SYCL based Lattice Boltzmann flow solver for Heterogeneous Computing using oneAPI
This whitepaper demonstrates oneAPI’s promise of functional and performance portability using a highly parallelizable Lattice Boltzmann Solver based on SYCL programming model in a heterogeneous platform based on Intel® CPU and GPU architectures.
This work uses the former approach, i.e. Data Parallel C++ with SYCL. While adoption of this programming model as the de-facto choice for implementing production codebases appears lucrative, a software development team might ask three pertinent questions:
- How easy is it to port existing C++ codes to SYCL?
- How much performance overhead is incurred by SYCL due to the abstractions, compared to C++ codes on CPUs?
- Finally, how much speedup does SYCL offer on GPUs?
In this paper, we delve into these three questions. We demonstrate how a physics solver can be ported from C++ to SYCL and report benchmarks comparing the performance between C++ and SYCL solvers on CPUs and GPUs.
Speaker(s)
Dr. Mahan Raj Banerjee - Senior Software Development Engineer at SankhyaSutra Labs Ltd.
Dr. Mahan has a Doctor of Philosophy (Ph.D.) from JNCASR. He holds an M.Tech from IIT Kanpur. His Thesis includes the flow topology around anisotropic particles in planar shearing flows and Discretization of stochastic partial differential equations and Electrohydrodynamics: Jets. He holds an academic Excellence Award (IIT Kanpur). His project at IIEST, Shibpur, Mechanical Engineering includes an Artificial Neural Network model to predict the quality of the outputs of a Pulsed MIG process.
