PolyBenchMT

中文 | English

PolyBenchMT is a modified version of PolyBench benchmarks, specifically adapted for the MT-3000 platform using hthreads. This project represents a platform migration of the original PolyBench suite, optimized for parallel execution on MT-3000.

Available Benchmarks

Convolution

• 2DCONV: 2D Convolution
• 3DCONV: 3D Convolution

Linear Algebra

• 2MM: 2 Matrix Multiplications
• 3MM: 3 Matrix Multiplications
• ATAX: Matrix Transpose and Vector Multiplication
• BICG: BiCG Sub Kernel of BiCGStab Linear Solver
• DOITGEN: Multiresolution Analysis Kernel
• GEMM: Matrix-Multiply
• GEMVER: Vector Multiplication and Matrix Addition
• GESUMMV: Scalar, Vector and Matrix Multiplication
• GRAMSCHMIDT: Gram-Schmidt Decomposition
• LU: LU Decomposition
• MVT: Matrix Vector Product and Transpose
• SYR2K: Symmetric Rank-2K Update
• SYRK: Symmetric Rank-K Update

Datamining

• CORRELATION: Correlation Computation
• COVARIANCE: Covariance Computation

Stencils

• ADI: Alternating Direction Implicit Solver
• FDTD-2D: 2D Finite Different Time Domain Kernel
• JACOBI-1D: 1D Jacobi Stencil Computation
• JACOBI-2D: 2D Jacobi Stencil Computation

Building and Running

Directory Structure

.
├── bin/
│   ├── MINI_DATASET/
│   ├── SMALL_DATASET/
│   ├── STANDARD_DATASET/
│   ├── LARGE_DATASET/
│   └── EXTRALARGE_DATASET/
└── results/
    └── YYYY-MM-DD-HH-MM/
        ├── all_results/
        └── <dataset>_result.csv

Compilation

To compile a single dataset size:

./compile.sh <dataset_size>

To compile all dataset sizes:

./compile.sh ALL

Available dataset sizes:

• MINI_DATASET
• SMALL_DATASET
• STANDARD_DATASET
• LARGE_DATASET
• EXTRALARGE_DATASET

Execution

To run benchmarks for a single dataset:

./run.sh <dataset_size>

To run all datasets:

./run.sh ALL

The results will be saved in the results/YYYY-MM-DD-HH-MM/ directory, with CSV files containing performance metrics including CPU time, DSP time, and speedup for each operator.

Modifying Codes

Key parameters that can be modified:

DATA_TYPE

• Default: float
• Can be changed to double by modifying the DATA_TYPE typedef

PERCENT_DIFF_ERROR_THRESHOLD

• Defines the acceptable percentage difference between DSP and CPU results
• Range: 0.0-100.0
• Can be adjusted per operator in the input code file

After modifications, run:

make clean
make

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.