, , , , ,

[SOLVED] COMP4601 chapter 5 lab

$25

File Name: COMP4601_chapter_5_lab.zip
File Size: 207.24 KB

5/5 - (1 vote)

In this lab you will study and accelerate three versions of a 1024-point FFT. First you will
optimize typical FFT software code for an FPGA implementation. Next, you’ll compare its
performance with that of a dataflow design. And finally, a streaming version will be
implemented.The goal is to achieve a 100X improvement in the task interval compared with
the unoptimized hardware implementation of the software code.For this lab, you are encouraged to complete Chapter 7 of the 2020.1 Vivado HLS Tutorial
and to carry out exercises relating to Ch. 5 of the text. Your answers to indicated exercises
should be submitted electronically by 5:00 pm Monday 17 July. This hand in is worth 10% of
your mark in the course. Please keep your answers concise – I expect you will need between
10 and 20 pages for your hand in.Completing the exercises below should take 3 – 4 hours after you have completed Chs 2 – 7
of the Vivado HLS Tutorial. The files needed for this lab are contained in the ch5.zip zipfile
attached to the Week 6 & 7 lab section of the course website.Code for the lab is contained in three file sets: fft_sw, fft_stages and
fft_stages_loop, corresponding to software, dataflow and streaming implementations
of the FFT. Each file set comprises a Tcl command script to create a project for the
implementation, C++ code describing the FFT algorithm, and a C header file.In the following exercises, unless instructed to do otherwise, place all directives in the
directives script for the current solution.Implementing the software code in hardware
solution1
Run vivado_hls -f fft_sw.tcl from the Vivado HLS command prompt to create
your first implementation. Open the synthesized baseline design using the command
vivado_hls -p fft_sw.projBriefly outline the performance and resource utilization of this baseline before
adding any directives. Note that the latency is not reported because the compiler
was unable to determine how many iterations of the dft_loop and
butterfly_loop need to be executed. Refer to the grey text box at the top of
page 115 in the text for an explanation of this phenomenon.solution2
The first exercise is to obtain a reasonably precise estimate of the task latency.
Create a new solution from solution1.Apply the LOOP_TRIPCOUNT directive to the dft_loop and butterfly_loop
to obtain a reasonably precise estimate of the task latency. You may find it useful to
try using a median rather than mean iteration tripcount for each loop. It is desirable
to reduce the difference between min & max latency to below 1.5%.
What do you estimate the latency to be? How did you obtain your estimate?
Has the resource utilization changed as a consequence of adding the directives?
If so, by how much?
Next, we will try reducing the latency of the bit_reverse function.solution3
Create a new solution from solution2.
Compare the performance and utilization of the bit_reverse function from
solution2 with alternatively (1) unrolling the reverse_bits_loop, (2) pipelining
the reverse_bits_loop, or (3) pipelining the function reverse_bits.Tabulate your results. The synthesis results for each solution can be found by
navigating to solution2|3 → syn → report → bit_reverse_synth.rpt from the Explorer
pane after you have provided the necessary directives and synthesized the design.
Can the performance/utilization of function bit_reverse be improved beyond
that achieved by unrolling reverse_bits_loop?solution4
Create a new solution from solution3, which has unrolled the
reverse_bits_loop.
Pipeline the dft_loop and run synthesis.
You will notice an II violation that prevents an II=1 from being achieved. Explain how
this violation comes about. Explain why the loop-carried dependency is a false
(invalid) dependency?
Next, you’ll try to eliminate the false dependency.solution5
Create a new solution from solution4.
Refer to pages 137 – 138 of the 2020.1 Vivado Design Suite User Guide on High-Level
Synthesis (UG902) linked to the General Guides section of the Labs page of the
course website for an explanation on how to use the DEPENDENCE directive to
remove false loop-carried dependencies so as to improve loop pipelining.
What is the minimum II and iteration latency you are able to achieve for the
dft_loop? How did you achieve this result?Is it worthwhile optimizing the outer loops of function fft? What constraints are
there on improving the performance any further?solution6
Create a new solution from solution5.
Edit fft_sw.h to change the data type of DTYPE from float to
ap_fixed<22,11>. Does this improve performance/resource utilization any
further?Please include a comparison report for solutions 1–6 in your submission.
Please include a copy of your directives.tcl file as well as the following
sections of your synthesis report for solution6 in your report:
Performance Estimates – Timing and Latency (Summary & Details of
Instances & Loops);
Utilization Estimates – Summary
Implementing the dataflow code in hardwaresolution1
Run vivado_hls -f fft_stages.tcl from the Vivado HLS command prompt to
create the project and solution1. Open the synthesized baseline dataflow design
using the command vivado_hls -p fft_stages.proj
Briefly comment on the performance and resource utilization of this baseline design
before adding any directives.solution2
Create a new solution from solution1.
Apply the directives you used in fft_sw/solution6 and compare the performance and
resource utilization of fft_stages/solution2 with fft_sw/solution6. Comment on any
problematic aspects of the fft_stages/solution2 results.Please include a copy of your directives.tcl file as well as the following
sections of your synthesis report for fft_stages/solution2 in your report:
Performance Estimates – Timing and Latency (Summary & Details of
Instances & Loops);
Utilization Estimates – Summary
Implementing the streaming code in hardwaresolution1
Run vivado_hls -f fft_stages_loop.tcl from the Vivado HLS command
prompt to create the project and first HLS solution. Open the synthesized baseline
streaming design using the command vivado_hls -p fft_stages_loop.proj
Briefly comment on and compare the performance and resource utilization of this
baseline design with that of fft_stages/solution2. Explain the improvement in
performance. What is the task latency and the task interval reported by the tools?solution2
Create a new solution from solution1.
The goal of this exercise is to improve the task interval and latency by pipelining the
bit_reverse_loop.If you were able to improve the performance, what measures did you take? What
impact did your changes have on resource utilization?
Overall, what improvement did you measure in task latency and interval between
fft_sw/solution2 and fft_stages_loop/solution2?How would you assess the performance of the fft_stages_loop/solution2
implementation, were it implemented as predicted on the specified device, relative
to what you think could be achieved with current high-performance general-purpose
processor? What factors are you ignoring in your assessment?Please include a copy of your directives.tcl file as well as the following
sections of your synthesis report for fft_stages_loop/solution2 in your report:
Performance Estimates – Timing and Latency (Summary & Details of
Instances & Loops);
Utilization Estimates – Summary

Reviews

There are no reviews yet.

Only logged in customers who have purchased this product may leave a review.

Shopping Cart
[SOLVED] COMP4601 chapter 5 lab
$25