Algorithm Harmonization #5.0 SYCL CKF, main branch (2026.01.23.)

[krasznaa]

This is a first step in synchronizing the CKF algorithms. First just synchronizing the SYCL algorithm with the CUDA one.

I had to turn 3 CUDA-only kernels into generic device code. Introducing the following new functions:

• traccc::device::gather_best_tips_per_measurement: Since the kernel's payload technically depends on the used algebra type, I made it into a template function. The logic of the code didn't really change.
• traccc::device::gather_measurement_votes: Here it's really a quite literal transcription of traccc::cuda::kernels::gather_measurement_votes .
• traccc::device::update_tip_length_buffer: This last function however is a bit different from the current traccc::cuda::kernels::update_tip_length_buffer kernel. Instead of managing a "resizable vector buffer" by hand, I rather introduced a resizable vecmem::data::vector_buffer into the code, and modified the code to use vecmem::device_vector<T>::push_back(...) instead of managing a buffer and a size variable by hand.

This latter change also required a small change in vecmem::device::build_tracks. But that was really just a small change. (And since that kernel changed, I also had to tweak the Alpaka CKF in a single line to keep it compiling.)

My first local tests were successful. Though I'll do some further tests as well. But as long as the CI doesn't complain, I'd ask you @stephenswat to launch the usual CI performance tests as well.

Let me also tag @flg. I believe with this PR included we'll be able to run some reasonable full chain tests on our AMD devices. 🤞

[stephenswat]

My first local tests were successful. Though I'll do some further tests as well. But as long as the CI doesn't complain, I'd ask you @stephenswat to launch the usual CI performance tests as well.

These are running!

🤦 I knew that at one point I'd mistakenly add this god-forsaken file... 😭

FYI there is a local equivalent to .gitignore in every git repository in .git/info/exclude, so you could just put this file there.

[krasznaa]

@flg, hopefully 3a7dc91 will make the CUDA and SYCL codes be more on par with each other. Though I didn't do any serious performance tests.

The c42ce55 update should help the CUDA and SYCL codes both by a tiny amount. Maybe it will be measurable, to be seen.

[stephenswat]

@flg, hopefully 3a7dc91 will make the CUDA and SYCL codes be more on par with each other. Though I didn't do any serious performance tests.

The c42ce55 update should help the CUDA and SYCL codes both by a tiny amount. Maybe it will be measurable, to be seen.

Does this explain the suspicious number of tracks in the CI output?

Total number of found tracks went from 20044 to 50259 (+150.7%)

[krasznaa]

I was not expecting to get more tracks, no. 😦 If anything, I was fearing that we would get too few. 😕

Something for tomorrow to debug then...

[stephenswat]

I was not expecting to get more tracks, no. 😦 If anything, I was fearing that we would get too few. 😕

Something for tomorrow to debug then...

What I'm saying is that 3a7dc91 might already resolve that issue!

[krasznaa]

Ahh... Could you re-run the tests? It should indeed fix that issue. 🤔

[stephenswat]

Performance summary

Here is a summary of the performance effects of this PR:

Graphical

Tabular

Kernel	Reciprocal Throughput			Parallelism
	4cec30b	adcc228	Delta	4cec30b	adcc228
propagate_to_next_surface	7.80 ms	7.80 ms	-0.0%	3.46	3.46
find_tracks	1.72 ms	1.72 ms	-0.4%	1.84	1.84
ccl_kernel	827.56 μs	826.03 μs	-0.2%	1.37	1.37
count_doublets	819.31 μs	822.42 μs	0.4%	1.61	1.61
count_triplets	567.51 μs	568.74 μs	0.2%	1.02	1.02
find_doublets	535.50 μs	544.70 μs	1.7%	3.08	3.08
Thrust::sort	379.90 μs	379.57 μs	-0.1%	7.32	7.32
find_triplets	169.87 μs	169.88 μs	0.0%	1.31	1.32
build_tracks	125.11 μs	124.60 μs	-0.4%	3.72	3.71
select_seeds	53.81 μs	53.66 μs	-0.3%	1.34	1.34
populate_grid	24.00 μs	23.97 μs	-0.1%	1.22	1.22
remove_duplicates	23.23 μs	23.24 μs	0.1%	26.33	26.33
count_grid_capacities	22.14 μs	22.14 μs	-0.0%	1.22	1.22
fill_sorted_measurements	19.69 μs	19.76 μs	0.3%	1.13	1.13
update_triplet_weights	14.77 μs	14.78 μs	0.1%	1.27	1.27
apply_interaction	13.88 μs	13.87 μs	-0.1%	6.71	6.71
estimate_track_params	11.76 μs	11.80 μs	0.4%	2.69	2.69
fill_finding_propagation_sort_keys	8.80 μs	8.80 μs	-0.0%	7.67	7.67
form_spacepoints	8.31 μs	8.32 μs	0.2%	1.48	1.48
reduce_triplet_counts	5.63 μs	5.68 μs	0.9%	3.08	3.08
unknown	5.08 μs	5.08 μs	0.0%	4.26	4.26
fill_finding_duplicate_removal_sort_keys	1.57 μs	1.57 μs	-0.2%	38.03	38.11
Total	13.16 ms	13.17 ms	0.0%	2.99	2.99

