Status meeting October 25th, 2012¶
Soon: Optimization leads to very good number. Performance improved from 35 to 55 (200 without uploads).
Dongwook: Working on debugging the bremstralung postDoIt CUDA version.
Build system.¶Marc: we should stop using the
-fast-mathoption until we really need them. We should also using
includecoming only from a single cuda SDK.
We need a build system that will prevent inconsistent builds. Marc has already engaged Lynn's support for:
- rewriting the
g4hpcbenchmarksbuild files to use
Lynn will be using the
new-build-systembranch in the repository for this purpose. Anyone who wants to track the work can do so, but don't expect things to build immediately on that branch.
We have decided that the
nvcc product with not set up a specific version of
gcc during its setup. Rather, a product like
g4hpcbenchmarks will have its qualifiers specify what version of
nvcc and what version of
gcc are to be used to produce a specific build.