Hi,
I have added a tiny new feature called "Probe" that allows more precise
benchmarks and simpler benchmark programs.
A probe is a small compile-time history of the CPU cycle counter.
Currently two probes are installed in SkalarFunction.cc to measure one
iteration of the inner loop of skalar functions.
A little workspace showing its usage is workspaces/SkalarBenchmark.apl
as of SVN 236.
It shows that the effect of the first measurement being much slower that
subsequent ones
is also present if the code is completely sequential, and probably not a
problem with OpenMP.
The column 1st= shows the number of cycles for the first iteration of a
loop while μ= column
shows the average of the subsequent iterations. The GNUPLOTTING line is
a placeholder for
an interface to gnuplot and does nothing yet.
As to our previous measurements, I believe we may have misinterpreted
the results and have to
look closer at how we measured.
/// Jürgen
