Vitis AI Custom Platform Development

1. Introduction to Vitis Acceleration Platform
   
2. Create the Vivado Hardware Component
   
3. Configure Platform Interface Properties and Generate XSA
   
4. Create the PetaLinux Software Component
   
5. Create the Vitis Platform
   
6. Prepare for the DPU Kernel
   
7. Create and Build a Vitis Application
   
8. Prepare the Network Deployment File
   
9. Run Application on Board
   

Introduction to Vitis Acceleration Platform

The Vivado Design Suite is used to generate XSA containing a few additional IP blocks and metadata to support kernel connectivity. The following figure shows the acceleration kernel application development flow:

For Vitis AI platform, DPU is integrated as RTL kernel. To create a Vitis AI platform on MPSoC and run ConvNet on that, you need to create a Vivado HW platform, a PetaLinux SW platform, a Vitis platform which contains both the HW/SW platform you created. Then create a Vitis application based on this Vitis platform, import DPU kernel & ARM deployment code and build the Vitis application to be a HW-SW cowork design. Vitis would generate a SD card folder as output which would contain all the files needed to boot up from a target board. In the meanwhile to cross-compile the application and run it on board you may need Vitis AI library and DNNDK, you should install them both on the host and target board.

Create the Vivado Hardware Component and Generate XSA

1. Source <Vitis_Install_Directory>/settings64.sh, and call Vivado out by typing "vivado" in the console.
   

2. Create a Vivado project named zcu102_custom_platform.
   a) Select File->Project->New.
   b) Click Next.
   c) In Project Name dialog set Project name to zcu102_custom_platform.
   d) Click Next.
   e) Leaving all the setting to default until you goto the Default Part dialog.
   f) Select Boards tab and then select Zynq UltraScale+ ZCU102 Evaluation Board
   g) Click Next, and your project summary should like below:
   
   h) Then click Finish
   

3. Create a block design named system.
   a) Select Create Block Design.
   b) Change the design name to system.
   c) Click OK.
   

4. Add MPSoC IP and run block automation to configure it.
   a) Right click Diagram view and select Add IP.
   b) Search for zynq and then double-click the Zynq UltraScale+ MPSoC from the IP search results.
   c) Click the Run Block Automation link to apply the board presets.
   In the Run Block Automation dialog, ensure the following is check marked:
   

   • All Automation
     
   • Zynq_ultra_ps_e_0
     
   • Apply Board Presets
     

   d) Click OK. You should get MPSoC block configured like below:
   
   

Note: At this stage, the Vivado block automation has added a Zynq UltraScale+ MPSoC block and applied all board presets for the ZCU102. Add the IP blocks and metadata to create a base hardware design that supports acceleration kernels.

5. Re-Customizing the Processor IP Block
   a) Double-click the Zynq UltraScale+ MPSoC block in the IP integrator diagram.
   b) Select Page Navigator > PS-PL Configuration.
   c) Expand PS-PL Configuration > PS-PL Interfaces by clicking the > symbol.
   d) Expand Master Interface.
   e) Uncheck the AXI HPM0 FPD and AXI HPM1 FPD interfaces.
   f) Click OK.
   g) Confirm that the IP block interfaces were removed from the Zynq UltraScale+ MPSoC symbol in your block design.
   
   

Note: This is a little different from traditional Vivado design flow. When trying to make AXI interfaces available in Vitis design you should disable these interface at Vivado IPI platform and enable them at platform interface properties. We will show you how to do that later

6. Add clock block:
   a) Right click Diagram view and select Add IP.
   b) Search for and add a Clocking Wizard from the IP Search dialog.
   c) Double-click the clk_wiz_0 IP block to open the Re-Customize IP dialog box.
   d) Click the Output Clocks tab.
   e) Enable clk_out1 through clk_out3 in the Output Clock column, rename them as clk_100m, clk_200m, clk_400m and set the Requested Output Freq as follows:
   

   • clk_100m to 100 MHz.
     
   • clk_200m to 200 MHz.
     
   • clk_400m to 400 MHz.
     

   f) At the bottom of the dialog box set the Reset Type to Active Low.
   g) Click OK to close the dialog.
   The settings should like below:
   
   Note: So now we have set up the clock system for our design. This clock wizard use the pl_clk as input clock and geneatate clocks needed for the whole logic design. In this simple design I would like to use 100MHz clock as the axi_lite control bus clock, 200MHz clock as DPU AXI interface clock and 400MHz as DPU core clock. You can just modifiy these clocks as you like and remember we should "tell" Vitis what clock we can use. Let's do that later.(And after creating this example I learn that the Vitis AI DPU can only have 2 clock domains and the axi_lite control bus and DPU AXI interface share same clock. So the 100MHz clock can't be used as axi_lite control bus now. The design still can work. But between 100MHz clock and 200MHz clock Vitis would add a clock convertor inside the axi_interconnect.)
   
