CSC4700-GPU Programming, the C++ way

About this Video

• Video Title: CSC4700-GPU Programming, the C++ way
• Channel: Stellar at LSU
• Speakers: The instructor's name is not explicitly mentioned in the transcript.
• Duration: 01:14:21

Introduction

This lecture introduces GPU programming using C++ and the Kokkos library. The primary goal is to demonstrate how to write portable C++ code that can run efficiently on both CPUs and GPUs, addressing the challenges of diverse HPC architectures and programming models. The lecture covers Kokkos's API, parallel patterns (algorithms), views for data management, and integration with HPX for asynchronous operations.

Key Takeaways

1. GPU Architecture Differences: GPUs have a fundamentally different architecture than CPUs, employing warps of processing units that execute instructions in lock-step. This necessitates different programming models.
2. Kokkos Library: Kokkos is a C++ library designed for data-parallel programming, offering a single API that translates to various backends (architectures and programming models) at compile time. It provides portability and performance across different platforms.
3. Kokkos Views: Kokkos views are lightweight representations of arrays that enable efficient data management and transfer between CPU and GPU memory spaces. They allow shallow copies, avoiding the overhead of copying large datasets.
4. Execution and Memory Spaces: Kokkos manages execution spaces (where code runs—CPU, GPU, etc.) and memory spaces (where data resides). The lecture demonstrates how to use these features, including mirroring data across spaces for efficient GPU computation.
5. HPX Integration: The lecture shows how to integrate Kokkos with HPX, leveraging HPX's parallel algorithms and asynchronous capabilities to further enhance performance and code efficiency, particularly for large datasets requiring data transfer between CPU and GPU.