On Wed, Jan 18, 2012 at 08:34:41PM +0100, Thomas Koenig wrote: > Hello, > > I have set up a semi-automatic lapack tester on > powerpc64-unknown-linux-gnu (gcc110 on the gcc compile farm). > It downloads the current trunk, compiles it, then uses that > compiler to compile the reference BLAS and Lapack and run the LAPACK > test suite, cycling through the Lapack build with different compiler > options. > > An example of the results is at > > http://gcc.gnu.org/ml/gcc-testresults/2012-01/msg01757.html > > It is inexpensive to add another set of options to the scripts. > If there is interest, I could easily do so. > > I have found one problem, PR 51751, which could also be an upstream bug. >
PR 51751 seems to concern only complex*16 failures. Have you had a chance to look at the real and double precision errors? Being semi-automatic, can you run the timing tests, too? These would obviously reveal possible performance regressions. It would also be interesting to see if -ftree-vectorize helps performance. Against my better judgement, I'll suggest a run with '-O3 -ffast-math'. It seems a large number of gfortran users use -ffast-math. -- Steve
