PARCEL SYCL v1.2 Search Download 日本語

PARCEL SYCL v1.2 Manual

The manual consists of two documents. The link at the top of each page switches between the Japanese and English versions.

DocumentContents
User GuideFor users of the library: quick start, basic structure, SYCL compiler and installation (including spack and Homebrew on macOS), usage from C++, C, Fortran and Python, algorithms (linear solvers, eigenvalue solvers, preconditioners, QR factorization), choosing a solver, performance, versions and compatibility.
API ReferenceSpecifications of the public classes and functions: language-independent common items (solvers, input matrix, meaning and defaults of the settings, constants, results, convergence states, convergence by solver) and the C++ API (data types, comm, matrix, vector, member functions of each solver), C, Fortran and Python APIs, in a uniform format per function. The common items and the chapters of one language are enough to read. Switch the language with the tabs at the top of each page. Includes a function index.
SearchKeyword search over the documents (User Guide, API Reference).

Download

The latest version (v1.2, released on 2026-10-23) is in PARCEL-release.

Past versions: SYCL version v1.0 is in PARCEL-release, and PARCEL 1.0–1.2 before the SYCL version (Fortran, CUDA) are in PARCEL-legacy.

License

Open source under the GNU Lesser General Public License (LGPL-3.0).

Purpose and Overview

Today's computing environments use many kinds of hardware: GPUs, FPGAs and AI accelerators in addition to CPUs. These devices require different programming models (CUDA, OpenCL, etc.), which makes portability and maintenance of programs difficult. SYCL, an open standard defined by the Khronos Group, addresses this problem: it is based on C++ and enables single-source parallel programming on heterogeneous platforms including CPUs and GPUs.

The PARallel Computing ELements (PARCEL) library provides parallel Krylov subspace solvers suited to parallel computers whose computing performance is high but whose inter-node communication performance is relatively low. PARCEL had been developed separately for CPUs and NVIDIA GPUs; reimplementing it in SYCL improves performance portability and maintainability. v1.2 has been confirmed on Intel / Arm CPUs (including A64FX and Apple M4) and NVIDIA / AMD GPUs, and support for Intel GPUs, FPGAs and others is planned.

Because the performance and characteristics differ from hardware to hardware, choosing appropriate precisions (half, single, double, double-double) is important. SYCL PARCEL supports mixed-precision computation that flexibly combines different precisions depending on the algorithm and data, achieving fast and numerically stable solvers that keep the accuracy while reducing the cost of memory access.

Changes in v1.2

Solvers

KindSolverProblem
Linear solverspcg: preconditioned conjugate gradient methodLinear systems with symmetric positive definite matrices; the lightest
pbicgstab: preconditioned stabilized biconjugate gradient methodLinear systems with nonsymmetric matrices, when each iteration should be light
pgmres: preconditioned generalized minimal residual methodLinear systems with nonsymmetric matrices, or symmetric but not positive definite ones, when stable convergence is wanted
pcagmres: preconditioned communication-avoiding generalized minimal residual methodThe same problems as pgmres, when many ranks make collective communication the bottleneck
Eigenvalue solverspower: power methodOne eigenvalue of largest magnitude
inverse_power: inverse power methodOne eigenvalue of smallest magnitude
ks_arnoldi: Krylov-Schur methodSeveral eigenvalues, complex conjugate pairs (nonsymmetric matrices), eigenvalues near a shift by shift-invert
PreconditionersPoint Jacobi preconditionerThe cheapest; small effect
Block Jacobi preconditionerFor CPUs; its effect depends on the numbers of ranks and blocks
Jacobi preconditionerFor GPUs; its effect does not depend on the partitioning

APIs for calling these solvers are provided for C++, C, Fortran and Python: C++ as a header-only library, C and Fortran as an API library, and Python as the package parcel (API Reference).

