Run a two-parameter Gray-Scott reaction-diffusion simulation using Kokkos for performance portability
The following works for me on my AMD Ryzen 3950X:
git clone --recurse-submodules https://github.com/markstock/grayscott
cd grayscott
mkdir build && cd build
cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF -DKokkos_ENABLE_OPENMP=ON -DKokkos_ARCH_ZEN2=ON ..
make
And for CUDA:
cmake -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF -DKokkos_ENABLE_CUDA=ON ..
Both programs take one argument: the size of the array in each dimension.
./src/grayscott2d.bin 4096
./src/grayscott3d.bin 256
The first command will run a 2D simulation and write out files that are 4096x4096 pixels. The second will run a 3D simulation of a 256x256x256 cube, writing a center plane every 100 steps, and at the end will write each plane as a png file. These sizes were selected because each runs the simulation over 2^24 cells. This symmetry is also achievable with 32768 and 1024, respectively (1B cells, 16 GB memory needed).
On a 16-core AMD Ryzen 9 3950X, the 4096^2 problem solves at ~0.24s per step, and a 256^3 solves in ~0.075s per step. On a 3070Ti with Cuda 11.2, the 4096^2 problem achieves 0.005s and the 256^3 requires 0.008s.