README.md

This repo is DECREPIT - see https://github.com/opesci/devito instead

Opesci-FD

Opesci-FD is a software package to automatically generate Finite Difference models from a high-level description of the model equations. It allows the rapid development, analysis and optimisation of propagator codes for use in seismic imaging.

Dependencies

Before installation, ensure that all the dependencies specified in the requirements.txt file are installed, using

pip install -r requirements.txt

In addition, CMake must also be installed. This can either be done using

sudo apt-get install cmake

or by building CMake from source.

Installation

Opesci-FD can be installed with:

pip install --local git+https://github.com/opesci/opesci-fd.git

This will install the latest version of the opesci python package locally. To get the latest updates for your local copy simply add --upgrade to the above command. For a developer checkout of the code run:

git clone https://github.com/opesci/opesci-fd.git
cd opesci-fd
export PYTHONPATH=`pwd`:$PYTHONPATH
python setup.py build_clib --build-clib=`pwd`

Getting started

An example of the high-level model description is provided under tests/eigenwave3d.py. This script generates a stencil code that models the propagation of an eigenwave on a unit cube by solving the 3D elastic wave equation on a staggered grid. To generate the model code run:

python tests/eigenwave3d.py

This will generate the mode source code in tests/src/eigenwave3d.cpp. The source code can be compiled and executed either manually or automatically.

Automatic compilation and execution

Opesci-FD provides automatic compilation and execution that allows developers to test their code directly from the Python environment. To compile the generated source code:

python tests/eigenwave3d.py --compiler <cc>

where <cc> is either g++,clang or icpc, indicating which compiler to use. Make sure your clang compiler supports openmp for multithreaded execution. To compile and execute the above test case in parallel run:

python tests/eigenwave3d.py --compiler <cc> --execute --nthreads <nt>

where <nt> is the number of threads to use during execution. For additional options please see:

python tests/eigenwave3d.py --help

Profiling

If the PAPI library is found on your system during the initial build, Opesci-FD can also provide profiling information, such as the achieved number of MFlops/s during automated runs. To enable this feature simply add the --profiling flag to the example command above. The user can also supply a list of PAPI event names, for example PAPI_TOT_CYC or PAPI_FP_OPS, via the --papi-events flag:

python tests/eigenwave3d.py -c g++ -x --n 4 --profiling --papi-events PAPI_TOT_CYC PAPI_FP_OPS

Please note that the availability of PAPI events is highly dependent on the hardware you are running on and the local PAPI install.

Manual compilation

The generated source file can also be compiled by hand using the provided CMake file:

mkdir tests/build
cd tests/build
cmake ../src
make
bin/eigenwave3d

The outputs are grid spacings (dx, dy, dz), followed by L2 norms between numerical and analytical solutions for the stress and velocity fields. To switch off this output, change output_convergence from True to False in the Python test definition. Further parameter switches for controlling model input, the data type used (single of double precision or explicit vectorisation are also provided.

Auto-tuning for pluto

Here is a script to test the best tile size to get the best optimisation effect:

python tests/pluto_tile_test.py -b -s -l -- nthreads=4 tile_size="4 4 8"

The results will be in the "results" folder. The list "sizes" contains all the tile sizes that will be tested. This script depends on pybench.py

Loop Fission optimisation

To enable the fission of the kernel inner loop type:

python tests/eigenwave3d.py -so <order> --fission