   

7. Add the Processor System Reset blocks:
   a) Right click Diagram view and select Add IP.
   b) Search for and add a Processor System Reset from the IP Search dialog
   c) Add 2 more Processor System Reset blocks, using the previous step; or select the proc_sys_reset_0 block and Copy (Ctrl-C) and Paste (Ctrl-V) it four times in the block diagram
   d) Rename them as proc_sys_reset_100m, proc_sys_reset_200m, proc_sys_reset_400m
   

8. Connect Clocks and Resets:
   a) Click Run Connection Automation, which will open a dialog that will help connect the proc_sys_reset blocks to the clocking wizard clock outputs.
   b) Enable All Automation on the left side of the Run Connection Automation dialog box.
   c) Select clk_in1 on clk_wiz_0, and set the Clock Source to /zynq_ultra_ps_e_0/pl_clk0.
   d) For each proc_sys_reset instance, select the slowest_sync_clk, and set the Clock Source as follows:
   

   • proc_sys_reset_100m with /clk_wiz_0/clk_100m
     
   • proc_sys_reset_200m with /clk_wiz_0/clk_200m
     
   • proc_sys_reset_400m with /clk_wiz_0/clk_400m
     

   e) On each proc_sys_reset instance, select the ext_reset_in, set Board Part Interface to Custom and set the Select Manual Source to /zynq_ultra_ps_e_0/pl_resetn0.
   f) Make sure all checkboxes are enabled, and click OK to close the dialog and create the connections.
   g) Connect all the dcm_locked signals on each proc_sys_reset instance to the locked signal on clk_wiz_0.
   Then the connection should like below:
   
   
   Now we have added clock and reset IPs and configure and connect them. Some would be used in creating the hardware platform and some would be called in Vitis high level design
   
   

9. Add Kernel Interrupt Support
   You can provide kernel interrupt support by adding an AXI interrupt controller to the base hardware design and connecting the output of the interrupt controller to the input of the processor block interrupt. The interrupt inputs of the AXI interrupt controller are initialized to a de-asserted state by wiring them to ground. When the v++ linker adds acceleration kernels to the base hardware design, the dynamic_postlink.tcl script is used to wire the interrupt output of the kernel to the AXI interrupt controller.
   a) In the block diagram, double-click the Zynq UltraScale+ MPSoC block.
   b) Select PS-PL Configuration > PS-PL interfaces > Master interface.
   c) Select the AXI HPM0 LPD check box, keep the AXI HPM0 LPD Data width settings as default 32.
   d) Click OK to finsih the configuration.
   e) Connect maxihpm0_lpd_aclk to /clk_wiz_0/clk_100m.
   f) Right click Diagram view and select Add IP, search and add Concat IP.
   g) Double-click the Concat block to open the Re-Customize IP dialog box.
   h) Set the number of ports to 32.
   i) Right click Diagram view and select Add IP, search and add Constant IP.
   j) Double-click the Constant IP, set Const Width = 1 & Const Val = 0, click OK.
   k) Connect the xlconstant_0 dout[0:0] output to all 32 inputs of xlconcat_0 like below:
   
   
   l) Select the xconstant_0 IP block,in the Block Properties, General dialog box, change the name to xlconstant_gnd.
   m) Select the xlconcat_0 IP block,in the Block Properties, General dialog box, change the name to xlconcat_interrupt_0.
   These names should match the ones in the dynamic_postlink.tcl script.
   n) Right click Diagram view and select Add IP, search and add AXI Interrupt Controller IP.
   o) Double-click the AXI Interrupt Controller block, set the Interrupts type to Level by changing the button to Manual and entering 0x0 text field, Set the Level type to High by changing the button to Manual and entering 0xFFFFFFFF, Set the Interrupt Output Connection to Single, click OK.
   The configuration of axi_intc should like below:
   
   
   p) Click Run Connection Automation
   q) Leave the default values for Master interface and Bridge IP.
   

   • Master interface default is /zynq_ultra_ps_e_0/M_AXI_HPM0_LPD.
   • Bridge IP default is New AXI interconnect.

   r) For the clock source for driving Bridge IP/Slave interface/Master interface, select /clk_wiz_0/clk_100m.
   You can select other clock resources if you want. But for axi_lite bus for register control i would recommend a lower frequency.
   s) Connect the interrupt_concat/dout[31:0] to the axi_intc_0/intr[0:0] input.
   t) Connect the axi_intc_0/irq output to the Zynq UltraScale+ MPSoC pl_ps_irq0[0:0] input.
   
   
   u) Press and hold the Shift button on keyboard, then left click xlconstant_gnd and xlconcat_interrupt_0 to select these 2 IPs. Release the Shift button, right click one of these 2 IPs, select Create Hierachy ..., use interrupt_concat as Cell Name.
   v) Click (+) to expand the interrupt_concat block, click the connection network between xlconstant_gnd and xlconcat_interrupt_0, modify the System Network Properties->Name to xlconstant_gnd_dout.
   
