On Fri Apr 22, 2016 at 09:34:38AM +0100, Stuart Henderson wrote: > On 2016/04/21 19:05, Michael McConville wrote: > > Stuart Henderson wrote: > > > On 2016/04/21 20:58, Rafael Sadowski wrote: > > > > On Wed Apr 20, 2016 at 05:30:26AM -0400, Michael McConville wrote: > > > > > It's on by default, but can be disabled with the below cmake > > > > > argument. > > > > > > > > Are test results with and without SSE3 same or better? If yes okay > > > > for me. > > > > > > I think it would have to be "better" otherwise what's the point in > > > disabling this in opencv (which does runtime detection)? > > > > Are you sure? Looking at the build logs, it seems like -msse3 is in the > > build-global C and C++ flags, so it'll be used for autovectorization > > (see build snippet below). > > Hmm. Can you check with upstream then? They're doing runtime detection > based on cpuid in their code so it seems they're trying to be portable. > I'd sooner see if they'll fix it and backport that fix instead, > otherwise this change is likely to get forgotten and left in even if > they fix things later - there is clearly code in a program like this > which is going to benefit from fast matrix ops if available.
Upstream looks not smarter with GCC flags handling. My 3.1.0 port is almost ready but needs more runtime tests with nomacs and digikam. I run tests with and without Michael McConville's patch. Everything looks good. Okay for me till 3.1.0 is finally completed but Stuart has the last word. Best regards, Rafael
