PARCEL SYCL v1.2
/
User Guide
User Guide
API Reference
Search
Download
日本語
Contents
1
Quick Start
1.1
Overview of the Program
1.2
Installing a SYCL Compiler
1.3
C++
1.4
C
1.5
Fortran
1.6
Python
1.7
Where to Go Next
2
Basic Structure
2.1
Data Structures
2.2
GPUs, MPI Ranks and Data
2.3
Precision Configurations
3
Installing a SYCL Compiler
3.1
s86 (SGI8600)
3.2
Odyssey (Wisteria-Odyssey, A64FX)
3.3
Aquarius (Wisteria-Aquarius, A100)
3.4
Miyabi-C (Xeon Max)
3.5
Miyabi-G (GH200)
3.6
PS-A (Plasma Simulator Subsystem A)
3.7
PS-B (Plasma Simulator Subsystem B)
3.8
macOS (Homebrew)
3.9
Generic Linux (RHEL Family, Ubuntu Family)
3.10
Installing LAPACKE
4
Installation
4.1
Package Layout
4.2
Checking After the Build
4.3
Compile Options
4.4
Macros
4.5
Recommended Options
4.6
Default of MPI
4.7
Build without MPI
4.8
Python Package
4.9
Installing with spack
5
Usage
5.1
C++
5.1.11
Passing Your Own Matrix
5.2
C
5.2.10
Passing Your Own Matrix
5.3
Fortran
5.3.10
Passing Your Own Matrix
5.4
Python
5.4.11
Passing Your Own Matrix
6
Choosing a Solver
6.1
Linear Solvers
6.2
Eigenvalue Solvers
6.3
Preconditioners
7
Running on GPUs and in Jobs
7.1
Running on GPUs
7.2
Running in Jobs
8
Linear Solvers
8.1
Preconditioned Conjugate Gradient Method [1]
8.2
Preconditioned Biconjugate Gradient Stabilized Method [1]
8.3
Preconditioned Generalized Minimal Residual Method [1]
8.4
Preconditioned Communication-Avoiding Generalized Minimal Residual Method [2,3]
8.5
Precision of Orthogonalization and of the Basis
9
Eigenvalue Solvers
9.1
Power Method [4]
9.2
Inverse Power Method [4]
9.3
Krylov-Schur Method [5]
10
Preconditioners
10.1
Scaling
10.2
Point Jacobi Preconditioner [1]
10.3
Jacobi Preconditioner [1]
10.4
Block Jacobi Preconditioner [1]
11
QR Factorization
11.1
Classical Gram-Schmidt Method [1]
11.2
Modified Gram-Schmidt Method [1]
11.3
Classical Gram-Schmidt Method with Reorthogonalization (CGS2) [6]
11.4
Cholesky QR Method [7]
11.5
Cholesky QR Method by the Moment Method (QR_CAGMRES_CHOLESKYQR)
12
Numerical Accuracy and Reproducibility
12.1
Checking the Convergence Status
12.2
Recurrence Residual and True Residual
12.3
Differences by Degree of Parallelism and Execution Environment
12.4
Known Limits at Low Precision
12.5
double-double
12.6
Do Not Use -ffast-math
13
Optimization
13.1
Defaults per Environment
13.2
Blocking
13.3
Ladder Splitting
13.4
Decomposing gemv
13.5
Optimizations for A64FX (A to F)
13.6
Register Limit on V100
13.7
Choosing Between Memory and Speed
14
Performance
14.1
Measurement Conditions
14.2
Time to Convergence
14.3
Mixed Precision (precon_dataF)
14.4
Comparison with PETSc / SLEPc and v1.0
14.5
Measuring Your Own Environment
15
Versions and Compatibility
15.1
Compatibility Policy
15.2
Checking the Version
15.3
Citation and Support
15.4
Verifying the Distribution
15.5
Changes from v1.0 to v1.2
16
Glossary
16.1
Systems
16.2
Processors and GPUs
16.3
Components and Resource Units
16.4
Software
16.5
PARCEL Terms
16.6
Computational Terms
References
Quick Start →
PARCEL SYCL v1.2 User Guide
For users of the library.
