Package Base Details: tensorflow-computecpp

Git Clone URL: (read-only, click to copy)
Keywords: computecpp opencl sycl tensorflow-opencl
Submitter: enihcam
Maintainer: mirh
Last Packager: mirh
Votes: 0
Popularity: 0.000000
First Submitted: 2018-04-25 12:51 (UTC)
Last Updated: 2023-05-02 13:56 (UTC)

Latest Comments

1 2 Next › Last »

mirh commented on 2023-05-02 01:21 (UTC)

After much painstaking python bisecting, I found out this was the culprit..

Yet I don't really have clues about how to proceed from here.

mirh commented on 2023-04-27 23:12 (UTC) (edited on 2023-04-27 23:13 (UTC) by mirh)

So.. after more (numpy and python) or less (the whitespace error) ado I managed to sort of revive the package. Too bad that for reasons, python 3.10 is giving me this last woe.

Traceback (most recent call last):
  File "//tensorflow/tools/api/generator/", line 27, in <module>
    from tensorflow.python.util import tf_decorator
  File "//tensorflow/python/", line 63, in <module>
    from tensorflow.python.framework.framework_lib import *  # pylint: disable=redefined-builtin
  File "//tensorflow/python/framework/", line 52, in <module>
    from tensorflow.python.framework.importer import import_graph_def
  File "//tensorflow/python/framework/", line 27, in <module>
    from tensorflow.python.framework import function
  File "//tensorflow/python/framework/", line 36, in <module>
    from tensorflow.python.ops import resource_variable_ops
  File "//tensorflow/python/ops/", line 37, in <module>
    from tensorflow.python.ops import math_ops
  File "//tensorflow/python/ops/", line 1130, in <module>
    _OverrideBinaryOperatorHelper(gen_math_ops.logical_or, "or")
  File "//tensorflow/python/ops/", line 887, in _OverrideBinaryOperatorHelper
    clazz_object._override_operator("__%s__" % op_name, binary_op_wrapper)
  File "//tensorflow/python/framework/", line 644, in _override_operator
    _override_helper(Tensor, operator, func)
  File "//tensorflow/python/framework/", line 101, in _override_helper
    raise ValueError("operator %s cannot be overwritten again on class %s." %
ValueError: operator __or__ cannot be overwritten again on class <class 'tensorflow.python.framework.ops.Tensor'>.

Eirikr commented on 2023-04-03 15:21 (UTC) (edited on 2023-04-03 16:08 (UTC) by Eirikr)

Unable to install. The Bazel patch does not apply anymore is what I am guessing? The fix below (to remove the space at line 12) is no longer valid. There is no space at line 12.

Removing the space between line 12 and line 13 results in an "invalid patch" error.

Used makepkg -si --skipcheck otherwise had a checksum error from editing the gcc2.patch file.

Any other tips would be appreciated!

-> Extracting bazel-0.17.2-1-x86_64.pkg.tar.xz with bsdtar
==> Starting prepare()...
error: patch failed: tensorflow/workspace.bzl:397
error: tensorflow/workspace.bzl: patch does not apply
==> ERROR: A failure occurred in prepare().
 -> error making: tensorflow-computecpp (python-tensorflow-computecpp tensorflow-computecpp)

mirh commented on 2020-11-30 10:03 (UTC)

@alkis05 you are free to commit the fixes I don't have that much use for this anymore

alkis05 commented on 2020-11-29 10:29 (UTC) (edited on 2020-11-29 19:19 (UTC) by alkis05)

@grdgkjrpdihe I have the same error, but there is no space left at the end of gcc2.diff:12. What worked for me was to just comment out the PKGBUILD line that apply this patch.

Also, there was a change to numpy that made the build to fail (at the very end of a very long build, goddamn it). I complains about not being able to change a varialble into a const. It is related to issues #40688, #40654. The solution was to use const inputs in a specific tensorflow function and create a function overload for the new signature. The patch is this one:

If I'm able to build, I will make a push request for the patch to be included in the package, if that is possible. This thing takes forever to build from source.

EDIT: Success! with those alterations version 1.19 compiles well and I can confirm that my crappy intel HD 5500 (broadwell) is being used (for all it is worth -.-), thanks to SYCL. Next time I go to sleep, I will try to compile the latest release (v2.1.2) overnight and see how it goes. If all goes well, maybe we could update this package. Or create a tensorflow2 or something, if there is the need for keeping a 1.x and 2.x versions available.

grdgkjrpdihe commented on 2020-08-18 15:08 (UTC) (edited on 2020-08-21 14:50 (UTC) by grdgkjrpdihe)

prepare failed

==> Starting prepare()...
error: patch failed: tensorflow/workspace.bzl:397
error: tensorflow/workspace.bzl: patch does not apply
==> ERROR: A failure occurred in prepare().

edit: removing the space at the end of gcc2.diff on line 12 solved the problem

mirh commented on 2019-02-01 20:16 (UTC) (edited on 2019-08-13 21:33 (UTC) by mirh)

¿It runs on Intel hardware you know? What's the matter? I mean, at least Neo driver was reported working/fixed last year. (also, I know some random fix was pushed to beignet in the meantime, so who knows on <Gen8 hardware)

Then if it's just a matter of time instead.. Well, I'm in no really better position.

EDIT: before any update to this, I'd like upstream to add support for python 3.7. I'd loathe having to make patches myself to backport it. In the meantime you can follow the instructions below

enihcam commented on 2019-01-31 02:36 (UTC)

@mirh. I agree but now I only have an intel laptop. I don't have resource to maintain (test) this package. Can I disown this package? :(

mirh commented on 2019-01-31 02:20 (UTC)

It's the only thing that would work on "generic whatever" ARM hardware though (I mean, yes we aren't currently providing the package for that, but still)

It's working fine here once you downgrade bazel to 0.17, python to 3.6 (and/or relative libraries, I still haven't precisely settled down the whole thing) and use Rbiessy branch. Also, mercurial has to be installed otherwise cmake 3.13 complains.

Aside of that, you are good.

enihcam commented on 2019-01-21 06:28 (UTC)

@ModYokosuka Sorry I'm going to deprecate this package because of various build issues, also the performance of computecpp approach is lower than tensorflow-mkl(intel)/tensorflow-rocm(amd)/tensorflow-cuda(nvidia).