TIM-VX - Tensor Interface Module

TIM-VX is a software integration module provided by VeriSilicon to facilitate deployment of Neural-Networks on Verisilicon ML accelerators. It serves as the backend binding for runtime frameworks such as Android NN, Tensorflow-Lite, MLIR, TVM and more.

Main Features

• Over 150 operators with rich format support for both quantized and floating point
• Simplified C++ binding API calls to create Tensors and Operations Guide
• Dynamic graph construction with support for shape inference and layout inference
• Built-in custom layer extensions
• A set of utility functions for debugging

Framework Support

• Tensorflow-Lite (External Delegate)
• Tengine (Official)
• TVM (Fork)
• Paddle-Lite (Official)
• OpenCV (In development)
• MLIR Dialect (In development)

Feel free to raise a github issue if you wish to add TIM-VX for other frameworks.

Architecture Overview

Get started

Build and Run

TIM-VX supports both bazel and cmake.

---

cmake

To build TIM-VX for x86 with prebuilt:

mkdir host_build
cd host_build
cmake ..
make -j8
make install

All install files (both headers and *.so) is located in : host_build/install

cmake options:

option name	Summary	Default
TIM_VX_ENABLE_TEST	Enable unit test case for public APIs and ops	OFF
TIM_VX_ENABLE_LAYOUT_INFER	Build with tensor data layout inference support	ON
TIM_VX_USE_EXTERNAL_OVXLIB	Replace internal with a prebuilt libovxlib library	OFF
OVXLIB_LIB	full path to libovxlib.so include so name, required if TIM_VX_USE_EXTERNAL_OVXLIB=ON	Not set
OVXLIB_INC	ovxlib's include path, required if TIM_VX_USE_EXTERNAL_OVXLIB=ON	Not set
EXTERNAL_VIV_SDK	Give external vivante openvx driver libraries	Not set
TIM_VX_BUILD_EXAMPLES	Build example applications	OFF
TIM_VX_ENABLE_40BIT	Enable large memory (over 4G) support in NPU driver	OFF

---

run unit test:

cd host_build/src/tim

export LD_LIBRARY_PATH=`pwd`/../../../prebuilt-sdk/x86_64_linux/lib:<path to libgtest_main.so>:$LD_LIBRARY_PATH
export VIVANTE_SDK_DIR=`pwd`/../../../prebuilt-sdk/x86_64_linux/lib
export VSIMULATOR_CONFIG=<hardware name should get from chip vendor>
# if you want to debug wit gdb, please set
export DISABLE_IDE_DEBUG=1
./unit_test

Build with local google test source

    cd <wksp_root>
    git clone --depth 1 -b release-1.10.0 git@github.com:google/googletest.git

    cd <root_tim_vx>/build/
    cmake ../ -DTIM_VX_ENABLE_TEST=ON -DFETCHCONTENT_SOURCE_DIR_GOOGLETEST=<wksp_root/googletest> <add other cmake define here>

---

Build for your evk-board

1. prepare toolchain file follow cmake standard
2. make sure cross build low-level driver with toolchain separately, we need the sdk from the low-level driver
3. add -DEXTERNAL_VIV_SDK=<low-level-driver/out/sdk> to cmake definitions, also remember -DCMAKE_TOOLCHAIN_FILE=<Your_Toolchain_Config>
4. then make

---

Important notice for integration

If you want to build tim-vx as a static library, and link it to your shared library or application, please be carefull with the linker, "-Wl,--whole-archive" is required.

@see samples/lenet/CMakeLists.txt for reference

Bazel

Install bazel 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.so

To 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_cc

Other

To build and run Tensorflow-Lite with TIM-VX, please see README

To build and run TVM with TIM-VX, please see TVM README

Reference board

Chip	Vendor	References
i.MX 8M Plus	NXP	ML Guide, BSP
A311D	Khadas - VIM3	A311D datasheet, BSP
S905D3	Khadas - VIM3L	S905D3 , BSP

Support

create issue on github or email to ML_Support@verisilicon.com