Demonstrator tracking chain for accelerators.
Category | Algorithms | CPU | CUDA | SYCL | Alpaka | Kokkos | Futhark |
---|---|---|---|---|---|---|---|
Clusterization | CCL / FastSv / etc. | β | β | β | π‘ | βͺ | β |
Measurement creation | β | β | β | π‘ | βͺ | β | |
Seeding | Spacepoint formation | β | β | β | π‘ | βͺ | βͺ |
Spacepoint binning | β | β | β | β | β | βͺ | |
Seed finding | β | β | β | β | βͺ | βͺ | |
Track param estimation | β | β | β | β | βͺ | βͺ | |
Track finding | Combinatorial KF | β | β | π‘ | π‘ | βͺ | βͺ |
Track fitting | KF | β | β | β | βͺ | βͺ | βͺ |
Ambiguity resolution | Greedy resolver | β | βͺ | βͺ | βͺ | βͺ | βͺ |
β : exists, π‘: work started, βͺ: work not started yet
The relations between datatypes and algorithms is given in the (approximately commutative) diagram shown below. Black lines indicate CPU algorithms, green lines indicate CUDA algorithms, blue lines indicate SYCL algorithms, and brown lines indicate Futhark algorithms. Solid algorithms are ready for use, dashed algorithms are in development or future goals. Data types for different heterogeneous platforms are contracted for legibility, and identities are hidden.
flowchart LR
subgraph clusterization [<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/clusterization/clusterization_algorithm.hpp'>Clusterization</a>]
direction TB
cell(Cells);
cluster(Clusters);
meas(Measurements);
end
subgraph trkfinding [Track Finding]
sp(Spacepoints);
bin(Bins);
seed(Seeds);
ptrack(Prototracks);
end
subgraph trkfitting [Track Fitting]
track(Track);
end
click cell href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/cell.hpp";
click cluster href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/cluster.hpp";
click meas href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/measurement.hpp";
click sp href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/spacepoint.hpp";
click seed href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/seed.hpp";
click ptrack href "https://github.com/acts-project/traccc/blob/main/core/include/traccc/edm/track_parameters.hpp";
%% CPU CCL algorithm
cell -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/clusterization/component_connection.hpp'>CCL</a>| cluster;
linkStyle 0 stroke: black;
%% SYCL CCL algorithm
cell -->|CCL| cluster;
linkStyle 1 stroke: blue;
%% CUDA CCL algorithm
cell -->|CCL| cluster;
linkStyle 2 stroke: green;
%% CPU clusterization
cluster -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/clusterization/measurement_creation.hpp'>Agg.</a>| meas;
linkStyle 3 stroke: black;
%% SYCL clusterization
cluster -->|Agg.| meas;
linkStyle 4 stroke: blue;
%% CUDA clusterization
cluster -->|Agg.| meas;
linkStyle 5 stroke: green;
%% CUDA CCA
cell -->|CCA| meas;
linkStyle 6 stroke: green;
%% CPU local to global
meas -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/clusterization/spacepoint_formation.hpp'>L2G</a>| sp;
linkStyle 7 stroke: black;
%% SYCL local to global
meas -->|L2G| sp;
linkStyle 8 stroke: blue;
%% CUDA local to global
meas -->|L2G| sp;
linkStyle 9 stroke: green;
%% CPU binning
sp -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/seeding/spacepoint_binning.hpp'>Binning</a>| bin;
linkStyle 10 stroke: black;
%% CUDA binning
sp -->|<a href='https://github.com/acts-project/traccc/blob/main/device/cuda/include/traccc/cuda/seeding/spacepoint_binning.hpp'>Binning</a>| bin;
linkStyle 11 stroke: green;
%% CPU seeding
bin -->|Seeding| seed;
linkStyle 12 stroke: black;
%% SYCL seeding
bin -->|<a href='https://github.com/acts-project/traccc/blob/main/device/sycl/include/traccc/sycl/seeding/seed_finding.hpp'>Seeding</a>| seed;
linkStyle 13 stroke: blue;
%% CUDA seeding
bin -->|<a href='https://github.com/acts-project/traccc/tree/main/device/cuda/include/traccc/cuda/seeding'>Seeding</a>| seed;
linkStyle 14 stroke: green;
%% CUDA binless seeding
sp -.->|Seeding| seed;
linkStyle 15 stroke: green;
%% CPU param est.
