2015 - Institute of Computing, Unicamp, Brazil
Authors: Luís Felipe Mattos, Rafael Cardoso & Márcio Pereira
The UniBench is a collection of open source benchmark suites organized in a simple and modular structure. This benchmark suites were converted to make use of the new target
directive on the OpenMP 4.0, and were developed to test our version of the Clang compiler, which translates OpenMP 4.0 into OpenCL programs.
To use the UniBench, the script needs the right permissions. If the UniBench script does not have the execution permission, just use:
chmod +x ./unibench
Use ./parboil help
to display the commands and the available benchmarks.
Use ./parboil list
to display the available benchmarks.
The UniBench script offers many options for the execution. If you want to execute a specific benchmark from a specific benchmark suite, you need to specify both the suite and the benchmark in the command line. For example, if you want to compile or execute the 2MM benchmark from the Polybench suite, you just need to use:
./unibench compile Polybench/2MM
./unibench run Polybench/2MM
The script also accepts the option to execute the complete benchmark suite, you can do this by specifying only the suite in the command line. For example, if you want to compile or execute the Parboil suite, just type:
./unibench compile Parboil
./unibench run Parboil
However, if you want to compile or execute all the available benchmarks, just use:
./unibench compile all
./unibench run all
An important point is that this script is case sensitive, if you want to be sure that your benchmark name is correct, use the list
command before executing the UniBench.
For executing this benchmark on mobile phones with Mali GPU devices, just add the flag -mali
, for example:
./unibench compile Parboil -mali
For proper compilation when executing on mobile phones, the environment must be working before the compilation!
In this case, binaries, kernels and temporary execution logs will be saved in the directory /data/local/tmp/<Benchmark Suite>/<Benchmark Name>
.
The script also offers the clean
command, with the same options from the previous examples. You can clean the binaries from a specific benchmark from a suite, from a whole suite or from all benchmarks. For example:
./unibench clean Polybench/2MM
./unibench clean Parboil
./unibench clean all
The UniBench outputs the execution time on the GPU device, the execution time for the serial version running on the CPU, and the comparison of the results with an acceptable threshold for the float precision. Here is an example of output for the execution of the Polybench-2MM:
<< Linear Algebra: 2 Matrix Multiplications (D=A.B; E=C.D) >>
GPU Runtime: 2.554422s
CPU Runtime: 9.622760s
Non-Matching CPU-GPU Outputs Beyond Error Threshold of 0.05 Percent: 0