3.9.1. Neo-AI Deep Learning Runtime

Introduction

Neo-AI-DLR is an open source common runtime for deep learning models and decision tree models compiled by TVM, AWS SageMaker Neo, or Treelite. Processor SDK Linux has integrated Neo-AI-DLR. DLR stands for Deep Learning Runtime. With this integration, the models compiled by AWS SageMaker Neo and TVM can run on all Arm core of all Sitara devices (AM3/AM4/AM5/AM6).

On A5729 and AM5749, TensorFlow models compiled by Neo, if supported by TIDL, are fully offloaded to EVE and DSP cores. If not, then they run on Arm core. In future release, a graph of any model supported by Neo compiler can be split into sub-graphs, where TIDL supported sub-graph will run on EVE/DSP cores, while unsupported layer sub-graph will run on Arm core.

Examples

Examples of running inference with Neo-AI-DLR are available in /usr/share/dlr of the target filesystem:

  • demos folder: contains examples of Neo-AI-DLR on top of TIDL, which can only run on AM5729/49 devices. To run the examples, a neural network model must be compiled to generate runtime artifacts according to instructions in Compiling Network Models. Once a model is compiled, copy the generated artifacts to this folder and run one of the examples, e.g:
cd /usr/share/dlr/demos
Copy artifact, e.g. scp -r <TVM_on_host_machine>/apps/tidl_deploy/output4/mobileNet2 .
Edit subgraph0.cfg and add the following lines at the end of the file:
    inConvType = 0
    inIsSigned = 1
    inScaleF2Q = 128.0
    inIsNCHW = 0
    outConvType = 0
    outIsSigned = 0
    outScaleF2Q = 255.0
    outIsNCHW = 1
./do_tidl4.sh mobileNet2

For more information about running examples of Neo-Ai-DLR with TIDL, please refer to Neo-ai-dlr Texas Instruments branch in github.

  • tests folder: contains examples of Neo runtime for ARM cores, which can run on any Sitara device. Go to folder python/integration and run the example (http_proxy needs to be set properly):
cd /usr/share/dlr/tests/python/integration/
python3 load_and_run_tvm_model.py

Compiling Network Models to Run with DLR

SageMaker Neo compiler can compile neural network models to run with DLR on Sitara devices. Neo compiler source code supporting Sitara devices is hosted at https://github.com/TexasInstruments/tvm and will be periodically upstreamed to https://github.com/neo-ai/tvm.

  • Build Neo Compiler from source
git clone --recursive https://github.com/TexasInstruments/tvm.git --branch dev
cd tvm/
mkdir build && cd build
cp ../cmake/config.cmake .
cmake ..
make –j8
  • Install Python Package

Follow instructions at https://docs.tvm.ai/install/from_source.html#python-package-installation for installation.

  • Compile Neural Network Models

Follow instructions at https://github.com/TexasInstruments/tvm/tree/dev/apps/tidl_deploy to compile neural network models. A simple example generating the artifacts needed to run the demo above is shown below:

python3 NeoTvmCodeGen.py mobileNet2

Note

  • Currently Neo compiler with Sitara support can compile any models supported by Neo, but only TensorFlow models can be compiled to run on TIDL for acceleration if the model can be supported by TIDL.

Benchmarking

Performance with and without TIDL offload is shown below for TensorFlow MobileNet v1 and v2. The performance depends significantly on batch size (if batch size is 1, only one EVE is operating and performance would be very poor).

Batch Size TIDL MobileNetV1 (fps) ARM MobileNetV1 (fps) TIDL MobileNetV2 (fps) ARM MobileNetV2 (fps)
4 30.1260 2.2018 30.5178 3.6843
16 34.8465 2.2210 36.0127 3.6775
32 35.5279   37.5482  

Note

  • This release only supports batch size up to 32.
  • There is no TVM auto-tuning for ARM (using default scheduling) and it is single A15 core execution.

Rebuilding DLR from Source

DLR for Sitara devices is included in Proc-SDK Linux target file system. Source code is hosted at https://github.com/TexasInstruments/neo-ai-dlr and will be periodically upstreamed to https://github.com/neo-ai/neo-ai-dlr. Users may rebuild the latest source code before official Proc-SDK release, following steps below:

  • Clone git repo on x86 host to target NFS (git cloning may not work on EVM):
git clone --recursive https://github.com/TexasInstruments/neo-ai-dlr.git --branch dev
  • Build and Install DLR on AM57x9 EVM:
cd neo-ai-dlr
mkdir build && cd build
cmake ..
make –j2
make demo democv
cd ../python
python3 setup.py install --user