gpuKalmanFitter

A R&D repository based on A Common Tracking Software repository. It simplifies and transcribes the KalmanFitter in the ACTS repository to make it working on heterogeous computing.

Code dependency

• GCC compiler (min version 7.5.0)
• Nvidia CUDA (min version 10.2.89)
• Eigen library
• CERN ROOT

The gpu-kf-noRoot branch contains the master code without the ROOT dependency.
All the dependecies can be installed through Spack package manager.

Build the code

git clone https://github.com/XiaocongAi/gpuKalmanFitter.git
cd gpuKalmanFitter
cmake -S . -B <build_directory>
cmake --build <build_directory> <options>
Invoke the executable:
./INSTALL/bin/KalmanFitterGPUTest -d gpu -t 1000 -g 5120x1 -b 8x8x1 -o 0

Running the executable in a container

A singularity container (1.04GB) with all the dependecies is made available to download either

• directly from the cloud website, or
• through singularity call: singularity pull library://hpc-uhh/default/gpu-kf:v1.0

Currently, it runs the executable from tag v2.0-noRoot.

Invocation examples

To check the runtime options for the executable:
singularity run --nv gpu-kf_v1.0.sif --help

To run the fitting for 10,000 tracks on your available Nvidia GPU, with default parameters:
singularity run --nv gpu-kf_v1.0.sif -d gpu -t 10000

To run the fitting for 10,000 tracks on the CPU instead of the GPU:
(Note that a CUDA driver and a CUDA runtime must be accessible for the executable to run!)
singularity run --nv gpu-kf_v1.0.sif -d cpu -t 10000 -a Intel_i6-5218

To run the fitting for 10,000 tracks on multiple GPUs (if available):
(Note that the implementation will distribute the workload among all reachable GPUs)
singularity run --nv gpu-kf_v1.0.sif -d gpu -t 10000 -u 1

Developing the code

Install the code dependecies listed below, fork a branch from master and submit a pull request to merge your changes at the end.