seed -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/seeding/track_params_estimation.hpp'>Param. Est.</a>| ptrack;
linkStyle 16 stroke: black;
%% CUDA param est.
seed -->|<a href='https://github.com/acts-project/traccc/blob/main/device/cuda/include/traccc/cuda/seeding/track_params_estimation.hpp'>Param. Est.</a>| ptrack;
linkStyle 17 stroke: green;
%% CPU CKF
ptrack -.->|CKF| track;
linkStyle 18 stroke: black;
%% CPU Kalman filter
track -->|<a href='https://github.com/acts-project/traccc/blob/main/core/include/traccc/fitting/fitting_algorithm.hpp'>Kalman filter</a>| track;
linkStyle 19 stroke: black;
%% CUDA Kalman filter
track -->|<a href='https://github.com/acts-project/traccc/blob/main/device/cuda/include/traccc/cuda/fitting/fitting_algorithm.hpp'>Kalman filter</a>| track;
linkStyle 20 stroke: green;
%% SYCL binning
sp -->|<a href='https://github.com/acts-project/traccc/blob/main/device/sycl/include/traccc/sycl/seeding/spacepoint_binning.hpp'>Binning</a>| bin;
linkStyle 21 stroke: blue;
%% SYCL track parameter est.
seed -->|<a href='https://github.com/acts-project/traccc/blob/main/device/sycl/include/traccc/sycl/seeding/track_params_estimation.hpp'>Param. Est.</a>| ptrack;
linkStyle 22 stroke: blue;
%% Futhark measurement creation
cell -->|<a href='https://github.com/acts-project/traccc/blob/main/device/futhark/src/measurement_creation.fut'>CCA</a>| meas;
linkStyle 23 stroke: brown;
%% Futhark spacepoint creation
meas -->|<a href='https://github.com/acts-project/traccc/blob/main/device/futhark/src/spacepoint_formation.fut'>L2G</a>| sp;
linkStyle 24 stroke: brown;
%% SYCL Kalman filter
track -->|<a href='https://github.com/acts-project/traccc/blob/main/device/sycl/include/traccc/sycl/fitting/fitting_algorithm.hpp'>Kalman filter</a>| track;
linkStyle 25 stroke: blue;
%% CUDA CKF
ptrack -.->|CKF| track;
linkStyle 26 stroke: green;
Please note that due to the complexity of this software and its build system, it may be somewhat fragile in the face of compiler version changes. The following are general guidelines for getting traccc to compile:
- The C++ compiler must support C++17
In addition, the following requirements hold when CUDA is enabled:
- The CUDA Toolkit version must be greater than major version 11
- The CUDA Toolkit must not be minor version 11.3 due to a bug in the front-end compiler of that version
- Ensure that the CUDA host compiler supports C++17 and is compatible with the
nvcc
compiler driver
The following table lists currently combinations of builds, compilers, and toolchains that are currently known to work (last updated 2022/01/24):
Build | OS | gcc | CUDA | comment |
---|---|---|---|---|
CUDA | Ubuntu 20.04 | 9.3.0 | 11.5 | runs on CI |
Clone the repository and setup the data directory.
