Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Conjugate Gradient Solvers with High Accuracy and Bit-wise Reproducibility between CPU and GPU using Ozaki Scheme

: Mukunoki, Daichi; Ozaki, Katsuhisa; Ogita, Takeshi; Iakymchuk, Roman

Fulltext ()

Online im WWW, 2020, Paper hal-02986873, 11 pp.
Japan Society for the Promotion of Science JSPS
19K20286; KAKENHI
European Commission EC
H2020; 842528; Robust
Robust and Energy-Efficient Numerical Solvers Towards Reliable and Sustainable Scientific Computations
Report, Electronic Publication
Fraunhofer ITWM ()

On Krylov subspace methods such as the Conjugate Gradient (CG), the number of iterations until convergence may increase due to the loss of computation accuracy caused by rounding errors in floating-point computations. Besides, as the order of operations is non-deterministic on parallel computations, the result and the behavior of the convergence may be non-identical in different environments, even for the same input. This paper presents a new approach for the CG method with high accuracy as well as bit-level reproducibility of computed solutions on many-core processors, including both x86 CPUs and NVIDIA GPUs. In our proposed approach, accurate and reproducible operations are installed into all the inner-product based operations such as matrix-vector multiplication and dot-product, which are the main sources that may disturb reproducibility in the CG method. The accurate and reproducible operations are performed using the Ozaki scheme, which is the error-free transformation for dot-product that can ensure the correct-rounding. As this method can be built upon vendor-provided linear algebra libraries such as Intel Math Kernel Library and NVIDIA cuBLAS/ cuSparse, it reduces the development cost. In this paper, showing some examples with the non-identical conver-gences and computed solutions on different platforms, we demonstrate the applicability and the effectiveness of the proposed approach as well as its performance on both CPUs and GPUs. Besides, we compare against an existing accurate and reproducible CG implementation based on the Exact BLAS (ExBLAS) on CPUs.