Current default build is MKL

[Craigacp]

This is painful for machines which don't have MKL installed. The MKL version bound against is the one using Intel's OpenMP, which is even less likely to be available on a machine (plus it's MKL-ML, not regular MKL nor MKL DNN - now DNNL).

We should make the build system a little more configurable, or at least set the default build to be something that works everywhere (even if the performance is worse).

I keep running into this when testing things on different machines here, none of which have MKL, and I have to get approvals for new inbound libraries.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux
• TensorFlow installed from (source or binary): source
• TensorFlow version: 2.0

[karllessard]

@saudet, any suggestion for this?

[saudet]

Sorry for the delay. GitHub notifications don't seem to work by default here for some reason.

MKL gets downloaded automatically by the Bazel build. I'm not sure I understand what the issue is. @Craigacp What's the exact error message that you're getting?

We can obviously add flags and stuff, that's fine.

[karllessard]

Is libmklml supposed to be packaged and distributed in our native artifacts?

After running mvn package, I see the following in my target repository:

$ ls target/native/org/tensorflow/internal/c_api/macosx-x86_64/lib
libiomp5.dylib                   libjnitensorflow.dylib           libmklml.dylib                   libtensorflow.2.dylib

But when listing the content of my native jar (macosx-x86_64 in my case), libiomp5 and libmklml have been discarded:

$ jar tvf target/tensorflow-core-api-0.1.0-SNAPSHOT-macosx-x86_64.jar 
    99 Sun Dec 01 21:30:42 EST 2019 META-INF/MANIFEST.MF
     0 Sun Dec 01 21:30:42 EST 2019 META-INF/
     0 Sun Dec 01 21:30:42 EST 2019 org/
     0 Sun Dec 01 21:30:42 EST 2019 org/tensorflow/
     0 Sun Dec 01 21:30:42 EST 2019 org/tensorflow/internal/
     0 Sun Dec 01 21:30:42 EST 2019 org/tensorflow/internal/c_api/
     0 Sun Dec 01 21:30:24 EST 2019 org/tensorflow/internal/c_api/macosx-x86_64/
     0 Sun Dec 01 21:30:42 EST 2019 META-INF/maven/
     0 Sun Dec 01 21:30:42 EST 2019 META-INF/maven/org.tensorflow/
     0 Sun Dec 01 21:30:42 EST 2019 META-INF/maven/org.tensorflow/tensorflow-core-api/
702664 Sun Dec 01 21:30:22 EST 2019 org/tensorflow/internal/c_api/macosx-x86_64/libjnitensorflow.dylib
218643976 Sun Dec 01 21:30:26 EST 2019 org/tensorflow/internal/c_api/macosx-x86_64/libtensorflow.2.dylib
 12792 Sun Dec 01 21:07:20 EST 2019 META-INF/maven/org.tensorflow/tensorflow-core-api/pom.xml
   108 Wed Nov 06 22:59:04 EST 2019 META-INF/maven/org.tensorflow/tensorflow-core-api/pom.properties

In addition, FWIW, I have this warning on my side when compiling:

[INFO] Copying /Users/klessard/Documents/Projects/MachineLearning/Sources/tensorflow-java/tensorflow-core/tensorflow-core-api/bazel-tensorflow-core-api/external/mkl_darwin/lib/libiomp5.dylib
[INFO] Copying /Users/klessard/Documents/Projects/MachineLearning/Sources/tensorflow-java/tensorflow-core/tensorflow-core-api/bazel-tensorflow-core-api/external/mkl_darwin/lib/libmklml.dylib
[WARNING] Could not find library mklml_intel

[saudet]

It gets extracted from the org.bytedeco package, but this is going away with newer versions of MKL-DNN/DNNL, so we don't need to worry about that.

[karllessard]

Any progress on this issue?