Tested Environments

MachineCPU / GPUOSSYCL environmentBackend environment
SGI8600 CPUIntel Xeon Gold 6242R (20 cores) ×2RHEL 8.10AdaptiveCpp 24.10.0 (LLVM 18.1.8)OpenMP
Open MPI 4.1.6
SGI8600 GPUIntel Xeon Gold 6248R (24 cores) ×2
NVIDIA V100-SXM2-32GB ×4
RHEL 8.10AdaptiveCpp 24.10.0 (LLVM 18.1.8)CUDA 12.6
Open MPI 4.1.6
Miyabi-CIntel Xeon CPU Max 9480 (56 cores) ×2RHEL 9.8AdaptiveCpp 24.10.0 (clang 17.0.6, built with Intel oneAPI 2023.2.0)OpenMP
Open MPI 4.1.6
Miyabi-GNVIDIA GH200 (Grace 72 cores + H100 96GB)RHEL 9.8 (aarch64)AdaptiveCpp 24.10.0 (LLVM 18.1.8)CUDA 12.6
Open MPI 4.1.6 (CUDA-aware)
OdysseyFUJITSU A64FX (48 cores)RHEL 8AdaptiveCpp 24.10.0 (omp.library-only, Fujitsu compiler FCC 1.2.43)OpenMP (Fujitsu)
Fujitsu MPI 1.2.43
AquariusIntel Xeon Platinum 8360Y (36 cores) ×2
NVIDIA A100-SXM4-40GB ×8
RHEL 8.10AdaptiveCpp 24.10.0 (LLVM 18.1.8)CUDA 12.2
Open MPI 4.1.1
Plasma Simulator AIntel Xeon 6980P (128 cores) ×2RHEL 9.6AdaptiveCpp 24.10.0 (LLVM 18.1.8)OpenMP
Intel MPI 2021.18.1
Plasma Simulator BAMD MI300A (CPU 24 cores + GPU) ×4RHEL 9.6AdaptiveCpp 24.10.0 (LLVM 18.1.8)HIP (ROCm 6.3.3)
Open MPI 5.0.7
MacApple M4 (10 cores: 4 performance, 6 efficiency)macOS 15AdaptiveCpp 25.10 (omp.library-only, Apple clang 17)OpenMP
Open MPI 5.0
Linux Workstation (WSL2)Intel Xeon Gold 6226R (16 cores) ×2
NVIDIA RTX A6000 ×2
Ubuntu 24.04AdaptiveCpp 24.10.0 (LLVM 18.1.8)OpenMP, CUDA 12.6
Open MPI 4.1.6
Linux Workstation (WSL2)Intel Xeon Gold 6226R (16 cores) ×2
NVIDIA RTX A6000 ×2
Rocky Linux 9.8AdaptiveCpp 24.10.0 (LLVM 18.1.8)OpenMP, CUDA 12.6
Open MPI 4.1.1

Each machine name links to its section in the User Guide (compilers, library versions, installation). On other general Linux systems, everything from LLVM to PARCEL can be installed with spack (scripts in the distribution package).

Performance

Time and memory per device of v1.2, v1.0 and PETSc 3.22 (conjugate gradient method with the Jacobi preconditioner, 3D Poisson equation, 512×512×512 grid, double precision, 16 devices). The conditions and the results for other solvers are in the "Performance" chapter of the User Guide.

Time [s] v1.2 is up to 1.56× faster than v1.0 and up to 1.76× faster than PETSc SGI8600 CPU (Xeon Gold 6242R) 24.2 s 30.3 s vs v1.0: 1.25× faster 42.6 s vs PETSc: 1.76× faster Odyssey (A64FX) 9.64 s 15.0 s vs v1.0: 1.55× faster 11.5 s vs PETSc: 1.19× faster Miyabi-C (Xeon CPU Max 9480) 6.82 s 7.32 s vs v1.0: 1.07× faster 7.77 s vs PETSc: 1.14× faster Plasma Simulator A (Xeon 6980P) 5.73 s 6.62 s vs v1.0: 1.15× faster 5.76 s vs PETSc: same SGI8600 GPU (V100) 7.18 s 9.47 s vs v1.0: 1.32× faster 7.67 s vs PETSc: 1.07× faster Miyabi-G (GH200) 1.34 s 1.47 s vs v1.0: 1.09× faster 1.35 s vs PETSc: 1.01× faster Aquarius (A100) 3.92 s 4.02 s vs v1.0: 1.02× faster 4.45 s vs PETSc: 1.13× faster Plasma Simulator B (MI300A) 2.11 s 3.29 s vs v1.0: 1.56× faster 2.13 s vs PETSc: 1.01× faster v1.2 v1.0 PETSc 3.22
Memory per device [GiB] (host on CPUs, device on GPUs) v1.2 uses up to 45% less memory than v1.0 and up to 80% less than PETSc SGI8600 CPU (Xeon Gold 6242R) 2.89 GiB 3.41 GiB vs v1.0: 15% less 6.20 GiB vs PETSc: 53% less Odyssey (A64FX) 2.97 GiB 3.93 GiB vs v1.0: 24% less 4.55 GiB vs PETSc: 35% less Miyabi-C (Xeon CPU Max 9480) 2.37 GiB 2.93 GiB vs v1.0: 19% less 9.39 GiB vs PETSc: 75% less Plasma Simulator A (Xeon 6980P) 4.74 GiB 4.65 GiB vs v1.0: 2% more 23.9 GiB vs PETSc: 80% less SGI8600 GPU (V100) 1.73 GiB 2.80 GiB vs v1.0: 38% less 1.56 GiB vs PETSc: 11% more Miyabi-G (GH200) 1.97 GiB 3.05 GiB vs v1.0: 35% less 1.82 GiB vs PETSc: 9% more Aquarius (A100) 1.84 GiB 2.92 GiB vs v1.0: 37% less 1.69 GiB vs PETSc: 9% more Plasma Simulator B (MI300A) 1.46 GiB 2.67 GiB vs v1.0: 45% less 1.50 GiB vs PETSc: 3% less v1.2 v1.0 PETSc 3.22

Developer

Center for Computational Science & e-Systems, Japan Atomic Energy Agency

Contact

Questions and bug reports: GitHub Issues, https://github.com/parcel-jaea/PARCEL-release/issues