xnnpack in tensorflow lite for imx8 plus

[dnozik]

Hi ,

I working on  Imx8 plus board. (zeus-5.4.70-2.3.0).

I trying to execute my models with Tensorflow Lite (version  imx_5.4.70_2.3.0 ) via XNNPack delegate. looks like in Makefile there is default value  CXXFLAGS += -DTFLITE_WITHOUT_XNNPACK

Is there a reason XNNPack disabled?

[Bio_TICFSL]

Hello dnozik,

The XNNPack is disable in the actual version of tensorflow lite, however with the TensorFlow 2.3 release, XNNPACK backend is included with the pre-built TensorFlow Lite binaries for Linux and Android BSP. 

Regards

[dnozik]

Hi , 

thanks for reply .

1. Where can I find tensorflow lite prebuild library.

2.Does this prebuild library will also support IMX8 plus NPU delegate ?

3.We building all our libraries as part of image yocto recipe build , is there recipe with tensorflow branch with NNAPI and XNNPack for NPU/CPU ?

thanks

[Bio_TICFSL · Accepted solution]

Hello,

For TFLITE_WITHOUT_XNNPACK, it's disabled because not built with Makefile as below.

commit b9b739285e46ca0bdc54641bf7351463cf8aa928
Author: Terry Heo <terryheo@google.com>
Date:   Thu Mar 19 05:03:41 2020 -0700

    Fix Makefile build

    Since XNNPACK isn't buildable with Makefile, it should be disabled for Makefile build.

For 1# and 3#, BSP 5.10.9_1.0.0 is embedded with Tensorflow v2.4 and XNNPACK is enabled by default.

As for 2#, from 5.10.9 user guide (IMXMLUG), XNNPACK goes through CPU but not NPU.

Regards

[Bio_TICFSL · Accepted solution]

Hello,

For TFLITE_WITHOUT_XNNPACK, it's disabled because not built with Makefile as below.

commit b9b739285e46ca0bdc54641bf7351463cf8aa928
Author: Terry Heo <terryheo@google.com>
Date:   Thu Mar 19 05:03:41 2020 -0700

    Fix Makefile build

    Since XNNPACK isn't buildable with Makefile, it should be disabled for Makefile build.

For 1# and 3#, BSP 5.10.9_1.0.0 is embedded with Tensorflow v2.4 and XNNPACK is enabled by default.

As for 2#, from 5.10.9 user guide (IMXMLUG), XNNPACK goes through CPU but not NPU.

Regards