   
   Note: Now we have finished the IPI design input, let's set some platform parameters and generate the DSA
   
   
   

Configuring Platform Interface Properties

1. Click Window->Platform interfaces to open the Platform Interfaces Window.
   

2. Select Platform-system->zynq_ultra_ps_e_0->S_AXI_HP0_FPD, in Platform interface Properties tab enable the Enabled option like below:
   
   
   

3. Select Options tab, set memport to S_AXI_HP and set sptag to HP0 like below:
   
   

4. Do the same operations for S_AXI_HP1_FPD, S_AXI_HP2_FPD, S_AXI_HP3_FPD, S_AXI_HPC0_FPD, S_AXI_HPC1_FPD and set sptag to HP1, HP2, HP3, HPC0, HPC1. And be noticed that for HPC0/HPC1 ports the memport is set to S_AXI_HPC in default, but actually we would use these ports without data coherency function enabled to get a high performance. So please modify it into S_AXI_HP manually.
   
   
   

5. Enable the M01_AXI ~ M08_AXI ports of ps8_0_axi_periph IP(The axi_interconnect between M_AXI_HPM0_LPD and axi_intc_0), and set these ports with the same sptag name to HPM0_LPD and memport type to M_AXI_GP
   

6. Enable the M_AXI_HPM0_FPD and M_AXI_HPM1_FPD ports, set sptag name to HPM0_FPD, HPM1_FPD and memport to M_AXI_GP.
   Now we enable AXI master/slave interfaces that can be used for Vitis tools on the platform
   
   

7. Enable clk_200m, clk_400m, clk_100m of clk_wiz_0, set id of clk_200m to 0, set id of clk_400m to 1, set id of clk_100m to 2, enable is default for clk_200m.
   

8. Create a xsa_gen folder inside your Vivado project.
   

9. Copy the dynamic_postlink.tcl file into that xsa_gen folder.
   Or you can just find this file from any of the MPSoC official platform example.
   

10. Create a file named xsa.tcl inside the xsa_gen folder.
    

11. Copy the following commands into the xsa.tcl file and save the file.
    

# Set the platform design intent properties
set_property platform.design_intent.embedded true [current_project]
set_property platform.design_intent.server_managed false [current_project]
set_property platform.design_intent.external_host false [current_project]
set_property platform.design_intent.datacenter false [current_project]

get_property platform.design_intent.embedded [current_project]
get_property platform.design_intent.server_managed [current_project]
get_property platform.design_intent.external_host [current_project]
get_property platform.design_intent.datacenter [current_project]

# Set the platform default output type property
set_property platform.default_output_type "sd_card" [current_project]

get_property platform.default_output_type [current_project]

# Add the platform property to use dynamic_postlink.tcl during the v++ link
set_property platform.post_sys_link_tcl_hook ./dynamic_postlink.tcl [current_project]

12. In your Vivado project, use the Tcl console to navigate to the xsa_gen folder, and run source ./xsa.tcl command.
    
    

13. Right-click and select Validate Design on IP integrator diagram
    

14. Select the Zynq UltraScale+ MPSoC IP block and set SELECTED_SIM_MODEL to tlm in the Block Properties view.
    

15. Create the HDL wrapper:
    a. Right-click system.bd in the Block Design, Sources view and select Create HDL Wrapper.
    b. Select Let Vivado manage wrapper and auto-update.
    c. Click OK.
    

16. Right-click system.bd in the Block Design, Sources view and select Generate Output Products.
    

17. Type the tcl command in tcl console like:
    write_hw_platform -unified -force -file <your_vivado_project_dir>/xsa_gen/zcu102_custom_platform.xsa
    If you use export Hardware function in Vivado GUI it would add -fixed option which would generate a XSA for traditional embedded platform which can't add DPU acceleration kernel here.

18. Check the <your_vivado_project_dir>/xsa_gen folder, you should find the zcu102_custom_platform.xsa generated there.
    

Now we finish the Hardware platform creation flow, then we should go to the Software platform creation

Create the PetaLinux Software Component

A Vitis platform requires software components. For Linux, the PetaLinux tools are invoked outside of the Vitis tools by the developer to create the necessary Linux image,Executable and Linkable Format (ELF) files, and sysroot with XRT support. Yocto or third-party Linux development tools can also be used as long as they produce the same Linux output products as PetaLinux.

1. source <petaLinux_tool_install_dir>/settings.sh
   
2. Create a PetaLinux project named zcu102_custom_plnx and configure the hw with the XSA file we created before:
   petalinux-create --type project --template zynqMP --name zcu102_custom_plnx
cd zcu102_custom_plnx
petalinux-config --get-hw-description=<you_vivado_design_dir>/xsa_gen/
   
3. A petalinux-config menu would be launched, select DTG Settings->MACHINE_NAME, modify it to zcu102-rev1.0.
   Note: If you are using a Xilinx development board, izcu102_dpu_pkg/DPU-TRD/prj/Vitis/binary_container_1t is recomended to modify the machine name so that the board confiugrations would be involved in the DTS auto-generation. Otherwise you would need to configure the associated settings(e.g. the PHY information DTS node) by yourself manually.
   
