Download

License

The source code is licensed under the GNU Lesser General Public License（LGPL）

Old Version

Click Here

Overview of PARCEL

In today’s computing environments, a variety of hardware types—such as CPUs, GPUs, FPGAs, and AI accelerators—are utilized. Since each of these devices requires a different programming model (such as CUDA or OpenCL), development must be done separately, which presents challenges in terms of portability and maintainability of programs. To address these issues, the open standard SYCL was introduced by the Khronos Group. Based on C++, SYCL enables single-source parallel programming across heterogeneous computing platforms, including CPUs and GPUs.
The PARallel Computing ELements (PARCEL) library provides a parallel Krylov subspace solver optimized for parallel computing systems that offer high computational performance but relatively low inter-node communication performance. Traditionally, PARCEL has been developed independently for CPUs and NVIDIA GPUs. By re-implementing it with SYCL, we can enhance both performance portability and software maintainability. The current version has been verified on Intel/AMD/Arm CPUs as well as NVIDIA GPUs. In the future, support will be extended to AMD and Intel GPUs, and even to FPGAs. Furthermore, since computational performance and characteristics vary by hardware, selecting the appropriate precision—such as half, single, double, or quadruple precision—is crucial. The SYCL version of PARCEL supports hybrid mixed-precision processing, allowing flexible combinations of different precisions depending on the computational algorithm and data types. This enables the development of a fast and numerically stable solver that maintains computational accuracy while minimizing memory access costs.

Performance