Preparing Ginkgo for AMD GPUs – A Testimonial on Porting CUDA Code to HIP

Abstract

With AMD reinforcing their ambition in the scientific high performance computing ecosystem, we extend the hardware scope of the Ginkgo linear algebra package to feature a HIP backend for AMD GPUs. In this paper, we report and discuss the porting effort from CUDA, the extension of the HIP framework to add missing features such as cooperative groups, the performance price of compiling HIP code for AMD architectures, and the design of a library providing native backends for NVIDIA and AMD GPUs while minimizing code duplication by using a shared code base.

Single‐ and multi‐GPU computing on NVIDIA‐ and AMD‐based server platforms for solidification modeling application

“…In the past, software efforts primarily concentrated 32 on NVIDIA GPUs due to the mature CUDA development environment and the widespread adoption in the HPC community. However, with the launching of new AMD Instinct series GPUs and especially powerful Instinct MI200 accelerators with the CDNA2 architecture, 10 developers of scientific software have started porting their codes to AMD GPUs 5,33–38 using the HIP programming interface 32 . Among the main conclusions from these works, two deserve special attention.…”

“…The idea behind HIP is to increase the software and performance portability. The HIP provides an interface that can be used to access the functionality of both ROCm and CUDA 32 . Consequently, the codes written using the HIP programming environment can be executed on both AMD and NVIDIA GPUs.…”

Providing performance portable numerics for Intel GPUs

“…We note that the "cuda" and "hip" backends are very similar in kernel design, so we have "shared" kernels that are identical for the NVIDIA and AMD GPUs up to kernel configuration parameters. 5 Extending GINKGO 's scope to support Intel GPUs via the DPC++ language, we add the "dpcpp" backend containing corresponding kernels in DPC++.…”

“…We note that “reference” contains sequential CPU kernels designed to validate the correctness of the parallel algorithms and for the unit tests realized using the Googletest framework. We note that the “cuda” and “hip” backends are very similar in kernel design, so we have “shared” kernels that are identical for the NVIDIA and AMD GPUs up to kernel configuration parameters 5 . Extending Ginkgo 's scope to support Intel GPUs via the DPC++ language, we add the “dpcpp” backend containing corresponding kernels in DPC++.…”

GPU-accelerated molecular dynamics: State-of-art software performance and porting from Nvidia CUDA to AMD HIP

“…They found HIP to be best performing high-level framework for AMD devices and confirmed that HIP has close to zero overhead over CUDA on Nvidia GPU and thus provides both performance and portability. Among the first published results of porting applied software from CUDA to HIP we can mention the following studies: porting of the finite-element solver (Zubair et al, 2019), porting of MD algorithms (Kuznetsov and Stegailov, 2019), and porting of the high-performance linear algebra library Ginkgo (Tsai et al, 2020).…”