4. Add user packages by appending the CONFIG_x lines below to the <your_petalinux_project_dir>/project-spec/meta-user/conf/user-rootfsconfig file.
   Packages for base XRT support:
   

CONFIG_xrt
CONFIG_xrt-dev
CONFIG_zocl
CONFIG_opencl-clhpp-dev
CONFIG_opencl-headers-dev
CONFIG_packagegroup-petalinux-opencv
CONFIG_packagegroup-petalinux-opencv-dev

Packages for DPU support:

CONFIG_glog
CONFIG_gtest
CONFIG_json-c
CONFIG_protobuf
CONFIG_python3-pip
CONFIG_apt
CONFIG_dpkg

Packages for building Vitis AI applications:

CONFIG_gtest-staticdev
CONFIG_json-c-dev
CONFIG_protobuf-dev
CONFIG_protobuf-c
CONFIG_libeigen-dev

Packages for native compiling on target board:

CONFIG_packagegroup-petalinux-self-hosted
CONFIG_cmake 

Packages mentioned at DPU integration lab for Vivado flow:

CONFIG_packagegroup-petalinux-x11
CONFIG_packagegroup-petalinux-v4lutils
CONFIG_packagegroup-petalinux-matchbox

5. Run petalinux-config -c rootfs and select user packages, select name of rootfs all the libraries listed above, save and exit.
   
   

6. Copy ref_files/opencv folder from this Git repository to <your_petalinux_project_dir>/project-spec/meta-user/recipes-ai in your platform source (if not exist, users need to create this directory).
   

7. Add custom opencv recipe. Edit <your_petalinux_project_dir>/project-spec/meta-user/conf/user-rootfsconfig and add the opencv recipe at the end:
   

CONFIG_opencv

8. Run petalinux-config -c rootfs and select user packages, enable opencv, save and exit.
   

9. Enable OpenSSH and disable dropbear
   Dropbear is the default SSH tool in Vitis Base Embedded Platform. If OpenSSH is used to replace Dropbear, it could achieve 4x times faster data transmission speed (tested on 1Gbps Ethernet environment). Since Vitis-AI applications may use remote display feature to show machine learning results, using OpenSSH can improve the display experience.
   a) Run petalinux-config -c rootfs.
   b) Go to Image Features.
   c) Disable ssh-server-dropbear and enable ssh-server-openssh.
   
   
   d) Go to Filesystem Packages-> misc->packagegroup-core-ssh-dropbear and disable packagegroup-core-ssh-dropbear.
   e) Go to Filesystem Packages -> console -> network -> openssh and enable openssh, openssh-sftp-server, openssh-sshd, openssh-scp.
   

10. In rootfs config go to Image Features and enable package-management and debug_tweaks option, store the change and exit.
    

11. Increase the size allocation for CMA memory to 512 MB (optional), disable CPU IDLE in the kernel configurations as follows:
    Default CMA size in PetaLinux project and Vitis Base Platform is 256MB. But for some models, 256MB is not enough to allocate DPU instructions/parameters/data area. Unless it's clear that your 256MB is sufficient for your model, it's recommended to set cma=512M which could cover all Vitis-AI models.
    CPU IDLE would cause CPU IDLE when JTAG is connected. So it is recommended to disable the selection.
    a) Type petalinux-config -c kernel
    b) Select Device Drivers > Generic Driver Options > DMA Contiguous Memory Allocator > Size in Mega Bytes.
    c) Press the Enter key and change 256 to 512.
    Ensure the following are TURNED OFF by entering 'n' in the [ ] menu selection for:
    

    • CPU Power Mangement > CPU Idle > CPU idle PM support
      
    • CPU Power Management > CPU Frequency scaling > CPU Frequency scaling
      

12. Update the Device tree to include the zocl driver by appending the text below to the project-spec/meta-user/recipes-bsp/device-tree/files/system-user.dtsi file.

&amba {
	zyxclmm_drm {
		compatible = "xlnx,zocl";
		status = "okay";
		interrupt-parent = <&axi_intc_0>;
		interrupts = <0  4>, <1  4>, <2  4>, <3  4>,
			     <4  4>, <5  4>, <6  4>, <7  4>,
			     <8  4>, <9  4>, <10 4>, <11 4>,
			     <12 4>, <13 4>, <14 4>, <15 4>,
			     <16 4>, <17 4>, <18 4>, <19 4>,
			     <20 4>, <21 4>, <22 4>, <23 4>,
			     <24 4>, <25 4>, <26 4>, <27 4>,
			     <28 4>, <29 4>, <30 4>, <31 4>;
	};
};

13. Modify the bsp config file:
    Open project-spec/meta-user/conf/petalinuxbsp.conf and add a line like below:
    

PACKAGE_CLASSES = "package_deb"

14. Modify the u-boot settings:
    Because we didn't use SD card to store the rootfs files. So that u-boot need to load a large image. We need to modify the u-boot so that it can load larger image. Open project-spec/meta-user/recipes-bsp/u-boot/files/platform-top.h and modify:
    

