Jean-Baptiste Keck

PhD in Applied Mathematics

Co-founder & CTO of RayMapr

About me

I am a computational scientist working on implementing numerical methods efficiently on accelerators, from multicore CPUs to NVIDIA, AMD and Intel GPUs, up to distributed hybrid CPU/GPU clusters. The main technologies I use are CUDA, SYCL, OpenCL, Vulkan (Slang), Python, C++ and Rust, with a strong taste for symbolic code generation and runtime compilation. Most of the codes I develop are available as open source software on my github and gitlab accounts.

I graduated in 2015 from the Ensimag (Grenoble INP institute of engineering at university Grenoble-Alpes). I then completed my PhD by the end of 2019 at the Jean Kuntzmann laboratory under the supervision of Georges-Henri Cottet, Christophe Picard and Iraj Mortazavi on the subject of numerical modelling and high performance computing for sediment flows.

I completed my postdoctoral research under the supervision of Boris Thibert and Quentin Mérigot in the field of computational geometry. Together, we are now founding RayMapr, a deeptech startup where I lead the technical development. RayMapr computes the freeform lenses and mirrors that send the light of an LED exactly where it is needed, sharp cut-off included, using optimal transport and GPU computing. The resulting surfaces are described in NURBS and go straight to CAD, machining or injection moulding.

Despite now working in the field of nonimaging optics, I still occasionally contribute to HySoP (Hybrid Simulation with Particles), an open source library designed for high performance distributed flow simulation targeting heterogeneous architectures, and to Rodin, a modern C++20 finite element library for shape and topology optimization. More recently, I got involved in the analysis and rendering of 3D subcellular imaging datasets.

On the open source side, I maintain a stack of GPU libraries that build on each other (all of them are detailed on the projects page):

  • symox, a fast computer algebra system written in Rust with a SymPy compatible Python API, generating code for CUDA, SYCL, Slang, PyTorch and JAX.
  • kernox, fused symbolic GPU reductions built on symox, a successor to PyKeOps, and remblais, a drop-in replacement for GeomLoss built on kernox (Sinkhorn divergences, kernel norms and optimal transport).
  • cusvna and its Vulkan port vusvna, rigorous interval and affine arithmetic on the GPU, used among others by cupmb, a GPU isosurface extraction library (marching cubes, marching squares, root finding).
  • CUDA and NVRTC ports of autodiff and Eigen, bringing automatic differentiation and dense linear algebra to runtime compiled kernels.

Quick access

  • A short overview of my research interests is available as slides (4 pages, 2MB).
  • My PhD manuscript is publically available here and here.
  • A short twelve page summary of my PhD is available here (in french, 2.4MB).
  • Self-hosted backups of my PhD are also available here: lowres (20MB), highres (31MB).

Publications