Important

All metrics in this report are given as reciprocal throughput, not as wallclock runtime.

Note

This is an automated message produced upon the explicit request of a human being.

[stephenswat]

Physics performance summary

Here is a summary of the physics performance effects of this PR. Command used:

traccc_seeding_example_cuda --input-directory=/data/Acts/odd-simulations-20240506/geant4_ttbar_mu200 --digitization-file=geometries/odd/odd-digi-geometric-config.json --detector-file=geometries/odd/odd-detray_geometry_detray.json --grid-file=geometries/odd/odd-detray_surface_grids_detray.json --material-file=geometries/odd/odd-detray_material_detray.json --input-events=10 --use-acts-geom-source=on --check-performance --truth-finding-min-track-candidates=5 --truth-finding-min-pt=1.0 --truth-finding-min-z=-150 --truth-finding-max-z=150 --truth-finding-max-r=10 --seed-matching-ratio=0.99 --track-matching-ratio=0.5 --track-candidates-range=5:100 --seedfinder-vertex-range=-150:150 --max-num-tracks-per-measurement=1

Seeding performance

Total number of seeds went from 298341 to 298344 (+0.0%)

Seeding plots

Track finding performance

Total number of found tracks went from 20042 to 20043 (+0.0%)

Finding plots

Track fitting performance

Fitting plots

Seeding to track finding relative performance

Seeding to track finding plots

Note

This is an automated message produced on the explicit request of a human being.

[krasznaa]

Well, the number of tracks is okay at least... 😢

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[krasznaa]

It turns out that I misunderstood the code yesterday. 😦 c42ce55 was actually introducing a bug, and not fixing a performance issue.

With that removed, at least locally, I get back the expected efficiency. This was all a good exercise for trying out how I would make physics efficiency plots during local code development. 🤔

[stephenswat]

Physics performance summary

Here is a summary of the physics performance effects of this PR. Command used:

traccc_seeding_example_cuda --input-directory=/data/Acts/odd-simulations-20240506/geant4_ttbar_mu200 --digitization-file=geometries/odd/odd-digi-geometric-config.json --detector-file=geometries/odd/odd-detray_geometry_detray.json --grid-file=geometries/odd/odd-detray_surface_grids_detray.json --material-file=geometries/odd/odd-detray_material_detray.json --input-events=10 --use-acts-geom-source=on --check-performance --truth-finding-min-track-candidates=5 --truth-finding-min-pt=1.0 --truth-finding-min-z=-150 --truth-finding-max-z=150 --truth-finding-max-r=10 --seed-matching-ratio=0.99 --track-matching-ratio=0.5 --track-candidates-range=5:100 --seedfinder-vertex-range=-150:150 --max-num-tracks-per-measurement=1

Seeding performance

Total number of seeds went from 298341 to 298341 (+0.0%)

Seeding plots

Track finding performance

Total number of found tracks went from 20042 to 20044 (+0.0%)

Finding plots

Track fitting performance

Fitting plots

Seeding to track finding relative performance

Seeding to track finding plots

Note

This is an automated message produced on the explicit request of a human being.

[stephenswat]

Plots look good!