Hardware-Efficient Linear Algebra for Radar and 5G

© 2021 The MathWorks, Inc.

1
Agenda

▪ Introduction
– Applications: Radar, Comms and Wireless
– Hardware Prototyping – live demo
▪ Theory and Implementation
– Linear algebra
– Matrix decomposition: QR vs Cholesky
– Latency vs. area tradeoffs
▪ HDL Coder Implementation
– HDL Coder implementation
– Resource mapping and utilization

2
Adaptive Beamforming

• Algorithm chooses optimal weights

Signal of
interest based on receive data statistics

w1 • Improve SNR by automatically

placing nulls at interference angles

w2

+
w3

Interferer w4

Adaptive
Algorithm

3
Applications: Radar

• Increase angular resolution

• Suppress interference

4
Applications: 5G

• Increase number of simultaneous users

• Improve throughput and coverage

5
Beamforming Demonstration
Test Setup
ZU28DR RFSoC
Apply
Signal of interest Signal of
Steering Angle Beam Steer
Interest Weights
4 ADC and DAC

+ channel loopback

Interference Apply
Steering Angle Interference Beam Steer
FPGA IO
TCP/IP:

Weights
MVDR
Data Capture

MVDR
Signal of interest Algorithm
Steering Angle

MATLAB

6
Agenda

▪ Introduction
– Applications: Radar, Comms and Wireless
– Hardware Prototyping – live demo
▪ Theory and Implementation
– Linear algebra
– Matrix decomposition: QR vs Cholesky
– Latency vs. area tradeoffs
▪ HDL Coder Implementation
– HDL Coder implementation
– Resource mapping and utilization

7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Agenda

▪ Introduction
– Applications: Radar, Comms and Wireless
– Hardware Prototyping – live demo
▪ Theory and Implementation
– Linear algebra
– Matrix decomposition: QR vs Cholesky
– Latency vs. area tradeoffs
▪ HDL Coder Implementation
– HDL Coder implementation
– Resource mapping and utilization

24
FPGA Implementation Challenges

▪ Fixed-Point Math

▪ Performance vs Area tradeoffs

▪ Data Rate vs Clock Rate

▪ Project Timeline

25
HDL Implementation Workflow

MATLAB
Reference
MATLAB

Hardware
Architecture Fixed Point
Designer
Fixed-point Simulink
Implementation
HDL Coder
HDL Code Generation
and Optimization
Integrated Verification
HDL Verification
and Targeting
26
MATLAB MVDR reference code

function Y = mvdr_beamform(X, sv)

% form covariance matrix

Ecx = X.'*conj(X);

% compute weight vector 100+ hours of

wp = Ecx\sv; design time saved!

% normalize response
w = wp/(sv'*wp);

% form output beam

Y = X*conj(w);

end

27
HDL Implementation of MVDR Beamforming

28
HDL Implementation of MVDR Beamforming

29
Form Covariance Matrix
▪ For Each subsystem % form covariance matrix
– Process elements independently Ecx = X.'*conj(X);
– Concatenate results into outputs

30
Moving Average

▪ Use HDL Implementation properties

to map large delays to Block RAM

31
Compute Weight Vector

▪ Use Complex Matrix Solve block from

Fixed-Point Matrix Linear Algebra Library % compute weight vector
wp = Ecx\sv;

100+ hours of
design time saved!

32
Normalize Response

▪ Perform divide using reciprocal and multiply

▪ Fixed-point CORDIC reciprocal “just works” % normalize response
w = wp/(sv'*wp);

33
Implementation Results
Xilinx ZCU111 RFSoC Eval Board
MATLAB
Simulink Ethernet
▪ Device: xczu28dr (ZCU111)

▪ Maximum frequency: 452 MHz

▪ Resource utilization: Resource Utilization (%)

LUT 47K 11.13
LUTRAM 989 0.5
FF 40K 4.7
BRAM 2 0.2
URAM 10 12.5
DSP 92 3.5

34
Resources to Get Started and Speed Adoption

▪ Getting started:
– MATLAB Onramp
– Simulink Onramp
– HDL pulse detector self-guided tutorial and videos

▪ Proof-of-concept guided evaluations

– FREE support via weekly WebEx meetings using custom sample designs
– MathWorks coaches customers on “how to fish” through weekly WebEx sessions

▪ Training & consulting services

– HDL code generation, FPGA signal processing & Zynq programming training courses
– Consulting service on deep technical coaching, custom design / hardware and more

35
Beamforming Demonstration

▪ ZCU111 RFSoC Adaptive Beamformer demo for 4x4 matrix solve for 4 channel ADC/DAC
▪ Places nulls in interference locations and maximizes beam pattern for steering direction
▪ Interactively steer angles for interference and beam pattern at run-time

Download from
File Exchange

36