#define CONFIG_SYS_BOOTM_LEN 0xF000000

to

#define CONFIG_SYS_BOOTM_LEN 0x80000000
#undef CONFIG_SYS_BOOTMAPSZ

15. From within the PetaLinux project (petalinux), type petalinux-build to build the project.
    
16. Create a sysroot self-installer for the target Linux system:
    

cd images/linux
petalinux-build --sdk

Note: We would store all the necessary files for Vitis platform creation flow. Here we name it zcu102_dpu_pkg . Then we create a pfm folder inside.
17. Type ./sdk.sh to install PetaLinux SDK, provide a full pathname to the output directory <full_pathname_to_zcu102_dpu_pkg>/pfm(here in this example I use /home/wuxian/wu_project/vitis2019.2/vitis_custom_platform_flow/zcu102_dpu_pkg/pfm) and confirm.
18. We would install Vitis AI library and DNNDK into this rootfs in the future.
19. After the PetaLinux build succeeds, the generated Linux software components are in the <your_petalinux_dir>/images/linux directory. For our example, the images/linux directory contains the generated image and ELF files listed below. Copy these files to the <full_pathname_to_zcu102_dpu_pkg>/pfm/boot directory in preparation for running the Vitis platform creation flow:

    - image.ub
    - zynqmp_fsbl.elf
    - pmufw.elf
    - bl31.elf
    - u-boot.elf

20. Add a BIF file (linux.bif) to the <full_pathname_to_zcu102_dpu_pkg>/pfm/boot directory with the contents shown below. The file names should match the contents of the boot directory. The Vitis tool expands these pathnames relative to the sw directory of the platform at v++ link time or when generating an SD card. However, if the bootgen command is used directly to create a BOOT.BIN file from a BIF file, full pathnames in the BIF are necessary. Bootgen does not expand the names between the <> symbols.
    

/* linux */
 the_ROM_image:
 {
 	[fsbl_config] a53_x64
 	[bootloader] <zynqmp_fsbl.elf>
 	[pmufw_image] <pmufw.elf>
 	[destination_device=pl] <bitstream>
 	[destination_cpu=a53-0, exception_level=el-3, trustzone] <bl31.elf>
 	[destination_cpu=a53-0, exception_level=el-2] <u-boot.elf>
 }

Note: Now we prepare the HW platform and SW platform, next we would create a Vitis Platform.

Create the Vitis Platform

1. Source Vitis and XRT settings
   

source <Vitis_Install_Directory>/settings64.sh
source /opt/xilinx/xrt/setup.sh

2. Go to the zcu102_dpu_pkg folder you created: cd <full_pathname_to_zcu102_dpu_pkg>.
   
3. Launch Vitis by typing vits in the console.
   
4. Select zcu102_dpu_pkg folder as workspace directory.
   
   
   
5. In the Vitis IDE, select File > New > Platform Project to create a platform project.
   
6. In the Create New Platform Project dialog box, do the following:
   a) Enter the project name. For this example, type zcu102_vai_custom.
   b) Leave the checkbox for the default location selected.
   c) Click Next.
   
7. In the Platform Project dialog box, do the following:
   a) Select Create from hardware specification (XSA).
   b) Click Next.
   
8. In the Platform Project Specification dialog box, do the following:
   a) Browse to the XSA file generated by the Vivado. In this case, it is located in vitis_custom_platform_flow/zcu102_custom_platform/xsa_gen/zcu102_custom_platform.xsa.
   b) Set the operating system to linux.
   c) Set the processor to psu_cortexa53.
   d) Leave the checkmark selected to generate boot components.
   e) Click Finish.
   
9. In the Platform Settings view, observe the following:
   
   • The name of the Platform Settings view matches the platform project name of zcu102_vai_custom.
     
   • A psu_cortexa53 device icon is shown, containing a Linux on psu_cortexa53 domain.
     
   • A psu_cortexa53 device icon is shown, containing a zynqmp_fsbl BSP.
     
   • A psu_pmu_0 device icon is shown, containing a zynqmp_pmufw BSP.
     
10. Click the linux on psu_cortexa53 domain, browse to the locations and select the directory or file needed to complete the dialog box for the following:

Linux Build Output:
    Browse to zcu102_dpu_pkg/pfm/boot and click OK.
    
Bif file:
    Browse to zcu102_dpu_pkg/pfm/boot/linux.bif file and click OK.

Image:
    Browse to zcu102_dpu_pkg/pfm/boot and click OK.

11. Click zcu102_vai_custom project in the Vitis Explorer view, click the Build button to generate the platform.

Note: The generated platform is placed in the export directory. BSP and source files are also provided for re-building the FSBL and PMU if desired and are associated with the platform. The platform is ready to be used for application development.

