Define Custom Board and Reference Design for Deep Learning Processor IP Core Workflow

This example shows how to define and register a custom board and reference design for the deep learning processor IP core workflow. If you have a system that has preprocessing and postprocessing design under test (DUT) blocks, you can iterate on those blocks to improve your performance and resource utilization by using HDL Coder™. Create a custom reference design that includes the DUT blocks and the definition for your custom deep learning processor IP core.

Introduction

First create a custom board and reference design. Register the deep learning processor IP core in your custom board and reference design. Deploy the deep learning processor IP core to your custom board by generating a custom bitstream. Use MATLAB® to interface with the deep learning processor IP core.

Workflow

The image shows the workflow to deploy the deep learning processor IP core to a custom board.

Register Custom Board

Register a custom board in HDL Workflow Advisor.

1. Create a board registration file with the name hdlcoder_board_customization.m and add it to the MATLAB path. The hdlcoder_board_customization.m function must return a second output for workflow. Set the workflow to:

workflow = hdlcoder.Workflow.DeepLearningProcessor;

function [boardList, workflow] = hdlcoder_board_customization
% Board plugin registration file
% 1. Any registration file with this name on MATLAB path will be picked up
% 2. Registration file returns a cell array pointing to the location of
%    the board plugins
% 3. Board plugin must be a package folder accessible from MATLAB path,
%    and contains a board definition file

%   Copyright 2019 The MathWorks, Inc.

boardList = { ...
    'DLZCU102.plugin_board', ...
    'DLZC706.plugin_board', ...
    };

workflow = hdlcoder.Workflow.DeepLearningProcessor;

end

2. Create the board definition file.

%   Copyright 2019 The MathWorks, Inc.

% Board definition of ZCU102
function hB = plugin_board

% Construct board object
hB = hdlcoder.Board;

hB.BoardName    = 'Xilinx Zynq UltraScale+ MPSoC ZCU102 Evaluation Kit';

% FPGA device information
hB.FPGAVendor   = 'Xilinx';
hB.FPGAFamily   = 'Zynq UltraScale+';
hB.FPGADevice   = 'xczu9eg-ffvb1156-2-e';
hB.FPGAPackage  = '';
hB.FPGASpeed    = '';

% Tool information
hB.SupportedTool = {'Xilinx Vivado'};

% FPGA JTAG chain position
hB.JTAGChainPosition = 1;

%% Add interfaces
% Standard "External Port" interface
hB.addExternalPortInterface( ...
    'IOPadConstraint', {'IOSTANDARD = LVCMOS18'});

% GPIO LEDs
hB.addExternalIOInterface( ...
   'InterfaceID',    'LEDs General Purpose', ...
    'InterfaceType',  'OUT', ...
   'PortName',       'GPIO_LED', ...
    'PortWidth',      8, ...
    'FPGAPin',        {'AG14', 'AF13', 'AE13', 'AJ14', 'AJ15', 'AH13', 'AH14', 'AL12'}, ...
    'IOPadConstraint', {'IOSTANDARD = LVCMOS33'});

Register Custom Reference Design

Register the custom reference design in HDL Workflow Advisor.

1. Create a reference design registration file named hdlcoder_ref_design_customization.m containing a list of reference design plugins associated with an SoC board.

function [rd, boardName] = hdlcoder_ref_design_customization
% Reference design plugin registration file
% 1. The registration file with this name inside of a board plugin folder
%    will be picked up
% 2. Any registration file with this name on MATLAB path will also be picked up
% 3. The registration file returns a cell array pointing to the location of
%    the reference design plugins
% 4. The registration file also returns its associated board name
% 5. Reference design plugin must be a package folder accessible from
%    MATLAB path, and contains a reference design definition file

%   Copyright 2019 The MathWorks, Inc.

rd = {...
        'DLZCU102.matlab_libiio_3axi4_master_2019_1.plugin_rd', ...
		};

boardName = 'Xilinx Zynq UltraScale+ MPSoC ZCU102 Evaluation Kit';

end

2. Create the reference design definition file.

function hRD = plugin_rd
% Reference design definition

%   Copyright 2019-2020 The MathWorks, Inc.

% Construct reference design object
hRD = hdlcoder.ReferenceDesign('SynthesisTool','Xilinx Vivado');

hRD.ReferenceDesignName = 'AXI-Stream DDR Memory Access : 3-AXIM';
hRD.BoardName = 'Xilinx Zynq UltraScale+ MPSoC ZCU102 Evaluation Kit';

% Tool information
hRD.SupportedToolVersion = {'2019.1','2019.2','2020.1'};

%% Add custom design files
% add custom Vivado design
hRD.addCustomVivadoDesign( ...
    'CustomBlockDesignTcl','system_top.tcl');

% Add constraint files
hRD.CustomConstraints = {'ZCU102_DDR4.xdc'};

% Post Create Project Callback function to improve timing closure
hRD.PostCreateProjectFcn   = @DLZCU102.matlab_libiio_3axi4_master_2019_1.callback_PostCreateProject;

% Add HDL Verifier JTAG as AXI Master IP from support package
msg = message('hdlcommon:plugin:IPRepositoryHDLVerifierXilinxNotFound').getString;
hRD.addIPRepository( ...
	'IPListFunction','hdlverifier.fpga.vivado.iplist', ...
	'NotExistMessage', msg);

% Add AXI4-Stream to AXI4-Master DDR Access IP
hRD.addIPRepository( ...
	'IPListFunction','hdlcoder.fpga.vivado.hdlcoder_axis2axim_iplist', ...
	'NotExistMessage', 'AXI4-Stream to AXI4-Master IP not found.');

%% Add interfaces
% add clock interface
hRD.addClockInterface( ...
    'ClockConnection',     'core_clkwiz/clk_out1', ...
    'ResetConnection',     'sys_dut_rstgen/peripheral_aresetn',...
    'DefaultFrequencyMHz', 50,...
    'MinFrequencyMHz',     5,...
    'MaxFrequencyMHz',     500,...
    'ClockModuleInstance', 'core_clkwiz',...
    'ClockNumber',         1);

% add AXI4 and AXI4-Lite slave interfaces
hRD.addAXI4SlaveInterface( ...
    'InterfaceConnection', 'axi_cpu_interconnect/M03_AXI', ...
    'BaseAddress',         {'0xA0000000','0x00A0000000'}, ...
    'MasterAddressSpace',  {'hdlverifier_axi_master_0/axi4m','sys_cpu/Data'}, ...
    'InterfaceType',       'AXI4', ...
    'IDWidth',             13);

3. Add deep learning processor information to the reference design file. The reference design plugin file must contain information about the target interface and the deep learning processor IP core, the memory address space for the deep learning processor IP core, and a command to validate the reference design.

% Deep learning specific properties
hRD.registerDeepLearningTargetInterface("JTAG");
hRD.registerDeepLearningTargetInterface("Ethernet");
hRD.registerDeepLearningMemoryAddressSpace(0x80000000, 0x20000000); % 512 MB

hRD.ResourcesUsed.LogicElements = 35000;
hRD.ResourcesUsed.DSP = 3;
hRD.ResourcesUsed.RAM = 77.5;

The completed board and reference design registration files for the Xilinx® Zynq® Ultrascale+(R) MPSoC ZCU102 board can be found at,

supportpackages\R2021b\toolbox\dnnfpga\supportpackages\xilinx\boards and

supportpackages\R2021b\toolbox\dnnfpga\supportpackages\xilinx\boards\+DLZCU102.

For more details on how to set up, define, and register the custom board refer to the Getting Started with Targeting Zynq UltraScale+ MPSoC Platform (HDL Coder).