The regression test suite needs to be refactored to handle different implementations (cuda, OpenMP, TBB). It was written with the assumption that there was a single implementation (CUDA). Today, there can be different cuda implementation directories (for developer machines with more than one host device), an OpenMP implementation, and a TBB implementation. The regressions/Makefile needs to be parameterized to handle this.
This is a precondition for [1]
[1] https://sourceforge.net/p/mpgraph/tickets/2/ (Regression test suite needs to be updated for all implemented algorithms)
Here are the results for the GTX570 on Bryan's GPU workstation:
running on Bryan's GPU workstation as 'make -k test' for dev1 (K20). (The -k option is specified to make progress in case a GPU does not have enough RAM for a given data set.)
Here is a re-run of the K20 on the same host:
And here are the OpenMP timings for the same machine @ 3Ghz. Since this is not an Intel, that is all the operating modes that I can run against this machine. These runs are against r34 (basically, the same as the initial release except that some value reporting was commented out by Zhisong in r33).
Fix allows you to specify the implementation (cuda, open mp, or tbb) in r35.