call the float-only gpu::CAMath functions (Sqrt, ASin, ATan2, ...), so TrackParD / PropagatorD is effectively float. Compared with a std:: double reference, propagateParamTo(x, b[3]) from x = 85 to 250 cm is off by up to 74 µm after 100 steps (pt 0.3 GeV, small transverse field), and finite-difference derivatives through it are noisy. This affects the barrel alignment (Detectors/Align) and the ITS3 alignment; no reconstruction workflow uses the double types. Call math_utils::detail::{sqrt, asin, ...}<value_t> instead, as Propagator.cxx already does: for double these dispatch to the double functions, for float to the same functions as before. Add GPUCommonMath::Expd and math_utils::detail::exp for the remaining Exp calls. Float results are bit-identical; TrackParD now agrees exactly with the double reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
matthias-kleiner
added a commit
to matthias-kleiner/AliceO2
that referenced
this pull request
Oct 20, 2026
Follow-up to AliceO2Group#15986. TrackParametrization, TrackParametrizationWithError and TrackUtils call the float-only gpu::CAMath functions (Sqrt, ASin, ATan2, ...), so TrackParD / PropagatorD is effectively float. Compared with a std:: double reference, propagateParamTo(x, b[3]) from x = 85 to 250 cm in 100 steps is off by 0.4 to 74 µm, depending on pt and the field (toy fields with a transverse component, no material), and finite-difference derivatives through it are noisy. This affects the barrel alignment (Detectors/Align) and the ITS3 alignment; no reconstruction workflow uses the double types. Call math_utils::detail::{sqrt, asin, ...}<value_t> instead, as Propagator.cxx already does: for double these dispatch to the double functions, for float to the same functions as before. Add GPUCommonMath::Expd and math_utils::detail::exp for the remaining Exp calls. Float results are bit-identical; TrackParD now agrees exactly with the double reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
matthias-kleiner
added a commit
to matthias-kleiner/AliceO2
that referenced
this pull request
Oct 20, 2026
Follow-up to AliceO2Group#15986. TrackParametrization, TrackParametrizationWithError and TrackUtils call the float-only gpu::CAMath functions (Sqrt, ASin, ATan2, ...), so TrackParD / PropagatorD is effectively float. Compared with a std:: double reference, propagateParamTo(x, b[3]) from x = 85 to 250 cm in 100 steps is off by 0.4 to 74 µm, depending on pt and the field (toy fields with a transverse component, no material), and finite-difference derivatives through it are noisy. This affects the barrel alignment (Detectors/Align) and the ITS3 alignment; no reconstruction workflow uses the double types. Call math_utils::detail::{sqrt, asin, ...}<value_t> instead, as Propagator.cxx already does: for double these dispatch to the double functions, for float to the same functions as before. Add GPUCommonMath::Expd and math_utils::detail::exp for the remaining Exp calls. This restores the math_utils::detail calls of cd9835a (switched back to CAMath in bf65d3c) and extends them to TrackParametrizationWithError and TrackUtils. Float results are bit-identical; TrackParD now agrees exactly with the double reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
shahor02
pushed a commit
that referenced
this pull request
Oct 21, 2026
Follow-up to #15986. TrackParametrization, TrackParametrizationWithError and TrackUtils call the float-only gpu::CAMath functions (Sqrt, ASin, ATan2, ...), so TrackParD / PropagatorD is effectively float. Compared with a std:: double reference, propagateParamTo(x, b[3]) from x = 85 to 250 cm in 100 steps is off by 0.4 to 74 µm, depending on pt and the field (toy fields with a transverse component, no material), and finite-difference derivatives through it are noisy. This affects the barrel alignment (Detectors/Align) and the ITS3 alignment; no reconstruction workflow uses the double types. Call math_utils::detail::{sqrt, asin, ...}<value_t> instead, as Propagator.cxx already does: for double these dispatch to the double functions, for float to the same functions as before. Add GPUCommonMath::Expd and math_utils::detail::exp for the remaining Exp calls. This restores the math_utils::detail calls of cd9835a (switched back to CAMath in bf65d3c) and extends them to TrackParametrizationWithError and TrackUtils. Float results are bit-identical; TrackParD now agrees exactly with the double reference. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
…, host-device calls
Named double functions for GPUCommonMath::Sqrt,Sin,Cos,Tan,ASin,ACos,ATan,ATan2,Log,Copysign.
Clz64, Ctz64, Popcount64: 64-bit counterparts of Clz, Ctz and Popcount on the same builtins and CUDA/HIP intrinsics.
And AtomicLoad/Store.