Andreas Schäfer
Andreas Schäfer
| Function | Memory | Parallelism | Grid | Threading | Model | TestClass | | --- | --- | --- | --- | --- | --- | --- |...
...because global variables are ugly. Replace find_all_localities() by find_all_basename() and let the HpxSimulator hold the promises for UpdateGroup speeds and indices. Positive side effect: eliminates further races in initialization.
In short: we should replace sequential code which enforces implicit barriers by futurized code where the runtime is able to extract the maximum degree of parallelism and improve efficiency. Right...
HPX can migrate components between localities. This migration is transparent to all objects in the code as components are accessed through IDs and the ID would not change after migration....
...which represents a performance degradation. Proposed solution: - add HPX-based performance tests - add LBM test - add n-body test - add unstructured grid test - rework threading in UpdateFunctor
It's a source for nasty errors that Writers and Steerers need to save their own copy of maxSteps, as the user may easily forget to initialize them correctly. This change...
Quick demo to verify autotuning effectiveness, should use SimulatorFactory, CacheBlockingSimulator, CudaSimulator and Jacobi3D model. Moved from https://bitbucket.org/gentryx/libgeodecomp/issues/21/implement-compiler-flag-autotuning-demo
We have all the infrastructure for the SoA code ready, but ATM there is no test for the HiParSimulator which ensures that this actually works. Suggestion: implement this test in...
Should use single wavefront, not one per core Originally reported here: https://bitbucket.org/gentryx/libgeodecomp/issues/2/finish-cacheblockingsimulator
Moved from https://bitbucket.org/gentryx/libgeodecomp/issues/20/visitwriter-should-work-with-unstructured