git clone git@github.com:acts-project/traccc.git
cd traccc
./data/traccc_data_get_files.sh
cmake -S . -B <build_directory>
cmake --build <build_directory> <options>
Option | Description |
---|---|
TRACCC_BUILD_CUDA | Build the CUDA sources included in traccc |
TRACCC_BUILD_SYCL | Build the SYCL sources included in traccc |
TRACCC_BUILD_TESTING | Build the (unit) tests of traccc |
TRACCC_BUILD_EXAMPLES | Build the examples of traccc |
TRACCC_USE_SYSTEM_VECMEM | Pick up an existing installation of VecMem from the build environment |
TRACCC_USE_SYSTEM_EIGEN3 | Pick up an existing installation of Eigen3 from the build environment |
TRACCC_USE_SYSTEM_ALGEBRA_PLUGINS | Pick up an existing installation of Algebra Plugins from the build environment |
TRACCC_USE_SYSTEM_DFELIBS | Pick up an existing installation of dfelibs from the build environment |
TRACCC_USE_SYSTEM_DETRAY | Pick up an existing installation of Detray from the build environment |
TRACCC_USE_SYSTEM_ACTS | Pick up an existing installation of Acts from the build environment |
TRACCC_USE_SYSTEM_GOOGLETEST | Pick up an existing installation of GoogleTest from the build environment |
TRACCC_USE_ROOT | Build physics performance analysis code using an existing installation of ROOT from the build environment |
<build_directory>/bin/traccc_seq_example --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --input-events=10
<build_directory>/bin/traccc_throughput_mt --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --cold-run-events=100 --processed-events=1000 --threads=1
- Users can generate CUDA examples by adding
-DTRACCC_BUILD_CUDA=ON
to cmake options
<build_directory>/bin/traccc_seq_example_cuda --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --input--events=10 --run-cpu=1
<build_directory>/bin/traccc_throughput_mt_cuda --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --cold-run-events=100 --processed-events=1000 --threads=1
- Users can generate SYCL examples by adding
-DTRACCC_BUILD_SYCL=ON
to cmake options
<build_directory>/bin/traccc_seq_example_sycl --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --input--events=10 --run-cpu=1
<build_directory>/bin/traccc_throughput_mt_sycl --detector-file=tml_detector/trackml-detector.csv --digitization-config-file=tml_detector/default-geometric-config-generic.json --input-directory=tml_pixels/ --cold-run-events=100 --processed-events=1000 --threads=1
Users can generate muon-like particle simulation data with the pre-built detray geometries:
# Generate telescope geometry data
<build_directory>/bin/traccc_simulate_telescope --gen-vertex-xyz-mm=0:0:0 --gen-vertex-xyz-std-mm=0:0:0 --gen-mom-gev=100:100 --gen-phi-degree=0:0 --gen-events=10 --gen-nparticles=2000 --output-directory=detray_simulation/telescope_detector/n_particles_2000/ --gen-eta=1:3
# Generate toy geometry data
<build_directory>/bin/traccc_simulate_toy_detector --gen-vertex-xyz-mm=0:0:0 --gen-vertex-xyz-std-mm=0:0:0 --gen-mom-gev=100:100 --gen-phi-degree=0:360 --gen-events=10 --gen-nparticles=2000 --output-directory=detray_simulation/toy_detector/n_particles_2000/ --gen-eta=-3:3 --constraint-step-size-mm=1 --search-window 3:3
# Generate drift chamber data
<build_directory>/bin/traccc_simulate_wire_chamber --gen-vertex-xyz-mm=0:0:0 --gen-vertex-xyz-std-mm=0:0:0 --gen-mom-gev=2:2 --gen-phi-degree=0:360 --gen-events=10 --gen-nparticles=100 --output-directory=detray_simulation/wire_chamber/n_particles_100/ --gen-eta=-1:1 --constraint-step-size-mm=1 --search-window 3:3
The simulation will also generate the detector json files (geometry, material and surface_grid) in the current directory. It is user's responsibility to move them to an appropriate place (e.g. <detector_directory>
) and match them to the input file arguments of reconstruction chains.
