CHaiDNN-v2
Analysis and Eval
Supported Layers	Performance/Resource Utilization
Performance Eval
Design and Development
API Reference	Quantization User Guide for CHaiDNN	Model Zoo	Running Inference on new Network
Creating SDx GUI Project	Configurable Parameters	Custom Platform Generation	Software Layer Plugin
SDSoC Environment User Guide	Hardware-Software Partitioning for Performance

Introduction

CHaiDNN is a Xilinx Deep Neural Network library for acceleration of deep neural networks on Xilinx UltraScale MPSoCs. It is designed for maximum compute efficiency at 6-bit integer data type. It also supports 8-bit integer data type.

The design goal of CHaiDNN is to achieve best accuracy with maximum performance. The inference on CHaiDNN works in fixed point domain for better performance. All the feature maps and trained parameters are converted from single precision to fixed point based on the precision parameters specified by the user. The precision parameters can vary a lot depending upon the network, datasets, or even across layers in the same network. Accuracy of a network depends on the precision parameters used to represent the feature maps and trained parameters. Well-crafted precision parameters are expected to give accuracy similar to accuracy obtained from a single precision model.

What's new in CHaiDNN-v2

• 4x GOPS compared to CHaiDNN-v1 (2017.4) (Performance numbers)

• 2x MAC on DSPs at int6

• Double-Pumped DSPs allowing the DSPs to be clocked at twice the core clock (Some configs can go upto 350/700Mhz)

• Introducing DietChai - A miniature version of CHai for smaller MPSoC/ Zynq devices

• 128, 256, 512, 1024 DSP design configs verified for ZU9

• Support for URAM

• 128, 256, 512 DSP configs verified for ZU7

• ModelZoo of 6 networks at int8 and int6 precision

• Support for two quantization modes - Dynamic fixed point and Xilinx Quantizer

• Enhanced API to enable better hardware- software partitioning for users

• Support for software custom layer plug-ins

• Fully Connected layers on CPU

• More documentation

Performance Benchmarks(fps)

^{* Source: https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/}

Hardware and Software Requirements

The CHaiDNN library is designed to work with Zynq UltraScale+ MPSoCs. The library has been verified on zcu102 and zcu104 boards. Xilinx SDSoC 2018.2 Development Environment is required to work with the library.

How to Download the Repository

To get a local copy of the CHaiDNN repository, configure git-lfs and then, clone this repository to the local system with the following command:

git clone https://github.com/Xilinx/CHaiDNN.git CHaiDNN

Where CHaiDNN is the name of the directory where the repository will be stored on the local system. This command needs to be executed only once to retrieve the latest version of the CHaiDNN library.

GitHub Repository Structure

CHaiDNN/
|
|-- CONTRIBUTING.md
|-- LICENSE
|-- README.md
|-- SD_Card
|   |-- lib
|   |-- cblas
|   |-- images
|   |-- opencv
|   |-- protobuf
|   |-- zcu102
|   `-- zcu104
|-- design
|   |-- build
|   |-- conv
|   |-- deconv
|   |-- pool
|   `-- wrapper
|-- docs
|   |-- API.md
|   |-- BUILD_USING_SDX_GUI.md
|   |-- CONFIGURABLE_PARAMS.md
|   |-- CUSTOM_PLATFORM_GEN.md
|   |-- HW_SW_PARTITIONING.md
|   |-- MODELZOO.md
|   |-- PERFORMANCE_SNAPSHOT.md
|   |-- QUANTIZATION.md
|   |-- RUN_NEW_NETWORK.md
|   |-- SOFTWARE_LAYER_PLUGIN.md
|   |-- SUPPORTED_LAYERS.md
|   `-- images
|-- software
|   |-- bufmgmt
|   |-- checkers
|   |-- common
|   |-- custom
|   |-- example
|   |-- imageread
|   |-- include
|   |-- init
|   |-- interface
|   |-- scheduler
|   |-- scripts
|   |-- swkernels
|   `-- xtract
`-- tools
    |-- SETUP_TOOLS.md
    `-- tools.zip

Run Inference

Using Pre-built binaries

To run inference on example networks, follow these steps: 1. Download the example network 6-bit GoogleNet with Xilinx Quantization scheme. More networks are available as part of the ModelZoo. 1. Place the downloaded and unzipped contents at "SD_Card/models" directory. Create SD_Card/models directory if not present already.

