This tutorial includes several implementations of a Mandelbrot set Fractal code. The code originated from Dr. Martin Burtscher of the Efficient Computing Lab at Texas State University. You can find more here: https://userweb.cs.txstate.edu/~burtscher/research.html
In the "extras" directories, there are a few other RAJA implementations such as OpenMP and even a native CUDA implementation just for comparison. You can reference these implementations to study the differences in implementation and runtime comparison. However, anything beyond that is outside the scope of this tutorial.
In the following lessons, you will compare RAJA implementations of the fractal generating code. We will start with a sequential implementation of the fractal and gradually build our way up to a more complex RAJA launch implementation. You will notice that these lessons will employ the RAJA kernel and launch abstractions. Additionally, as the lessons progress, we will be exploring the performance portability of RAJA by looking at how we can change the targeted backend from CUDA to HIP. (Refer to lessons 02-CUDA and 03-HIP).
As before, the exercises have COMPILE macro guards. To compile the code, uncomment the COMPILE define at the top of the file.
If you are doing this tutorial outside of the RADIUSS tutorial series, be sure to build
the tutorial within a newly created, empty build directory located
in the raja-suite-tutorial repo. If you're on a LC machine, you can run these commands:
module load cuda/11.2.0
module load cmake/3.20.2
module load gcc/8.3.1
cmake -DENABLE_CUDA=On -DENABLE_OPENMP=Off -DCMAKE_CUDA_ARCHITECTURES=70 -DCMAKE_CUDA_COMPILER=/usr/tce/packages/cuda/cuda-11.2.0/bin/nvcc -DCUDA_TOOLKIT_ROOT_DIR=/usr/tce/packages/cuda/cuda-11.2.0 -DBLT_CXX_STD=c++14 -DCMAKE_BUILD_TYPE=Release -DRAJA_ENABLE_EXERCISES=On -DRAJA_ENABLE_OPENMP=Off -DCMAKE_CUDA_FLAGS=--extended-lambda -DCUDA_ARCH=sm_70 ../
I am building this code on LC's lassen machine. If these build instructions don't work for you, you can refer to the build documentation from RAJA's ReadTheDocs or use one of the provided build scripts.
Now, we can build the RAJA loop-exec implementation with ./bin/fractal 1024. The
first argument is the width of the fractal (1024). It may be interesting to see how
the fractal changes with different width values.
To verify your results in each lesson, you can look at the resulting .bmp file output. If you
have completed everything correctly, you will see a complete image of the fractal.
Currently, there is an if statement that makes sure the writeBMP function
is only called for smaller fractal runs (of width <= 2048). You can edit this if statement, but be careful because trying
to write a .bmp file that is too large will take a very long time.
Continue on to the first lesson located in the 00-BASE directory.