On clang/MacOSX i get an impressive improvement in performance, around 4x on the operations, even just using the vector types for copying data; my impression is that the compiler use some kind of vector load/store instruction that properly use the available memory bandwidth, but unfortunately i do not know more about the x86 architecture. On gcc/Linux, (gcc 4.5.2) the same code produce a *slow down* of around 2.5x.

Maybe. A better comparison would be: clang/Linux vs. gcc/Linux and clang/MacOSX vs gcc/MacOSX compiled binaries. Also as Dan already pointed out: gcc has a whole lot of optimization flags which are not enabled by default. try '-O3 -msse2 -ffast-math'.  '-ftree-vectorizer-verbose=2' is handy while optimizing code.

Short resume of my initial post: i found that using the gcc vector extensions induced a 2x slow down using gcc, and a 4x speed up in clang. I made more tests, isolating a small code example, on Mac OS and Ubuntu, and i found out the origin of the problem, even if i do not know what exactly happening. My original test used vectors of float of size 8; the gcc vector extension documentation says that if the vector size do not match the hardware vector size, the code is synthesized in some way. With a vector size of 8 i found the above results under Mac OS X, using clang and gcc4.2, and under Ubuntu 11.04, using clang and gcc4.5.2. When i move to a vector size of 4, things go better; clang slow down wrt the size of 8 of around 2x, and gcc obtains the same result; the interesting point is that gcc obtains essentially the same speed with and without vector extensions, meaning probably that the compiler is good enough in vectorizing the code, at least in the the test cases i used. I include the code, results and scripts to run the tests in a small zip if anybody want to make other tests; the test code compute an arbitrary vector computation (essentially 100 million multiply add), starting from a seed given as argument. The code is modeled around the way jmax compute, i.e. one vector operation at a time on vectors passed by pointers, and it is not designed to be the fastest possible code to implements this computation. Thanks for the help, Maurizio

Thanks for coming up with and sharing the tests!

would need to look into asm output.

BTW. I'm pretty much impressed that some tests run faster on vmware virtualization compared to native OSX. But yeah it's also no big surprise either (and may further be related to ubuntu using a never version of gcc).

So are you now considering use some #ifdef to select float/4 instead of double/8 vectors in jMax or just change all of them?