4.3. THE ARCHITECTURE AND OPERATION
37
the time required to finish these FFT processes will be 21.185 µs and 1.14
µs respectively. We can see that the first and the second FFT processes
meet the timing constraints, however the third FFT process fails to meet the
requirement. It requires more concurrent functioning FFT hardware blocks
to achieve the real-time performance for the third FFT. In this case, we need
at least 6 32 point FFT cores to meet the real-time constraints. However,
considering the sequential processing of FFTs, it is not mandatory for our
implementation.
38
CHAPTER 4. SYSTEM IMPLEMENTATION
Chapter 5
Results and Analysis
In the previous chapter, we discussed the algorithm and the implemented
architecture. This chapter describes the results of the implemented architec-
ture tested on Virtex-6 FPGA.
5.1
Results
Multiple tests have been carried out with the implemented architecture. As
specified in Chapter 4, the Microblaze core generates input data based on
the provided parameters such as range, velocity and angle information of
the target, number of transmitting and receiving antennas of the RF front
end and number of range and velocity bins in the FFT process. The test
results have been compared with the Matlab implementation. In addition,
the measurements have been taken to find out execution time required for
each process and to determine the bottleneck of the implementation.
5.1.1
Hardware Resource Usage
As it was mentioned in Chapter 1, the Xilinx ML605 board is used as a
hardware platform for the implementation. The boards contains Virtex-6
XC6VLX240T-1FFG1156 FPGA which is the main processing unit. The
hardware resources of this FPGA are: 241152 Logic Cells or LUTs, 37680
Logic Slices, 768 Digital Signal Processing Slices (DSP48E1) and 416 36
Kbit Block RAM blocks. Single look-up-table (LUT) has 6 input ports and
1 output port which can optionally be registered in a flip-flop. Each logic
slice consists of four LUTs, eight flip-flops and additional multiplexer and
arithmetic carry logic.
The Table 5.1 shows the hardware resource usage and utilization of the
39
40
CHAPTER 5. RESULTS AND ANALYSIS
implemented architecture. We can see that the memory interface core and the
AXI4 interconnect together take almost half of the total resources. Another
thing to note is that the FFT cores use considerable amount of resources. In
fact, the hardware resources required for 1024 point FFT core is much more
than the resources needed for the Microblaze processor core. One explanation
of this can be that both FFT cores have been configured to perform a single-
precision floating-point processing with a pipelined-streaming architecture
which require usage of significantly more hardware resources.
HW Component
Logic Slices
LUTs
BRAM
DSP48E1
Memory Interface Core
3044
6567
-
-
AXI DMA (x2)
995
2028
14
-
FFT Core - 1024 pt.
1795
5066
16
36
FFT Core - 32 pt.
1023
2869
2
16
Microblaze
1065
2565
6
5
AXI4 Interconnect
3102
7735
18
-
AXI4-Lite Interconnect
139
326
-
-
Total
12158
30184
70
57
Utilization (%)
32.3
12.5
16.8
7.4
Table 5.1: Resource usage of the architecture
5.1.2
Tests
The architecture was tested by generating input data with varying values of
range, velocity and angle parameters. The Table 5.2 contains the performed
test cases and their outputs.
We can observe that the range output values have 0.15 m difference be-
tween consequent ranges. This is due to the fact that range resolution of the
radar is equal to 0.15 m. It was derived in Chapter 2 and can be found by
Equation 2.18;
∆R =
c
2B
=
3 · 10
8
m/s
10
9
Hz
= 0.15 m
(5.1)
The data associated with each FFT bin represents a different range where
FFT bins range from 0 to 511. Consequently, we can find the maximum range
that the radar can detect which will be equal to 76.65 m (511 · 0.15).
On the other hand, the velocity of the moving target has 1.66 m/s differ-
ence between two consequent bins. The value represents the velocity resolu-
tion of the radar which was derived in Chapter 2 and can be found by using
5.1. RESULTS
41
Equation 2.20;
∆v =
c
2f
c
nT
=
3 · 10
8
2 · 79 · 10
9
· 32 · 35.6 · 10
−
6
= 1.6668 m/s
(5.2)
In a similar way we can find that the maximum velocity that the radar can
detect. It will be equal to 25.002 m/s (15 · 1.6668) due to the fact that the
FFT bins that represent positive frequencies range from 0 to 15. Therefore,
the radar can be used to detect velocities in -25 - 25 m/s range. Here the
negative velocities depict a target that is approaching to the radar and the
positive a one that is moving away.
We can see from the table that the angular resolution of the radar is not
constant over the range. It is higher in the angles closer to zero (boresight)
and smaller in the angles that are far away. This is due to the fact that
the angular resolution is dependent on the beam width of the antenna which
is smaller at angles closer to zero. We can also observe it in the shape of
arcsin() function which is used to calculate the angle FFT bin values. It has
a sharper change in the higher angles in 0
o
- 90
o
range as well as in the lower
angles in 0
o
- −90
o
range which results in having smaller resolution in that
angles.
Input
Output
Test #
Distance
Velocity
Angle
Distance
Velocity
Angle
1
4 m
4 m/s
60.2
o
4.05 m
3.33 m/s
61.0
o
2
4.16 m
5 m/s
57.3
o
4.20 m
5.00 m/s
54.3
o
3
4.33 m
6 m/s
50.0
o
4.35 m
6.66 m/s
48.6
o
4
4.55 m
7 m/s
45.0
o
4.50 m
6.66 m/s
43.4
o
5
4.68 m
-8 m/s
-5.7
o
4.65 m
-8.33 m/s
-7.2
o
6
5.07 m
-9 m/s
-11.5
o
5.10 m
-8.33 m/s
-10.8
o
7
5.81 m
-10 m/s
-15.0
o
5.85 m
-10.00 m/s
-14.5
o
8
6 m
-11 m/s
-19.5
o
6.00 m
-11.66 m/s
-18.2
o
Table 5.2: Radar test results
5.1.3
Performance
The performance of the implementation was measured by finding the execu-
tion times of the processes. For this purpose, Xilinx’s LogiCORE IP AXI
TIMER core [21] was used. The results of the measurements can be found
in the Table 5.3. It should be noted that the FFT processes also include the
time spent on DMA transfers. We can see that the time needed for 1024
Dostları ilə paylaş: |