GPU

In this chapter, we will present the GPU support in AMReX. AMReX targets NVIDIA, AMD and Intel GPUs using their native vendor language and therefore requires CUDA, HIP/ROCm and SYCL, for NVIDIA, AMD and Intel GPUs, respectively. Users can also use OpenMP and/or OpenACC in their applications.

AMReX supports NVIDIA GPUs with compute capability >= 6 and CUDA >= 10. While HIP and SYCL compilers are in development in preparation for Frontier and Aurora, AMReX only supports the latest publicly released versions of those compilers.

For complete details of CUDA, HIP, SYCL, OpenMP and OpenACC languages, see their respective documentations.

Be aware, this documentation is currently focused on CUDA. HIP and SYCL documentation is forthcoming.

A number of tutorials can be found at Tutorials/GPU.

Contents:

• Overview of AMReX GPU Strategy
• Building GPU Support
• Gpu Namespace and Macros
• Memory Allocation
• GPU Safe Classes and Functions
• Kernel Launch
• Stream and Synchronization
• An Example of Migrating to GPU
• Assertions and Error Checking
• Particle Support
• Profiling with GPUs
• Performance Tips
• Inputs Parameters