Fedora Account System
Red Hat Associate
Red Hat Customer
Description of problem: The libcblas.a is built, but not installed. Please install it. The header for cblas is already in the devel package, but it is unusable if there is no cblas library. Version-Release number of selected component (if applicable): atlas-3.10.2-1.fc23 How reproducible: Always Steps to Reproduce: 1. Check buildlog for libcblas.a 2. Check resultin RPMs for libcblas.a Actual results: libcblas is built but not packed to the RPM Expected results: libcblas is packed in the RPM Additional info:
It's blocking building of the hpl package on secondary arches.
Also isn't possible to built the libs as shared?
Shared lib is built and shipped as customary in Fedora. Build log for x86_64 (https://kojipkgs.fedoraproject.org//packages/atlas/3.10.2/1.fc23/data/logs/x86_64/build.log) says: """ /usr/bin/gcc -g -Wa,--noexecstack -fPIC -m64 -shared -o libsatlas.so.3.10 \ \ -Wl,-soname,"libsatlas.so.3" \ -Wl,--whole-archive liblapack.a libf77blas.a libcblas.a libatlas.a /builddir/build/BUILD/ATLAS/liblapack_pic_pruned.a -Wl,--no-whole-archive -L/usr/lib/gcc/x86_64-redhat-linux/5.1.1 -lgfortran -lc -lpthread -lm -lgcc """ Latest upstream decision is to link everything together and create only one big library to simplify linking. Actually there are two in Fedora, one single-threaded and one multithreaded, choose one (choose single-threaded if I can suggest). You are right, atlas-static subpackage is useless with libcblas and other libraries missing.
(In reply to Frantisek Kluknavsky from comment #3) Thanks for info, I used libsatlas.so fallback on ARM.
http://koji.fedoraproject.org/koji/taskinfo?taskID=9805910 Included single-threaded interfaces.
Tried to benchmark the libraries to be sure that the correct optimal versions are packaged and not some artifacts from optimalization process. The results are not sure on my laptop. Could you please test that the static library gives the same performance as the dynamic library?
Thanks. Built-in check in hpl package returns the same results for both dynamic/static libraries and the non-optimized HPL benchmark also seems to return similar numbers.