If I build our repo, I get this message as well when running a program:

Caused by: java.lang.UnsatisfiedLinkError: /Users/klessard/.javacpp/cache/sample-suite-1.0-SNAPSHOT-jar-with-dependencies.jar/org/tensorflow/internal/c_api/macosx-x86_64/libjnitensorflow.dylib: dlopen(/Users/klessard/.javacpp/cache/sample-suite-1.0-SNAPSHOT-jar-with-dependencies.jar/org/tensorflow/internal/c_api/macosx-x86_64/libjnitensorflow.dylib, 1): Library not loaded: @rpath/libmklml.dylib
  Referenced from: /Users/klessard/.javacpp/cache/sample-suite-1.0-SNAPSHOT-jar-with-dependencies.jar/org/tensorflow/internal/c_api/macosx-x86_64/./libtensorflow.2.dylib
  Reason: image not found
	at java.base/java.lang.ClassLoader$NativeLibrary.load0(Native Method)
	at java.base/java.lang.ClassLoader$NativeLibrary.load(ClassLoader.java:2430)
	at java.base/java.lang.ClassLoader$NativeLibrary.loadLibrary(ClassLoader.java:2487)
	at java.base/java.lang.ClassLoader.loadLibrary0(ClassLoader.java:2684)
	at java.base/java.lang.ClassLoader.loadLibrary(ClassLoader.java:2617)
	at java.base/java.lang.Runtime.load0(Runtime.java:767)
	at java.base/java.lang.System.load(System.java:1831)
	at org.bytedeco.javacpp.Loader.loadLibrary(Loader.java:1457)

Should/could we distribute libmklml ourself? If not, can we easily turn the --config=mkl off at compile time (maybe using a maven profile)?

[saudet]

Like I said, it is already distributed with the org.bytedeco artifact. This looks like some loading issue on Mac. Including it in the artifact for TensorFlow is going to end up with the same error. I'll look into that later, but like I said, this isn't something that Intel is maintaining anymore anyway, so let's not worry about it. I'll create a patch to disable this config by default, and include it in an extension for JavaCPP instead that the user needs to enable explicitly.

[Craigacp]

I do get this issue on Linux, it requires me to have MKL & Intel's OpenMP installed on the library path. After NeurIPS I can clean up a machine and try again to get the exact error message.

[saudet]

Ah, I think I understand what the issue is. You guys are probably adding dependencies on tensorflow-core instead of tensorflow-core-platform. The former won't add a dependency on MKLML, that's intentional, for users that know they have it on their system. Casual users should use tensorflow-core-platform which will get everything needed. But that's going to move to tensorflow-core-platform-mkl and such.

[zhangruiskyline]

Hi, I would like also explore the Tensorflow Java JNI lib with MLK DNN, can I know which Tensorflow version already have MLK-DNN built in? @saudet , can I know which release have your change? Thanks

[saudet]

Add a dependency on tensorflow-core-platform-mkl, that will pull in those binaries.

[zhangruiskyline]

Thanks @saudet

we already have tensorflow and libtensorflow_jni in our POM.xml, should we just add tensorflow-core-platform-mkl or replace some other dependency?

       <dependency>
            <groupId>org.tensorflow</groupId>
            <artifactId>tensorflow</artifactId>
            <version>${tensorflow.version}</version>
        </dependency>

        <dependency>
            <groupId>org.tensorflow</groupId>
            <artifactId>libtensorflow_jni</artifactId>
            <version>${tensorflow.version}</version>
        </dependency>

Also we have tensorflow version 1.13.1, is it compatible with tensorflow-core-platform-mkl?

Thanks

[saudet]

No, it's not compatible. You'll need to replace all that.

[zhangruiskyline]

Thanks @saudet , can I have a rough idea how to replace all? do I need to have all dependency in https://github.com/tensorflow/java/tree/master/tensorflow-core/tensorflow-core-platform-mkl? or how can I work out some solution? new to Java TF so would like to have some advice. Thanks