On Mon, 2022-08-08 at 18:02 +0200, Böszörményi Zoltán wrote: > Hi, > > I am trying to create a recipe for pytorch > (https://github.com/pytorch/pytorch) > with GPU accceleration but it seems I ran into a few issues. > > The intel-oneapi-mkl recipe simply repackages the contents of a DEB > package. > > There are some problems with this as this recipe cannot be used to > build > others that would link to e.g. libmkl_intel_lp64.so > > * the contents are installed into /opt so sysroot-destdir is empty > * if I add /opt to SYSROOT_DIRS, then the libraries are carried > over intel-oneapi-mkl into recipe-sysroot but they are > in /opt and this directory is not used by default to find > libraries
I think it was originally done because the oneAPI installer installs everything in /opt right now and it also installs libraries which may or may not already be installed in the system in standard locations (like tbb, opencv). I looked at mkl and it doesn't look like it installs any such library so we can probably change at least mkl to have it installed properly. > * even if I convince pytorch to look for the library in the > correct location, it stil looks for mkl_cblas.h which is > not present in the intel-oneapi-mkl package. It seems this header is in intel-oneapi-mkl-common-devel so that might also need to be fetched and installed. Can you please check if this helps? Thanks, Anuj
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#7610): https://lists.yoctoproject.org/g/meta-intel/message/7610 Mute This Topic: https://lists.yoctoproject.org/mt/92895333/21656 Group Owner: [email protected] Unsubscribe: https://lists.yoctoproject.org/g/meta-intel/unsub [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