1. Copy the required contents of "SD_Card" folder into a SD-Card.
   • opencv
   • protobuf
   • cblas
   • images
   • bit-stream, boot loader, lib & executables (either from SD_Card/zcu102 or SD_Card/zcu104)

2. Insert the SD-Card and power ON the board.

   :pushpin: NOTE: A serial port emulator (Teraterm/Minicom) is required to interface the user commands to the board

3. Attach a USB-UART cable from the board to the host PC. Set the UART serial port to Baud rate: 115200 Data: 8 bit Parity: none Stop: 1 bit Flow control: none

4. After boot sequence, set LD_LIBRARY_PATH env variable. sh export OPENBLAS_NUM_THREADS=2 export LD_LIBRARY_PATH=lib/:opencv/arm64/lib/:protobuf/arm64/lib:cblas/arm64/lib
5. Create a folder "out" inside the network directory to save the outputs sh cd /mnt mkdir models/<network>/out

6. Execute "*.elf" file to run inference

   • The format for running these example networks is described below: sh ./<example network>.elf <quantization scheme> <bit width> <img1_path> <img2_path>
   • For GoogleNet 6-bit inference with Xilinx quantization scheme execute the following sh ./googlenet.elf Xilinx 6 images/camel.jpg images/goldfish.JPEG
   • Sync after execution
     sh cd / sync umount /mnt
   • Output will be written into text file inside respective output folders. Ex : models/<network>/out

     :pushpin: NOTE: Failing to run sync might corrupt the file system and cause crash on subsequent runs.

:pushpin: NOTE: For running inference on a new network, please follow the instructions in Run new Network using CHaiDNN.

Build from Source

CHaiDNN can be built using Makefiles OR using SDx IDE. The below steps describe how to build CHaiDNN using Makefiles. For steps to build using SDx IDE, see the instructions in Build using SDx IDE.

Build CHaiDNN Hardware

Please follow the steps to build the design for zcu102 (ZU9 device based board)

1. Please generate a custom platform with 1x and 2x clocks using the steps described here. With Chai-v2, we now have the DSPs operating at twice the frequency of the rest of the core.

2. Go to CHaiDNN/design/build folder.

3. Set SDx tool environment

4. For BASH: sh source <SDx Installation Dir>/installs/lin64/SDx/2018.2/settings64.sh
5. For CSH sh source <SDx Installation Dir>/installs/lin64/SDx/2018.2/settings64.csh
6. To build the design, run Makefile. (By default this will build 1024 DSP design @ 200/400 MHz) sh make ultraclean make

:pushpin: NOTE:
- To build DietChai, run make DIET_CHAI_Z=1. This builds a design with 128 compute DSPs and 64-bit AXI interface. Run make DIET_CHAI_ZUPLUS=1 to build a design with 128 compute DSPs and 128-bit AXI interface. - To exclude deconv Kernel, set DECONV_ENABLE=0 in Makefile. Default is DECONV_ENABLE=1. - To exclude Pool Kernel, set POOL_ENABLE=0 in Makefile. With this setting, Pooling functionality embedded in Convolution accelerator is used. Default is POOL_ENABLE=1. - When building DietChai, do not change POOL_ENABLE, DECONV_ENABLE values in Makefile.

1. After the build is completed, copy the libxlnxdnn.so file and other build files (BOOT.BIN, image.ub and _sds directory) inside build/sd_card to SD_Card directory. sh make copy
2. The hardware setup is now ready.

:pushpin: NOTE:
- The 1024 DSP config was timing closed at 250/500Mhz with an iterative synthesis and P&R strategy. In the first iteration, the design was taken through the SDx flow (all the way till the bitstream Gen) at 200/400Mhz. In the second iteration the post-routed design from the first iteration was re-routed at 250/500Mhz. We believe that this is a general strategy that can be applied for other configs also. We would definitely like to hear from you on this if you are able to crank the frequency further up on other configs with this strategy. - Please note that when you try building some of the configs that are mentioned in the performance table, you might see some negative slack reported by the tools but we encourage you to try the bitstreams generated on hardware for functionality. These timing closure issues can be cleaned up with some special synthesis and P&R strategies. (You are welcome to try the timing-closure strategies that have worked for you in the past on other designs.)

Build CHaiDNN Software

Follow the steps to compile the software stack.

1. Copy libxlnxdnn.so to SD_Card/lib directory. The libxlnxdnn.so file can be found in the design/build/sd_card directory once the HW build is finished. You can skip this step if have already copied the libxlnxdnn.so file to the suggested directory.

2. Set the SDx tool environment.

3. CSH sh source <SDx Installation Dir>/installs/lin64/SDx/2018.2/settings64.csh

4. BASH sh source <SDx Installation Dir>/installs/lin64/SDx/2018.2/settings64.sh

5. Go to the software directory. This contains all the files to generate software libraries (.so).
   sh cd <path to CHaiDNN>/software

6. Go to scripts directory, open Makefile and update the SDx_BUILD_PATH variable. See example below. SDx_BUILD_PATH = <SDx Installation Dir>/installs/lin64/SDx/2018.2

7. Now run the following commands.

   sh make ultraclean make

:pushpin: NOTE:
- To build DietChai, run make DIET_CHAI_Z=1. This builds a design with 128 compute DSPs and 64-bit AXI interface. Run make DIET_CHAI_ZUPLUS=1 to build a design with 128 compute DSPs and 128-bit AXI interface. - To exclude deconv Kernel, set DECONV_ENABLE=0 in Makefile. Default is DECONV_ENABLE=1. - To exclude Pool Kernel, set POOL_ENABLE=0 in Makefile. With this setting, Pooling functionality embedded in Convolution accelerator is used. Default is POOL_ENABLE=1. - When building DietChai, do not change POOL_ENABLE, DECONV_ENABLE values in Makefile.

**:pushp