From: Ian S. <ian...@st...> - 2007-09-12 08:46:01
|
Amitha Perera wrote: > Ian, > > SSE-2 and related SIMD instruction support is great! Thanks to you > guys for the work. > > On Tue 11 Sep 2007, Ian Scott wrote: >> At present because of the generally disappointing performance for very >> small vectors and matrices, we don't want to turn on >> VNL_CONFIG_ENABLE_SSE2 by default. If anyone thinks we should - then >> please let me know. > > Is an if-statement that chooses sse2 instructions only when the vector > is large enough too much of a performance hit for the common case? > For a vnl_vector, I'd think not. For a vnl_vector_fixed, the if > statement could be eliminated using template magic. > > It's even possible write a small test case that automatically > determines the best N for the given processor. Maybe even run this > test at CMake configure time. > > Have you guys done any tests on this? It would certainly makes sense - However work on tuning the criteria would be time consuming. The bigger problem is that we don't really understand why it is so much slower for the small matrices. They aren't that small and the results imply a huge overhead. Either way we have no plans or time to investigate this in the near future. Ian. |