The school will deal with the optimisation of calculations on different types of hardware (CPU, GPU), presenting their respective characteristics, architectures and bottlenecks. It will focus on generic optimisation methods applicable to all types of hardware, and on the different libraries, technologies and languages available to achieve the best possible performance. Ideally, the peak performance of the machine.
- Considered equipment : CPU, GPU.
- Considered languages: C++17, C++20, CUDA, Fortran, Julia, Rust,
- Considered libraries: Sycl, Eve.
- Considered compilers: G++, Clang++, nvc++, nvfortran.
All methods will be illustrated on simple examples, such as Hadamard products, reductions, barycenter calculations and matrix products, to be applied to a single problem: the simulation of a Gray Scott reaction. This problem is simple enough to be understood quickly and complex enough that compilers will have difficulty optimising it without help. Each method will have a simple version, using default options, and one or more advanced versions, which will allow their advantages and disadvantages to be discussed and quantified.
All the details are available on the school’s website (programme, registration, etc.).
Pre-registration is open until 19 May, with a limited number of places available: https://indico.in2p3.fr/event/29755...