About me

Thibault Cimic

2026-10-25 dim. 00:00

How to Use

  • Press 'o' for navigation map and 'o' again to quit navigation map
  • Press arrows keys to navigate
  • Press space to follow the indented way

Studies (2014-2018)

Master (2017-2018)

From LJLL at Sorbonne Université, specialisation in modelling focused on :

  • Functional analysis :
    • Sobolev/ Hilbert analysis
  • PDEs modelisation :
    • Finite differences/volumes/elements methods
    • Discontinuous Galerkin methods
  • Optimal control
  • Shape optimization

Before that (2014-2017)

  • First year of master :
    • Large linear systems in C++
    • Finite element method in python
    • PDE analysis
    • Statistics
  • Licence :
    • ODE in python
    • Graph theory
    • Topology theory
    • Group theory

INRIA, Alpines (2018-2021)

website

Enlarged Conjugate Gradient (ECG)

  • Resume :

    ECG(k) is a method for solving (SPD) linear systems as if one have deflated the smallest eigenvalue (without constructing any preconditioner).

  • Build \( (K_k(A,b) \subset) EK_k(A,b)\) with :
\begin{equation} T_m : \mathbb{R}^n \rightarrow \mathbb{R}^{n \times m},\\ \forall x\in \mathbb{R}^n, T_m(x) \mathbb{1}_m = x \end{equation}
  • Build \( x_k \) with a ~CG on this \(EK_k(A,b)\)
  • Converges as (p.44) :
\begin{equation} \|x_l - x_*\|_A \leq c \left(\frac{\sqrt{\lambda_{max}} - \sqrt{\lambda_k}}{\sqrt{\lambda_{max}} + \sqrt{\lambda_k}}\right)^{l} \end{equation}

Map-Making problem in CMB data analysis

  • Resume :
    • Astrophysic :

      One wants to build a high resolution map of the sky of properties (temperature, polarization) of the first lights in the universe (some 13.5 billions years ago).

    • Maths :

      Build a fully algebraic 2-lvl preconditioner for a specific form of linear systems to bound the spectrum and reduce iterations.

Ones want to solve \( A x = b \) with :

\begin{equation} A := P^T N P = \left( \sum_{i=1}^B P_{i}^T N_i P_i \right), b := P^T N d \end{equation}
  • Build \(Z_i = Span\left((v_{i,l})_{l=1...k}|\lambda_{i,k} \leq \frac{1}{\tau}\right)\) where :
\begin{equation} N_i v_{i,l} = \lambda_{i,l} D_i v_{i,l} \end{equation}

  • Deflate \(Z = \underset{i=1}{\overset{B}{\oplus}} P_i^T M_i Z_i =: Span(Z_0) \) with :
\begin{equation} M^{-1} := Z_0 \left( Z_0^T A Z_0 \right)^{-1} Z_0^T \end{equation}
  • End up with :
\begin{equation} \Lambda \left( M^{-1}A \right) \subset \left[C_l(\tau), C_u \right] \end{equation}

CEA, LGL (2023-2024)

CIVA's website

Distribution of Ondo

  • Ondo : C++ FE solver with high order, spectral polynomials
    • Macro-mesh with Macro-Elements (ME)
    • Micro-mesh inside each ME simulating one physic

Ondo_2x2ME.png

  • OndoMPI :
    • Design the distribution from scratch
    • Build the necessary structures and modify existing ones
    • Redesign construction algorithms and iterations
  • Goals :
    • Distribute the memory load
    • Scale to a few hundreds procs
    • Start the load balancing with Metis
  • Constraints :
    • Use blocking communication
    • Have no collectives communications

Mem_HIEME3D_44lin.png

StrongSca_OndoMPI.png

CEA, L_Sim & Maison de la Simulation (2025-now)

  • 2-year contract, jointly financed by :
  • Year 1 : a Kokkos backend for the Poisson solver of BigDFT (finished)
  • Year 2 (ongoing) : easing the adoption of Kokkos in scientific Fortran codebases

BigDFT and its Poisson solver

  • BigDFT : quantum chemistry code based on Density Functional Theory
    • Fortran, 500k to 1M lines, about 20 developers
  • psolver : generalized Poisson solver, computational heart of BigDFT
    • Solves the Poisson equation with a 3D FFT
    • One backend per hardware : Fortran/OpenMP, SYCL, CUDA, …

PSolverKokkos

  • A new Kokkos backend for psolver, packaged as the library libpsolverkokkos
  • Two execution paths, KOKKOSCPU and KOKKOSGPU
  • One implementation for any CPU and any GPU (NVIDIA, AMD, Intel)

Fortran and C++ side by side

  • Fortran / C++ interoperability through an hourglass interface :
    • Fortran bind(C), a C layer, then C++ with Kokkos
  • Autotools (BigDFT) and CMake (the library) working together
  • Validated by 24 tests of the public API and 8 tests of internal kernels, and the same energies on every backend

Results

  • Compared with the native backends of psolver on 2 GPUs (A100, MI250X) and 2 CPUs (Xeon, EPYC)
  • Performance and performance portability measured against the native backends, details in the talk
  • Presented at the Kokkos Tea Time (21 October 2026), recorded on the HPSF YouTube channel :

Personal project

  • Git tutorial : a hands-on tutorial to learn Git from the ground up, for researchers and students.
  • KokkosECG : the Enlarged Conjugate Gradient solver on top of Kokkos, for CPUs and GPUs.
  • opencode fork : adds a Teach mode to the open source coding agent, to learn with an LLM.
    • A read-only agent : it explains and asks questions instead of editing the code.

Older toy projects in C++

  • Lina : Linear Algebra package in C++.
  • MPIP++ : MPIP++ is a C++ API for MPI.
  • BinaryContainers : BinaryContainers is a library to store C++ structures as contiguous blocks of memory.

Skill set resume

  • CEA, L_Sim & Maison de la Simulation :
    • Port a Fortran solver to Kokkos in C++, for CPUs and GPUs
    • Intranode CPU/GPU optimisation
    • Fortran/C++ interoperability
  • CEA, LGL :
    • Distribute a commercial finite element solver in C++/MPI

Earlier positions

  • INRIA :
    • Solving distributed linear systems in C/MPI
    • Solving large eigenvalue problems in C
    • Build fully algebraic preconditioner in C/MPI
  • Master :
    • Solving large linear systems in C++
    • Helmholtz with finite element in Python
    • Navier-Stokes with finite element in C++

Years of experience

  • Kokkos : 1.5 years
    • 1.5 years at CEA, L_Sim
  • C++ : 4 years
    • 1.5 years at CEA, L_Sim
    • 1.5 years in CIVA commercial software
    • 1 years in master
  • Git : 6 years
    • 1.5 years at CEA, L_Sim
    • 1.5 years at CEA, LGL
    • 3 years INRIA

Years of experience (continued)

  • MPI : 3.5 years
    • 1.5 years in CIVA commercial software
    • 2 years at INRIA on the B3DCMB project
  • C : 2 years
    • 2 years at INRIA on the B3DCMB project
  • Python : 1 year
    • 1 year at INRIA

Thank you !