Issues building scipy
"Olivier B." <[email protected]> Tue, 9 Jul 2019 18:32:05 +0200
| Newsgroups | gmane.comp.python.scientific.user |
|---|---|
| Message-ID | <CA+cSArimYYOVQ72O4dCDVuegZUMZF5fjDUNA-re=XNKP5OE8rw@mail.gmail.com> |
I am creating this thread to discuss issues i encounter while trying to build scipy Python is a 3.5.2 built from sources Numpy is 1.16.4 built from sources. With MKL 2019 link Scipy is 1.3.0 gfortran is the MINGW64 version 8.1 C compiler is VS2015 First tried platform is win64 / VS2015 First issue i encountered was that gfortran was in a path with spaces, leading to incorrect command lines because the path to gfortran is not quoted. I did junbctions to get a path to the same place without spaces. I still see some paths with spaces inb /LIBPATH: arguments of cl, but strangely that does not seem to bring issues. Next issue seems to be about undefined symbol _dgtsv. Which seems to be a lapack symbol. scipy seems to have found site.cfg of numpy build with mkl config, but the gfortran command line that fails does not have link to mkl libs references in site.cfg. Am i right to think i need to pass extra link flags to mkl to gfortran? How can i do this from setup.py command line or modification of numpy's site.cfg? An other strange thing i see is that prior to launching my build script, i am in a console with environment to find x64 verion of VS2015 build tools. However, the cl command lines use the x86 version of the compiler to cross compile for x64 (...\VC\BIN\x86_amd64\cl.exe). Where does that come from? Numpy distutils?