A minimalistic kokkos example to compute Mandelbrot set and illustrate asynchronous memory copy (i.e. overlap between a computationnal functor and a deep copy operation).
Three versions provided:
-
basic : Mandelbrot set is computed in a single Kokkos functor. Can be used with either OpenMP or Cuda backend.
-
basic_mdrange : same version a basic, just for illustrating the use of Kokkos::Experimental::MDRange
-
pipeline0 performs computations piece by piece
-
pipeline1 (only meaningfull when used with CUDA+OpenMP) performs computations piece by piece, but the loop over the pieces is parallelized using an OpenMP Kokkos functor, so that the different pieces can be computed in different Cuda streams.
What is kokkos ?
A modern C++ based programming model for HPC applications designed for portability across multiple hardware architectures (multicore, GPU, KNL, Power8, ...) and also providing as efficient as possible performance.
Build the basic version
-
Need to have installed kokkos
- Kokkos backend can be CUDA, OpenMP, ...
- Compiler can be nvcc_wrapper, g++, xlc++, ...
-
Set env variable KOKKOS_PATH to the root directory where Kokkos is installed
-
cd basic; make
-
run
./mandelbrot.omp (or ./mandelbrot.cuda)
With default parameters (image of size 8192x8192), some performance of the basic version:
- Nvidia K80 : 1.5 seconds
- Power8 (g++ 4.8.5)
- 20 threads : 50.1 seconds
- 40 threads : 27.7 seconds
- 60 threads : 16.9 seconds
- 160 threads : 10.5 seconds
NB: version pipeline1 require kokkos to be configured with both CUDA and OPENMP backends, and lambda function enabled. Example command line configuration:
generate_makefile.bash --with-cuda --arch=${YOUR_CUDA_ARCH} --prefix=${SOMEWHERE} --with-cuda-options=enable_lambda --with-openmp