GitLab

m is declared outside the parallelized for-loops which makes it shared
by default according the OpenMP standard as well as by the
"default(shared)" statement.  This implies a race condition and the
actual value of m is undefined with in the loops.  The patch fixes this
issue.

Note that the compiler might not allocate storage for m on the stack and
use register instead, so the race condition may not manifest in wrong
output when optimization is enabled, but the error is present anyway.

Made number of real and virtual GPUs adjustable via command line arguments.

Keep kernel data small by separating POI related data in a dedicated class. This was necessary to make the code run on the older Tesla C1060.

Time series for POIs are now stored on GPU side and only copied back at the end of the simulation. This saves a lot of runtime especially for a large number of POIs.

- Flushing the output explicitly.

- The memory alignment was adjusted.
- The progress output is now flushed.

- Fixed 'svn export' in download.py.

- Added PDF files from ftp server.
- Added a python script that can be used to download test data on demand.

- A directory "tools/" was added. It includes programs that can be used to display the results created by EasyWave.
- A directory "tests/" was added that contains some python scripts for testing different versions of EasyWave.

This is an alpha version that dynamically creates a subset of the input grid according to the source location.
The number of steps that are necessary to run a complete simulation is hardcoded for now. Maybe there is a better
approach.

This feature is needed to use EasyWave from the web where a user can specify any source location and we have to
choose the right grid to answer his request.

- Run *Update- and *Boundary-kernels in different streams and thus concurrently.

- Fixed some indexing issues in kernel WaveBoundary (memory accesses of 1D-Arrays were not 0-based).
- Closed small memory leaks caused by cuda-calls.

The memory layout should change to map the grid contents directly.
The pre-calculation of some constant arrays could be done once for each grid.