Contents
1
Quick Start
1.1
Overview of the Program
1.2
Installing a SYCL Compiler
1.7
Where to Go Next
1.3
C++
1.4
C
1.5
Fortran
1.6
Python
2
Basic Structure
2.1
Data Structures
2.2
GPUs, MPI Ranks and Data
2.3
Precision Configurations
3
Installing a SYCL Compiler
3.1
s86 (SGI8600)
3.2
Odyssey (Wisteria-Odyssey, A64FX)
3.3
Aquarius (Wisteria-Aquarius, A100)
3.4
Miyabi-C (Xeon Max)
3.5
Miyabi-G (GH200)
3.6
PS-A (Plasma Simulator Subsystem A)
3.7
PS-B (Plasma Simulator Subsystem B)
3.8
macOS (Homebrew)
3.9
Generic Linux (RHEL Family, Ubuntu Family)
3.10
Installing LAPACKE
4
Installation
4.1
Package Layout
4.2
Checking After the Build
4.3
Compile Options
4.4
Macros
4.5
Recommended Options
4.6
Default of MPI
4.7
Build without MPI
4.8
Python Package
4.9
Installing with spack
5
Usage
5.1
C++
5.1.11
Passing Your Own Matrix
5.2
C
5.2.10
Passing Your Own Matrix
5.3
Fortran
5.3.10
Passing Your Own Matrix
5.4
Python
5.4.11
Passing Your Own Matrix
6
Choosing a Solver
6.1
Linear Solvers
6.2
Eigenvalue Solvers
6.3
Preconditioners
7
Running on GPUs and in Jobs
7.1
Running on GPUs
7.2
Running in Jobs
8
Linear Solvers
8.1
Preconditioned Conjugate Gradient Method [1]
8.2
Preconditioned Biconjugate Gradient Stabilized Method [1]
8.3
Preconditioned Generalized Minimal Residual Method [1]
8.4
Preconditioned Communication-Avoiding Generalized Minimal Residual Method [2,3]
8.5
Precision of Orthogonalization and of the Basis
9
Eigenvalue Solvers
9.1
Power Method [4]
9.2
Inverse Power Method [4]
9.3
Krylov-Schur Method [5]
10
Preconditioners
10.1
Scaling
10.2
Point Jacobi Preconditioner [1]
10.3
Jacobi Preconditioner [1]
10.4
Block Jacobi Preconditioner [1]
11
QR Factorization
11.1
Classical Gram-Schmidt Method [1]
11.2
Modified Gram-Schmidt Method [1]
11.3
Classical Gram-Schmidt Method with Reorthogonalization (CGS2) [6]
11.4
Cholesky QR Method [7]
11.5
Cholesky QR Method by the Moment Method (QR_CAGMRES_CHOLESKYQR)
12
Numerical Accuracy and Reproducibility
12.1
Checking the Convergence Status
12.2
Recurrence Residual and True Residual
12.3
Differences by Degree of Parallelism and Execution Environment
12.4
Known Limits at Low Precision
12.5
double-double
12.6
Do Not Use -ffast-math
13
Optimization
13.1
Defaults per Environment
13.2
Blocking
13.3
Ladder Splitting
13.4
Decomposing gemv
13.5
Optimizations for A64FX (A to F)
13.6
Register Limit on V100
13.7
Choosing Between Memory and Speed
14
Performance
14.1
Measurement Conditions
14.2
Time to Convergence
14.3
Mixed Precision (precon_dataF)
14.4
Comparison with PETSc / SLEPc and v1.0
14.5
Measuring Your Own Environment
15
Versions and Compatibility
15.1
Compatibility Policy
15.2
Checking the Version
15.3
Citation and Support
15.4
Verifying the Distribution
15.5
Changes from v1.0 to v1.2
16
Glossary
16.1
Systems
16.2
Processors and GPUs
16.3
Components and Resource Units
16.4
Software
16.5
PARCEL Terms
16.6
Computational Terms
References
Quick Start →