Section: New Software and Platforms

BOAST: Metaprogramming of Computing Kernels

BOAST aims at providing a framework to metaprogram, benchmark and validate computing kernels. BOAST is a programming framework dedicated to code generation and autotuning. This software allows the transformation from code written in the BOAST DSL to classical HPC targets like FORTRAN, C, OpenMP, OpenCL or CUDA. It also enables the meta-programming of optimization that can be (de)activated when needed. BOAST can also benchmark and do non regression tests on the generated kernels. This approach gives, both, performance gains and improved performance portability.

BOAST can be dowloaded at this address https://forge.imag.fr/projects/boast/.

BOAST was already used to generate and optimize the computing kernels of three scientific applications:

• BigDFT: A massively parallel electronic structure code using wavelet basis set.

• SPECFEM: Computational Infrastructure for Geodynamics.

• Gysela: Fusion plasma simulations.

BOAST is currently used in the context of the European H2020/HPC4E project. The computing kernels of two scientific applications are currently studied with BOAST:

• Alya: Large Scale Computational Mechanics.

• Hou10ni: Solutions to accoustics wave propagation problems. This code is developed by the Magique3D Inria team (Pau, Julien Diaz).

Frédéric Desprez presented BOAST at the CSCD workshop http://www.netlib.org/utk/people/JackDongarra/CCDSC-2016/ in October 2016. After this workshop, a paper was submited at the Internationaj Journal on High Performance Computing Applications (IJHPCA).

BOAST was also used in the Bulldog project during the last CERMACS summer school http://smai.emath.fr/cemracs/cemracs16/ in July 2016. A joint paper with CEA researchers from Cadarache and Maison de la Simulation was also submitted to present the results of the Bulldog project.