S. Joube, H. Grasland, D. Chamont, and J. Falcou, “Kiwaku, a C++20 library for multidimensional arrays Application to ACTS tracking,” in 26th International Conference on Computing in High Energy & Nuclear Physics (CHEP 2023), Norfolk, United States, May 2023

S. Joube, H. Grasland, D. Chamont, and E. Brunet, “Comparing SYCL data transfer strategies for tracking use cases,” Journal of Physics: Conference Series, vol. 2438, no. 1, p. 012 018, Feb. 2023, issn: 1742-6596. doi: 10.1088/1742-6596/2438/1/012018.

Acts Parallelization Meeting
Online

CHEP 2023: 26th International Conference on Computing in High Energy & Nuclear Physics
Norfolk, United States

14e Journées Informatiques IN2P3/IRFU
Le Croisic, France

4ème Reprise, Sorbonne Université
Paris, France

Université de Reims Champagne–Ardenne, France

Bertinoro, Italy

I am currently in the final few months of my PhD in efficient heterogeneous computing at LISN and IJCLab. I daily use a desktop version of Debian Linux on my personal and professional computer.

I am a developer of Kiwaku, a C++20 and onward collection of high performance data containers for heterogeneous computing based on generative metaprogramming. It targets massive parallelization on CPU with TBB, vectorization with EVE and multiple GPU architectures thanks to SYCL, since SYCL makes portable performance across compute accelerators much easier. Kiwaku is designed with usability in mind, fully taking advantage of the improvements provided by the C++20 standard.

As far as I can remember, I have always been fascinated by big physics experiments, and deeply value the research made at CERN. I was therefore delighted to use ATLAS data and Acts code (a charged particle tracking software) to evaluate the usability and performance of Kiwaku, leading to a talk at the CHEP 2023 conference called “Kiwaku, a C++20 library for multidimensional arrays, application to Acts tracking”. I wrote the corresponding article, now accepted for publication on the CHEP proceedings “EPJ Web of Conferences”. I used Python to process benchmark data and write plot scripts.

I also contributed to the CERN Acts/traccc repository which is “A single-source tracking demonstrator on CPUs/GPUs”, an R&D project of Acts designed for heterogeneous programming on CPUs, multi-core CPUs with TBB and GPUs with SYCL and CUDA. In particular, I have adapted the last step of the tracking chain, the ambiguity resolution algorithm, from Acts to traccc, and plan to develop a SYCL version for GPUs.

Six-month Master's degree internship in heterogeneous computing using the SYCL C++17 open standard and two of its implementations: Intel oneAPI and hipSYCL (now AdaptiveCpp).

I conducted an in-depth evaluation of the usability and performance of various SYCL memory models for both a memory-bound microbenchmark and a real use-case: my GPU adaptation of the first step of the Acts/traccc tracking chain (a charged particle tracking software). I measured the impact of memory access patterns on GPUs and CPUs, mainly profiling code execution with Nvidia Nsight and Intel VTune. I developed efficient data structures to better understand compute accelerator performance characteristics such as cache effects, memory bandwidth (PCIe bus, GPU memory, threads local memory, caches) and the implications of memory layout.

This led to the publication of my article in the ACAT 2021 proceedings called “Comparing SYCL data transfer strategies for tracking use cases” in the “Journal of Physics: Conference Series”.

Please refer to my work experience called “PhD: GPU and Portable Performances - Heterogeneous Approaches and Applications”.

I learned a lot during my master’s degree, in various fields related to heterogeneous programming. Here are a few relevant topics I studied: kernel C programming, CPU and GPU architectures, hardware and software CPU pipelines, assembly, distributed synchronization (consensus in a network of unreliable or fallible processors), distributed load balancing with Scala and Java, Git underlying blockchain architecture, operating system virtualization, C++ concurrent programming, formal methods.

This taught me a general knowledge on object oriented programming, data science, network and project management.

A few relevant topics: Probability on a finite universe, random variables and laws, expectation and variance ; Matrix calculus and linear systems - vector spaces, matrices and linear applications - endomorphisms of an Euclidean space ; Common algebraic structures - topology of normed vector spaces ; Limits, derivability, continuity and asymptotic analysis for functions - primitives and linear differential equations - integration on any interval ; Vector functions - suites and series of functions - integer, numerical and vector series ; And a few basics in physics: electromagnetism - signal processing - quantum physics - mechanics - optics - statistical thermodynamics

Stopped due to serious family problems.