lcai-tihu / sw

LCAI-TIHU SW is a software stack of the AI inference processor based on RISC-V

LCAI-TIHU SOFTWARE

Introduction

TIHU is an AI computing development platform based on the RISC-V instruction set. It provides software stack and hardware and is a complete set of solutions that support different AI frameworks and AI algorithms. It can be directly used for AI processor development and verification. It can also be used to explore the current AI open sources ecology, such as RISC-V instruction set, deep-learning accelerator, AI compiler & runtime, AI algorithms, and frameworks.

Currently, TIHU only supports TensorFlow. Pytorch will be supported very soon. As shown below, several CNN models have been tested on the Inspur F37X card.

CNN models and accuracy:

num	category	CNN model	TIHU(int8) accuracy	llvm(float) accuracy	data set
1	classification	LeNet	0.9884	0.992	mnist
2	classification	ResNet50	0.755	0.757	imagenet
3	classification	MobileNetv2	0.71	0.72	imagenet
4	classification	VGG16	0.627	0.629	imagenet
5	detection	YOLOv3	0.548	0.562	coco
6	NLP	bert-base	-	-	squad_v1.1

At present, the following operators are supported. Intrinsic operators will be done.

category	operators
convolution	Conv, Transposed Conv, Dilated Conv, Spatially Separable Conv, Depthwise Separable Conv, Grouped Convolution, Flattened Conv
activation	ReLU, ReLU6, P-ReLU, LeakyReLU, Sigmoid
fusion	CBR(Conv + BN + Relu), CB(Conv + BN)
pool	Max Pooling, Average Pooling, Upsample
dense	matrix multiplication, FC
Eltwise	Elemet-wise
other	Softmax, BatchNorm, Add, sub, div, multiply, power, slice, reshape, concate, expf, tanh, mean, cast, transpose, take, one_hot, expand, dense, batch_matmul, squeeze, resize, reduce, split, pool2D, etc.

Example of yolov3:

Code structure

├── compiler_runtime: TIHU software, include compiler and runtime
├── doc: TIHU user's guide
├── firmware: TIHU SoC firmware
├── xdma: driver and firmware download tools
├── docker: docker file
├── README.md
└── LICENSE

TIHU software structure

TIHU software include compiler, runtime, xdma driver, and firmware.

Compiler and runtime

The compiler and runtime are designed based on TVM. The compiler consists of the parser, optimizer, codegen, and operator libraries. CNN models are compiled into computing tasks with different device types and submitted to the device by the runtime. The host interacts with the device through PCIe.

Driver

TIHU driver is designed based on Xilinx PCIe DMA driver, so as the firmware download program. More driver information can be found in the readme of the xdma.

Firmware

TIHU hardware is driven by firmware to complete computing tasks submitted by the runtime. More firmware information can be found in the readme of the firmware.

TIHU software workflow

Build procedure on ubuntu

1. Build compiler and runtime without dockfile
   Install docker and then:

$ sudo docker run -it --network=host -v /your/project/path/on/host:/workspace -v /dev:/dev --privileged=true --name your_docker_name docker_image_name:version /bin/bash # run docker and load ubuntu image
$ apt-get update
$ apt-get install -y python3 python3-dev python3-setuptools python3-pip python3-venv python-scipy gcc libtinfo-dev zlib1g-dev build-essential cmake libedit-dev libxml2-dev llvm-dev libjpeg-turbo8-dev git python3-pip autoconf
$ pip3 install -U pip -i https://pypi.tuna.tsinghua.edu.cn/simple
$ pip3 install numpy decorator attrs pytest scipy  opencv-python-headless tqdm pycocotools 
$ pip3 install keras==2.6.0 tensorflow==2.6.1  
$ git clone https://github.com/LCAI-TIHU/SW.git
$ cd SW/compiler_runtime && source env.sh
$ source build.sh

2. build compiler and runtime with dockefile
   We provide two versions of dockerfile, which you can choose by yourself, which will positively affect your build efficiency.

$ git clone https://github.com/LCAI-TIHU/SW.git
$ cd SW/docker && sudo docker build -t tihu_software:v1 -f TIHU.Dockerfile .
$ sudo docker run -it --network=host -v /your/SW/path/:/workspace -v /dev:/dev --privileged=true --name tihu_software tihu_software:v1 /bin/bash
$ cd /workspace && git clone https://github.com/LCAI-TIHU/SW.git
$ cd /workspace/SW/compiler_runtime
$ source env.sh && source build.sh 

Build FPGA

Go to TIHU-HW.

Quickstart

There are some samples in /SW/compiler_runtime/demo, before run you should:

• Make sure the FPGA bitstream has been load, you can refer to HW project.
• Make sure firmware has been compiled and downloaded, you can refer to ./firmware README.
• Run docker and build compiler and runtime, make sure env.sh has been sourced.

$ mkdir -p /workspace/mnist && cd /workspace/mnist
$ wget http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz && gunzip t10k-images-idx3-ubyte.gz
$ wget http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz && gunzip t10k-labels-idx1-ubyte.gz
$ mkdir -p /workspace/tfmodel && cd /workspace/tfmodel
$ cp /workspace/SW/compiler_runtime/demo/lenet_test/frozen_lenet_3.pb ./
$ mkdir -p /workspace/testimage && cd /workspace/testimage
$ cp /workspace/SW/compiler_runtime/demo/lenet_test/00000_7.jpg ./
$ cd /workspace/SW/compiler_runtime/demo
$ python3 from_tensorflow_quantize_lenet.py 2>&1 | tee lenet.log

The number in the picture is 7, and you get the inference that the 8th data is the largest.

TODO

A serial of optimation will be done.

• support multi-process.
• Optimize the operators and compiler to improve the computational efficiency.
• ...

Acknowledgements

TIHU is designed based on a variety of open-source projects, including cva6, ara, nvlda, xdma, and TVM. If you want to learn more about these models, check out their website.

License

TIHU is licensed under the Apache-2.0 license. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

sw's Issues

请问硬件有计划开源吗？

很有意思的工作，请问硬件什么时候开源？