TIM-VX is a software integration module provided by VeriSilicon to facilitate deployment of Neural-Networks on OpenVX enabled ML accelerators. It serves as the backend binding for runtime frameworks such as Android NN, Tensorflow-Lite, MLIR, TVM and more.
Main Features
- Over 130 operators with rich format support for both quantized and floating point
- Simplified C++ binding API calls to create Tensors and Operations
- Dynamic graph construction with support for shape inference and layout inference
- Built-in custom layer extensions
- A set of utility functions for debugging
- Tensorflow-Lite Delegate (Unofficial)
- Tengine (Official)
- MLIR Dialect (In development)
- TVM (In development)
Roadmap of TIM-VX will be updated here in the future.
TIM-VX uses bazel build system by default. Install bazel first to get started.
TIM-VX needs to be compiled and linked against VeriSilicon OpenVX SDK which provides related header files and pre-compiled libraries. A default linux-x86_64 SDK is provided which contains the simulation environment on PC. Platform specific SDKs can be obtained from respective SoC vendors.
To build TIM-VX
bazel build libtim-vx.soTo run sample LeNet
# set VIVANTE_SDK_DIR for runtime compilation environment
export VIVANTE_SDK_DIR=`pwd`/prebuilt-sdk/x86_64_linux
bazel build //samples/lenet:lenet_asymu8_cc
bazel run //samples/lenet:lenet_asymu8_ccTo build and run Tensorflow-Lite delegate on A311D platform
# clone and cross build VeriSilicon tensorflow fork with TFlite delegate support
git clone --single-branch --branch vx-delegate.v2.4.1 [email protected]:VeriSilicon/tensorflow.git vx-delegate; cd vx-delegate
bazel build --config A311D //tensorflow/lite/tools/benchmark:benchmark_model
# push benchmark_model onto device and run
./benchmark_model --graph=mobilenet_v1_1.0_224_quant.tflite --use_vxdelegate=true