Age | Commit message (Collapse) | Author |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
libibverbs was merged into rdma-core.
References
----------
https://lists.archlinux.org/pipermail/aur-requests/2018-October/027305.html
|
|
|
|
|
|
nanopb submodule was removed by upstream.
References
----------
https://github.com/pytorch/pytorch/commit/c172ffb6321db128fb42f77554869f66406c06c4
|
|
This package now provides the non-cuda version (as known as the
'cpy only' build).
If you want caffe2 with cuda support, use package caffe2-cuda-git.
This new package naming schema will better reflect the package
contents. This conforms to tensorflow package naming from the
official repositories.
|
|
|
|
References
----------
https://github.com/pytorch/pytorch/commit/05c473b85cf40908ffefb16351bcc49ac582968f
|
|
cuda 9.2 compiler is gcc7, but compilation fails with it. It works
with gcc6 but it's not on the official repositories.
gcc54 is on the official repositories, so let's use it.
|
|
|
|
gcc54 is now a dependency for cuda, sot it will be already
pulled. Letting gcc5 (as it was until now) will cause a
package conflict with gcc54.
References
----------
https://git.archlinux.org/svntogit/community.git/commit/trunk?h=packages/cuda&id=2b0be9b7390fa897edb6e4f44224eaf8619399c9
|
|
|
|
Update git submodules to match latest upstream changes.
Note that 'third_party/nccl' is not a real git submodule
and thus it does not need to be cloned by makepkg (nccl
source code currently comes bundled in 'third_party/nccl'
when cloning the pytorch git repository).
$pkgver now uses the caffe2 version from the file
'caffe2/VERSION_NUMBER'. Since this version does
not coincide with the pytorch version mumber, the
revision number is now the total number of revisions
in the pytorch git repository.
Note
----
gcc5 is currently on the [community] repository. It was
moved from the AUR to [community] due to CUDA. Removed
gcc5 from the AUR block on makedepends.
|
|
Fix build: use gcc5
-------------------
Unfortunatelly, the current git master branch of caffe2
does not build with the current cuda 9.1 and gcc6 from the
official repositories. This prevented the package to be built
for a long time.
Details here:
https://github.com/caffe2/caffe2/issues/1459
https://github.com/caffe2/caffe2/issues/1636
But it is possible to build with gcc5 from the AUR. Be warned
that gcc5 from the AUR takes a lot of time to compile.
Switch from python2 to python3
------------------------------
Upstream now provides python3 support at git master branch.
Change source to pytorch repository to match upstream changes
-------------------------------------------------------------
Caffe2 source code moved to PyTorch repository.
References
----------
https://github.com/caffe2/caffe2/issues/315#issuecomment-350974915
https://github.com/caffe2/caffe2/commit/a062b05d70c75a6005c6e76109a1dde482729a53#diff-04c6e90faac2675aa89e2176d2eec7d8
|
|
References
----------
https://github.com/caffe2/caffe2/commit/acd1aef483c168e929ca56b2a6fe958ecb40e46b
|
|
References
----------
https://github.com/caffe2/caffe2/commit/01827c153db96349745a544148c1ff0386c5ef9e
|
|
|
|
The newly released NCCL version 2 requires registration at the
NVIDIA developer webiste and a manual download. As this breaks
automated builds by AUR helpers, NCCL was disabled by default in
this package. It can still be enabled by modifying the proper
lines in the PKGBUILD file (instructions are at the top of the
PKGBUILD file).
|
|
|
|
Also on this commit:
- explicitly added the options 'USE_METAL' and
'USE_MOBILE_OPENGL', setting both of them to 'OFF'.
|
|
Adding python-future fixes an error when importing the caffe2
python modules. Thanks to AUR user wangqr for pointing this.
|
|
|
|
Caffe2 currently does not compile with the newly released cudnn 7.0.
It builds fine with cudnn 6.0.21.
https://github.com/caffe2/caffe2/issues/1025
|
|
|
|
https://github.com/caffe2/caffe2/commit/f070eb99a4a36f12abbdb9af98bb68ff485c840b
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Currently disabled. Image and video processing seems to be done
only when enabling opencv.
|
|
Compiling with opencv currently is giving linking errors
with libtbb.so (intel-tbb, enabled in opencv).
Temporarily disabled opencv until it is fixed.
|
|
|
|
|
|
|
|
cudnn has been moved from AUR to [community].
|
|
cudnn already pulls cuda. The current version of cuda (8.0) already
pulls gcc5. But strictly adding them the dependencies can avoid
some problems in the future.
|
|
|
|
|