OK, I had a closer look at this.
Buccaneer is failing because it can't grow a single residue of chain
from the seed positions. This can only happen if the map is total garbage.
So the first thing to do is check the phases which are going into
buccaneer in both the gcc and intel cases. Calculate a map from each
set, and make sure they are both reasonably similar and interpretable
(they should be for the example data).
If the two maps, differ, then the problem is in the upstream program.
If they don't, then the most likely cause is that buccaneer is not
calculating the map correctly. Add 'mapout buc.map' to the buccaneer
command file, and rerun both the gcc and intel versions. If the
buccaneer map from the intel version is corrupted, then the problem is
most likely the linking of the fft library in the intel case.
Kevin
Nobuo OKAZAKI wrote:
Dear Bill,
Yes, the first problem I faced was precision. In double precision,
all of CLIPPER programs did not run. After that, I have used with single
precision MKL's FFT. And CLIPPER programs except cbuccaneer.exam
successfully ran.
So, I wonder why only the BUCCANEER fails under this environment. :(
Certainly, using pre-built package is easear and faster than building by
myself. But I want to compare performance between pre-built and Intel
MKL. :)
In my trial, test(make runtest) under Intel Compiler with Intel MKL(as
LAPACK without FFT) is 8% faster than pre-build package. I wornder how
much contributes with MKL's FFT library.
Nobuo OKAZAKI
Bill wrote:
I should let the experts answer definitively, but I think this might
be because the clipper-related stuff (and coot) need single-precision
s[r]fftw libraries.
I had to do something like this:
LDFLAGS="-Wl,-dylib_file,/sw/lib/libsfftw.2.0.7.dylib:/sw/lib/
libsrfftw.2.0.7.dylib"
It will probably be a bit different on Fedora (linux uses .so instead
of .dylib suffixes). The main point is to force-feed it the single-
precision libraries.
In practice it is probably just easier just to let ccp4 build its own,
unless you need the mpi versions. Single precision makes the
calculation go faster and you can't see the difference in the maps, in
my experience.
HTH,
