Missing dnnl_config.h when installing horovod

[sxjscience]

Description

I tried to compile MXNet using the cmake command:

cmake -GNinja -C ../config.cmake ..
cmake --build .

and then install horovod

HOROVOD_WITHOUT_TENSORFLOW=1 HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL pip3 install -e . --no-cache-dir --user

The error message is

/home/ubuntu/mxnet/python/mxnet/../../include/mkldnn/dnnl.hpp:23:10: fatal error: dnnl_config.h: No such file or directory

I find that we have these lines in mkldnn.mk https://github.com/apache/incubator-mxnet/blob/c2aff58c448bbe5755de0c8d35a1cc7d3fd9aa51/mkldnn.mk#L52-L53

The problem might be that we haven't copied the files to mxnet/include/mkldnn so the build system used in horovod is broken.

[sxjscience]

@leezu @apeforest Have you met problems when building horovod with cmake-built MXNet?

[sxjscience]

A temporary fix is to manually copy files inside the build directory to include/mkldnn. I've confirmed that this solves the problem in horovod installation.

cp build/3rdparty/mkldnn/include/dnnl_config.h include/mkldnn
cp build/3rdparty/mkldnn/include/dnnl_version.h include/mkldnn

[TaoLv]

I will take a look at this. I remember previously the recommended build system of mxnet was makefile when cooperating with horovod.
@sxjscience is this the only issue when using cmake to build mxnet and linking it to horovod?

[sxjscience]

That’s currently the only one I found for linking to horovod. However I just ran setup.py in horovod and hasn’t run distributed training yet. I’m having some other issues in the cmake-built MXNet which I’m still investigating. Get Outlook for iOS<https://aka.ms/o0ukef>

…

________________________________ From: Tao Lv <notifications@github.com> Sent: Wednesday, February 19, 2020 5:15:05 PM To: apache/incubator-mxnet <incubator-mxnet@noreply.github.com> Cc: Xingjian SHI <xshiab@connect.ust.hk>; Mention <mention@noreply.github.com> Subject: Re: [apache/incubator-mxnet] Missing dnnl_config.h when installing horovod (#17628) I will take a look at this. I remember previously the recommended build system of mxnet was makefile when cooperating with horovod. @sxjscience<https://github.com/sxjscience> is this the only issue when using cmake to build mxnet and linking it to horovod? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#17628>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ABHQH3VFOAX6T6T4WKRZDY3RDXKRTANCNFSM4KYACRAA>.

[leezu]

@TaoLv what's your status on this issue?
I think this issue is currently blocking Makefile build removal so we should fix it soon. (You mentioned you plan to work on it, so please let us know if you don't plan to work on it soon.)

[TaoLv]

@leezu I‘m working on a fix.