If users have a geometry json file already, it is also possible to run simulation with traccc_simulate
application
# Given that users have a geometry json file
<build_directory>/bin/traccc_simulate --output-directory=<output-directory> --detector-file=<geometry_file> --material-file=<material-file> --grid-file=<grid-file> --event=10 --constraint-step-size-mm=1
There are three types of partial reconstruction chain users can operate: seeding_example
, truth_finding_example
, and truth_fitting_example
where their algorithm coverages are shown in the table below. Each of them starts from truth measurements, truth seeds, and truth tracks, respectively.
Category | Clusterization | Seeding | Track finding | Track fitting |
---|---|---|---|---|
seeding_example |
β | β | β | |
truth_finding_example |
β | β | ||
truth_fitting_example |
β |
The dirft chamber will not produce meaningful results with seeding_example
as the current seeding algorithm is only designed for 2D measurement objects. Truth finding works OK in general but the combinatoric explosion can occur for a few unlucky events, leading to poor pull value distributions. The followings are example commands:
# Run cuda seeding example for toy geometry
<build_directory>/bin/traccc_seeding_example_cuda --input-directory=detray_simulation/toy_detector/n_particles_2000/ --check-performance --detector-file=<detector_directory>/toy_detector_geometry.json --material-file=<detector_directory>/toy_detector_homogeneous_material.json --grid-file=<detector_directory>/toy_detector_surface_grids.json --input-events=1 --track-candidates-range=3:30 --constraint-step-size-mm=1000 --run-cpu=1 --search-window 3:3
# Run cuda truth finding example for toy geometry
<build_directory>/bin/traccc_truth_finding_example_cuda --input-directory=detray_simulation/toy_detector/n_particles_2000/ --check-performance --detector-file=<detector_directory>/toy_detector_geometry.json --material-file=<detector_directory>/toy_detector_homogeneous_material.json --grid-file=<detector_directory>/toy_detector_surface_grids.json --input-events=1 --track-candidates-range=3:30 --constraint-step-size-mm=1000 --run-cpu=1 --search-window 3:3
# Run cuda truth finding example for drift chamber
<build_directory>/bin/traccc_truth_finding_example_cuda --input-directory=detray_simulation/wire_chamber/n_particles_100/ --check-performance --detector-file=<detector_directory>/wire_chamber_geometry.json --material-file=<detector_directory>/wire_chamber_homogeneous_material.json --grid-file=<detector_directory>/wire_chamber_surface_grids.json --input-events=10 --track-candidates-range=6:30 --constraint-step-size-mm=1 --run-cpu=1 --search-window 3:3
# Run cpu truth fitting example for drift chamber
<build_directory>/bin/traccc_truth_fitting_example --input-directory=detray_simulation/wire_chamber/n_particles_2000_100GeV/ --check-performance --detector-file=<detector_directory>/wire_chamber_geometry.json --material-file=<detector_directory>/wire_chamber_homogeneous_material.json --grid-file=<detector_directory>/wire_chamber_surface_grids.json --input-events=10 --constraint-step-size-mm=1 --search-window 3:3
Users can open the performance root files (with --check-performance=true
) and draw the histograms.
$ root -l performance_track_finding.root
root [0]
Attaching file performance_track_finding.root as _file0...
(TFile *) 0x3871910
root [1] finding_trackeff_vs_eta->Draw()
The following are potentially useful instructions for troubleshooting various problems with your build:
You may experience errors being issued about standard library features, for example:
/usr/include/c++/11/bits/std_function.h:435:145: note: β_ArgTypesβ
/usr/include/c++/11/bits/std_function.h:530:146: error: parameter packs not expanded with β...β:
530 | operator=(_Functor&& __f)
In this case, your nvcc
host compiler is most likely incompatible with your
CUDA toolkit. Consider installing a supported version and selecting it through
the CUDAHOSTCXX
environment variable at build-time.