dgesvd_ bechmark

[brada4]

from #1077
Shows some surplus threads at small samples:
_NUM_THREADS=1

           128x128 :    1554.22 MFlops   0.009000 sec
           256x256 :    3197.26 MFlops   0.035000 sec

_NUM_THREADS=4

           128x128 :    1076.00 MFlops   0.013000 sec
           256x256 :    2869.33 MFlops   0.039000 sec

[grisuthedragon]

Is it not better to have a C benchmark for this. Furthermore, the flopcount/rate for the SVD is a bit dangerous because although it works like a direct method, it involves an iterative process. So depending on the "quality" of the random input data this values may differ on the same machine.

[brada4]

You are right. I better make one script that drills through lapack calls that R core wraps closely in its simplest functions.