The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Workshop:MCHPC’21: Workshop on Memory Centric High Performance Computing
Authors: Christian Trott (Sandia National Laboratories)
Abstract: To make memory-centric compute devices impactful and sustainable for a wide range of customers in HPC and beyond, we need to find ways of leveraging their capabilities in language standards. One of the prime targets for such efforts is the ISO C++ standard, which has been the language of choice for HPC vendors to implement programming models for accelerators, such as CUDA, HIP and SYCL. This talk will discuss existing and upcoming capabilities in the C++ standard like std::mdspan and std::linalg, which enable memory-centric application design. Based on concepts popularized in the Kokkos ecosystem for performance portability, these new features allow the design of algorithms that are memory location and memory layout aware, that leverage advanced memory access capabilities, and that provide customization points to plug special hardware-specific functionality into C++ code without relying on non-standard APIs, such as intrinsics and vendor-specific libraries.
Back to MCHPC’21: Workshop on Memory Centric High Performance Computing Archive Listing