Prepare for the DPU Kernel

1. Download Vitis AI by calling command git clone https://github.com/Xilinx/Vitis-AI.git.
   
2. Navigate to the repository:cd Vitis-AI, set the tag to proper tag(here we use v1.1) by typing: git checkout v1.1.
   
3. If you don't want to destroy the TRD reference design. Copy DPU-TRD folder into another directory. For example I would copy that into my zcu102_dpu_pkg folder: cp -r DPU-TRD ~/wu_project/vitis2019.2/vitis_custom_platform_flow/zcu102_dpu_pkg/
   
4. Source Vitis tools setting sh file: source <vitis install path>/Vitis/2019.2/settings64.sh.
   
5. Source XRT sh file:source opt/xilinx/xrt/setup.sh.
   
6. Export SDX_PLATFORM with the directory of the custom platform xpfm file which you created before. Here in my project it would be: export SDX_PLATFORM=/home/wuxian/wu_project/vitis2019.2/vitis_custom_platform_flow/zcu102_dpu_pkg/zcu102_vai_custom/export/zcu102_vai_custom/zcu102_vai_custom.xpfm. Remember now this custom platform name is zcu102_vai_custom.
   
7. Navigate to the copy of the DPU-TRD folder, then go to the ./prj/Vitis folder.
   There are 2 files can be used to modify the DPU settings: The config_file/prj_config file is for DPU connection in Vitis project and the dpu_conf.vh is for other DPU configurations. Here we would modify the prj_config so that 2 DPU cores are enabled. And we would keep dpu_conf.vh in default.
   
8. Modify the config_file/prj_config like below:
   

[clock]

id=0:dpu_xrt_top_1.aclk
id=1:dpu_xrt_top_1.ap_clk_2
id=0:dpu_xrt_top_2.aclk
id=1:dpu_xrt_top_2.ap_clk_2

[connectivity]

sp=dpu_xrt_top_1.M_AXI_GP0:HPC0
sp=dpu_xrt_top_1.M_AXI_HP0:HP0
sp=dpu_xrt_top_1.M_AXI_HP2:HP1
sp=dpu_xrt_top_2.M_AXI_GP0:HPC1
sp=dpu_xrt_top_2.M_AXI_HP0:HP2
sp=dpu_xrt_top_2.M_AXI_HP2:HP3

[advanced]
misc=:solution_name=link
param=compiler.addOutputTypes=sd_card

#param=compiler.skipTimingCheckAndFrequencyScaling=1

[vivado]
prop=run.impl_1.strategy=Performance_Explore
#param=place.runPartPlacer=0

9. Generate the XO file by typing: make binary_container_1/dpu.xo DEVICE=zcu102_vai_custom.
   
10. Verify if the XO file is generated here: <zcu102_dpu_pkg directory>/DPU-TRD/prj/Vitis/binary_container_1/dpu.xo.
    

Create and Build a Vitis application

1. Open Vitis workspace you were using before.
   
2. Select File -> New -> Application Project.
   
3. Name the project hello_dpu, use *new system project and use the default name, click next.
   
4. Select zcu102_vai_custom as platform, click next.
   
5. Set Domain to linux on psu_cortexa53, set Sys_root path to <full_pathname_to_zcu102_dpu_pkg>/pfm/sysroots/aarch64-xilinx-linux(as you created by running sdk.sh) and click next.
   
6. Select Empty Application and click finish to generate the application.
   
7. Right click on the src folder under your hello_dpu application in the Expplorer window, and select "Import Sources"
   
   
8. Choose from directory <zcu102_dpu_pkg directory>/DPU-TRD/prj/Vitis/binary_container_1/ as the target location, and import the dpu.xo file that we just created.
   
9. Import sources again, and add the cpp, header and prj_config files from ref_files/src folder provided by this Git repository.
   
10. In the Explorer window double click the hello_dpu.prj file to open it, change the Active Build configuration from Emulation-SW to Hardware.
    
11. Under Hardware Functions, click the lightning bolt logo to Add Hardware Function.
    
    
    
12. Select the "dpu_xrt_top" included as part of the dpu.xo file that we included earlier.
    
13. Click on binary_container_1 to change the name to dpu.
    
14. Click on dpu_xrt_top and change the Compute Units from 1 to 2 because we have 2 dpu cores involved.
    
15. Right click on "dpu", select Edit V++ Options, add --config ../src/prj_config -s as V++ Options, then click OK.
    
16. Go back to the Explorer window, right click on the hello_dpu project folder select *C/C++ Building Settings.
    
17. In Propery for Hello_DPU dialog box, select C/C++ Build->Settings->Tool Settings->GCC Host Linker->Libraries , click the green "+" to add the following libraries:

opencv_core
opencv_imgcodecs
opencv_highgui
opencv_imgproc
opencv_videoio
n2cube
hineon

18. In the same page, modify the Library search path to add ${SYSROOT}/usr/lib/, click Apply
    
    
    
19. Then go to C/C++ Build->Settings->Tool Settings->GCC Host Compiler->Includes, remove the HLS include directory and add ${SYSROOT}/usr/include/ like below, then click Apply and Close to save the changes.
    
