DRStencil

GPU code generator for low-order stencil computations.
Given the neccessary stencil parameters in a .stc file, DRStencil generates highly data-reusing GPU codes for stencil computation. DRStencil provides many optimazation techniques, and users are allowed to adjust the optimazation configurations.

DEPENDENCIES

1. python3 (version 3.8.5 tested)
2. GCC (support for C++ 17 needed, version 9.3.0 tested)
3. CUDA (version 11.2 tested)
4. Nsight Compute (version 2020.3.1.0 tested)

SETUP

To build drstencil, you only need to run "make" in the main directory with the system supporting C++ 17.

cd drstencil
make

Then, you'll get an executable drstencil.

USAGE

Copy the executable drstencil into the test directory, 2d5pt_star for instance. By launching the script starter.sh, you can start tuning GPU code for the stencil 2d5pt_star. The script generates hundreds or thousands of optimazation configurations and profiles each of them with Nsight Compute. The profiling results will be located in prof/ directory.

cd benchmarks/2d5pt_star/
cp ../../drstencil .
./starter.sh

To collect the GPU metrics, launch the script getGpuMetrics.sh. (Note that the script is only supported by Nsight Compute 2020.3. Some modification need to be made for other versions.)

bash getGpuMetrics.sh

Usage of drstencil:

./drstencil [options] <input_stcfile>

Lanuch the following command for more detailed usage of drstenil.

./drstencil -h

To Create a new test, you need a .stc file like benchmarks/2d5pt_star/2d5pt_star.stc, which defines the stencil computation.

PUBLICATION

[HPCC 2021] DRStencil: Exploiting Data Reuse Within Low-Order Stencil on GPU.