NNAPI

Supported platforms

NNAPI is supported only on i350 and i500 soc family.

Introduction

Google has ported NNAPI from Android to their ChromeOS (NNAPI on ChromeOS). We basically adapted their work on yocto.

Currently NNAPI on Linux, supports only one HAL that needs to be built at compile time. We support three HALs : Cpu HAL, ArmNN HAL and Xtensa HAL (VP6).

A HAL is a dynamically shared library named libvendor-nn-hal.so.

Available HAL

Cpu HAL

This is a very basic HAL that declares to handle only a very limited set of operations and does them on the CPU. No acceleration. This is used mainly for integration testing.

ArmNN HAL

This is the HAL to use the ArmNN framework that can leverage the GPU and Neon acceleration.

Xtensa HAL (VP6)

This is the HAL to use the VP6 APU from Cadence.

Note

A licence agreement with Cadence is needed in order to use the Xtensa HAL.

Building and adding the package

If building an image for a supported platform, NNAPI will by default be included in the image. To ensure it is included, please ensure that the package packagegroup-rity-nnapi is installed.

The HAL to be compiled is specified using a virtual provider. You can specify it like this:

nnapi-cpu-hal (Cpu HAL) - default one

android-nn-driver (ArmNN HAL)

xtensa-ann-bin (Xtensa HAL)

Inside the build/conf/local.conf file:

PREFERRED_PROVIDER_virtual/libvendor-nn-hal = "android-nn-driver"

Using NNAPI

NNAPI is used as a tensorflow-lite delegate. To use it, please refer to the Tensorflow-lite page.

Testing NNAPI

NNAPI provides the Compatibility Test Suite (CTS) to verify the correct functioning of a vendor HAL via the NNAPI public interface (ANeuralNetworks_*).

IMAGE_INSTALL_append = " nnapi-cts "

Run CTS

To run the cts test suite, make sure it is installed. It should be installed by default. If not you can install with the following inside your local.conf:

IMAGE_INSTALL_append = " nnapi-cts "

On the target, run the following

ulimit -S -n 5120
cd /usr/bin/
./nnapi-cts --gtest_output=json:/home/root/

Note

Before launching CTS, please increase per process open files limit on Linux by command: ulimit -S -n 5120. If using default open files limit (1024), you might get the error: Failed to clone native_handle in hidl_handle when running CTS.