    
    These steps are used to make sure your application can call libs in rootfs directly on Vitis appilcation build
20. The Vitis AI library and DNNDK are not included in PetaLinux SDK rootfs, now let's install them into the rootfs directory:
    Note: We should follow the section Setting Up the Host For Edge of Vitis AI library readme file to install the Vitis AI library and section Setup cross-compiler for Vitis AI DNNDK and make samples of DNNDK readme file to install the DNNDK. Most of the time I would suggest you to use a release tag when visting Github resource, but there are some critical modifications after v1.1 release. So I would just suggest you to refer to master branch this time. If you feel difficult to following the official guide there you can refer to the following ones. Please just skip these steps if you already install the libs referring to the readme files:
    a) Set the PetaLinux SDK environment by running command: . <full_pathname_to_zcu102_dpu_pkg>/pfm/environment-setup-aarch64-xilinx-linux
    b) Download the vitis_ai_2019.2-r1.1.0.tar.gz to a particular directory(here we take ~/Downloads as example) and install it to the roofs folder:
    

    cd ~/Downloads # Or some place else you download the vitis_ai_2019.2-r1.1.0.tar.gz file
    tar -xzvf vitis_ai_2019.2-r1.1.0.tar.gz -C <full_pathname_to_zcu102_dpu_pkg>/pfm/sysroots/aarch64-xilinx-linux
    

    c) Download the glog package to ~/Downloads folder and untar it:
    

    cd ~/Downloads # Or some place else you download the file
    curl -Lo glog-v0.4.0.tar.gz https://github.com/google/glog/archive/v0.4.0.tar.gz
    tar -zxvf glog-v0.4.0.tar.gz
    cd glog-0.4.0
    

    d) Build it and install it to the rootfs folder:
    

    mkdir build_for_petalinux
    cd build_for_petalinux
    unset LD_LIBRARY_PATH; source <full_pathname_to_zcu102_dpu_pkg>/pfm/environment-setup-aarch64-xilinx-linux
    cmake -DCPACK_GENERATOR=TGZ -DBUILD_SHARED_LIBS=on -DCMAKE_INSTALL_PREFIX=<full_pathname_to_zcu102_dpu_pkg>/pfm/sysroots/aarch64-xilinx-linux/usr ..
    make && make install
    make package
    

    e) Download DNNDK runtime package vitis-ai_v1.1_dnndk.tar.gz to ~/Downloads and install it into rootfs

    cd ~/Downloads # Or some place else you download the file
    tar -xzvf vitis-ai_v1.1_dnndk.tar.gz
    cd vitis-ai_v1.1_dnndk
    ./install.sh <full_pathname_to_zcu102_dpu_pkg>/pfm/sysroots/aarch64-xilinx-linux
    

Now we install both the VAI lib and DNNDK packages into the rootfs set as Vitis sysroot, then we can build application on Vitis.

21. Right click the hello_dpu project folder and select Build Project
    

Prepare the Network Deployment File

1. Find HWH file from your Vitis application folderhello_dpu/Hardware/dpu.build/link/vivado/vpl/prj/prj.srcs/sources_1/bd/system/hw_handoff/system.hwh
   Or go to your Vitis application folder use command find -name *.hwh to search for the file.
   
2. Copy this HWH file into <Vitis-AI-download_directory>/Tool-Example folder.
   
3. Go to <Vitis-AI-download_directory> folder and launch the docker.
   
4. Use following command to activate TensorFlow tool conda environment:
   

conda activate vitis-ai-tensorflow

5. Go to /workspace/Tool-Example folder and run dlet -f ./system.hwh.
   You should get the running log like below:

(vitis-ai-tensorflow) wuxian@wuxian-Ubuntu1804:/workspace/Tool-Example$ dlet -f ./system.hwh 
[DLet]Generate DPU DCF file dpu-03-26-2020-13-30.dcf successfully.

The DCF file name should be associated with the time and date you generating this file.
6. Edit the 6_tf_compile_for_v2.sh file and modify the --options parameter to add dcf file like below:
--options "{'save_kernel':'', 'dcf':'./<generated_dcf_file_name>'}"
Take my project as example it is:
--options "{'save_kernel':'', 'dcf':'./dpu-03-26-2020-13-30.dcf'}"
7. Following the TensorFlow steps at https://github.com/Xilinx/Vitis-AI/blob/v1.1/Tool-Example/README.md to generate the ELF from ResNet model.
8. Check the generated ELF file from tf_resnetv1_50_imagenet_224_224_6.97G/vai_c_output_ZCU102/dpu_resnet50_0.elf.
9. Copy that file to the src folder of Vitis application hello_dpu
10. Right click on the hello_dpu project folder in Vitis select *C/C++ Building Settings.
11. In Propery for Hello_DPU dialog box, select C/C++ Build->Settings->Tool Settings->GCC Host Linker->Miscellaneous->Other objects, add a new object: "${workspace_loc:/${ProjName}/src/dpu_resnet50_0.elf}", click Apply and Close.
12. Right click the hello_dpu project folder and select Build Project
Now you should get an updated hello_dpu.exe with a size of about 20MB(the ConvNet model is involved).

Run Application on Board

1. Copy all the files from sd_card folder inside your Vitis application like <hello_dpu_application_directory>/Hardware/sd_card/ to SD card, set ZCU102 to SD boot mode and boot up the board, connect the board with serial port.
   
2. Connect SSH:
   a) Run ifconfig to get the IP address, here we take 172.16.75.189 as example.
   b) Using SSH terminal to connect ZCU102 with SSH: ssh -x root@172.16.75.189, or use MobaXterm in Windows.
   
3. Mount SD card to mnt folder by running command: mount /dev/mmcblk0p1 /mnt.
   
4. Go to the /mnt folder and create a new folder named "package":

cd /mnt
mkdir package

5. Since this is a custom design the Vitis AI library, DNNDK and test images are not installed. We need to install them on board.
   I would suggest you to refer to section "Setting Up the Target" of Vitis AI library readme file to install the Vitis AI library and refer to section "Setup Evaluation Board and run Vitis AI DNNDK samples" of DNNDK example readme file to install DNNDK and test images.(For the similar reason now I would suggest the master branch not v1.1 tag.) If you feel difficult to do that please follow the steps below:
   a) Download the Vitis AI Runtime 1.1 package vitis-ai-runtime-1.1.2.tar.gz
   b) Untar the packet and copy the following files to the board using scp by running the command on host:
   

   scp <path_to_untar'd_runtime_library>/unilog/aarch64/libunilog-1.1.0-Linux-build<xx>.deb root@172.16.75.189:~/package
   scp <path_to_untar'd_runtime_library>/XIR/aarch64/libxir-1.1.0-Linux-build<xx>.deb root@172.16.75.189:~/package
   scp <path_to_untar'd_runtime_library>/VART/aarch64/libvart-1.1.0-Linux-build<xx>.deb root@172.16.75.189:~/package
   scp <path_to_untar'd_runtime_library>/Vitis-AI-Library/aarch64/libvitis_ai_library-1.1.0-Linux-build<xx>.deb root@172.16.75.189:~/package
   

   c) Copy the glog-0.4.0-Linux.tar.gz from host to board with the following command:
   glog-0.4.0-Linux.tar.gz is built before when configure rootfs for Vitis application
   

   cd <path_to_glog-0.4.0_build_folder>/build_for_petalinux
   scp glog-0.4.0-Linux.tar.gz 172.16.75.189:/mnt/package
   

   d) Download the package vitis-ai_v1.1_dnndk.tar.gz and package vitis-ai_v1.1_dnndk_sample_img.tar.gz, copy them to board:
   

   scp vitis-ai_v1.1_dnndk.tar.gz root@172.16.75.189:/mnt/package
   scp vitis-ai_v1.1_dnndk_sample_img.tar.gz root@172.16.75.189:/mnt/package
   

   e) In SSH console go to the /mnt/package folder and install the packages you have uploaded:
   

   cd /mnt/package
   tar -xzvf glog-0.4.0-Linux.tar.gz --strip-components=1 -C /usr
   dpkg -i --force-all libunilog-1.1.0-Linux-build46.deb
   dpkg -i libxir-1.1.0-Linux-build46.deb
   dpkg -i libvart-1.1.0-Linux-build48.deb
   dpkg -i libvitis_ai_library-1.1.0-Linux-build46.deb
   

   Notice that the first dpkg command we use --force-all option to force install this package and ignore the warning messages. And the build version may be a little different depending on the download time.
   f) Install DNNDK package like below:
   

   cp vitis-ai_v1.1_dnndk.tar.gz ~/
   cd ~/
   tar -zxvf vitis-ai_v1.1_dnndk.tar.gz
   cd vitis-ai_v1.1_dnndk
   ./install.sh
   

   g) Go back to /mnt/package folder and untar the dnndk example file:
   

   cd /mnt/package
   tar -zxvf vitis-ai_v1.1_dnndk_sample_img.tar.gz
   

6. Go to the vitis_ai_dnndk_samples and run the hello_dpu.exe application:
   

cd /mnt/package/vitis_ai_dnndk_samples
mkdir test
cd test
cp /mnt/hello_dpu.exe ./
./hello_dpu.exe

We store the hello_dpu.exe to /mnt/package/vitis_ai_dnndk_samples/test folder to suit the relative path in my code, you can do that according to your code context. The hello_dpu.exe is generated in Vitis application build and was copied to sd card from previous operation.
7. You should see the result like below:

Reference

https://www.xilinx.com/html_docs/xilinx2019_2/vitis_doc/index.html
https://github.com/Xilinx/Vitis-AI
https://github.com/Xilinx/Vitis_Embedded_Platform_Source
https://github.com/Xilinx/Vitis-AI-Tutorials/tree/Vitis-AI-Custom-Platform
https://github.com/Xilinx/Edge-AI-Platform-Tutorials/tree/3.1/docs/DPU-Integration

Note: If you would like to try with one click creating VAI platform flow it is recommended to try with the official platform source code for zcu102_dpu and zcu104_dpu.

More Information about Install and Set Vitis and XRT Environment

https://www.xilinx.com/html_docs/xilinx2019_2/vitis_doc/Chunk2027126153.html#zks1565446519267
https://www.xilinx.com/html_docs/xilinx2019_2/vitis_doc/pjr1542153622642.html
https://www.xilinx.com/html_docs/xilinx2019_2/vitis_doc/rbk1547656041291.html