Project 3: Using Vitis HLS to Implement Matrix Multiplication

ELEC 522, Advanced VLSI
Project 3: Using Vitis HLS to Implement Matrix Multiplication
Goals:
In project 2, we have looked at how to build a VLSI signal processing system by
using the graphic-based Xilinx Model Composer tool. However, text-based highlevel synthesis (HLS) tools are becoming more and more popular due to their
capability of rapid system prototyping. In this project, we will use the Xilinx Vitis
HLS tool to design a matrix multiplication system which could be easily targeted
to ASIC/FPGA. Based on Project 2 we have become familiar with the design
flow and will use methods to optimize the architecture and boost the throughput
by exploring the potential parallelism of the system.
Project procedures:
1. Review the concepts from Project 2 and the VitisHLS tutorials and Model
Composer black-box integration. (Vitis HLS labs for ZedBoard are posted
on Canvas. Please see: ELEC 522 Files / CAD_Tool_Examples /
Xilinx_Vitis_HLS_Labs)
2. Use the Vitis HLS tool to design a matrix multiplication system. Write C
code for matrix multiplication. Generate HDL code with the appropriate
architecture constraints. Verify your C code and HDL code to make sure
that they function correctly. Check the scheduling result. Adjust
architecture constraint parameters to optimize your architecture (you might
need to slightly modify your C code to adapt it to your architecture
optimization). You can build this from the sample Vitis HLS matrix
multiplication code in the tutorials.
3. Repeat step 3 until you get a good balance between system throughput
and hardware resource costs, by changing different combinations of
optimization methods. Make your choice and generate HDL code for your
final system. You should try at least 2 optimization step iterations, that is,
the original plus 2. If there is no improvement over the original, please
explain in the documentation.
4. Your system should preserve the data streaming aspects of the systolic
array structure as much as possible. That is, you should maintain multiple
I/O ports that allow data streaming from the Model Composer “host” model.
We are trying to avoid an initial block copy of the matrix from the “host” to
one single memory in the FPGA. The reason we would avoid a block
copy is that it does not allow for the overlap/pipelining of communication
and computation as in a systolic array.
5. Import your Vitis HLS code into a Model Composer model by using
the Vitis HLS Model Composer block. Use a testbench to test your
system. (You could modify your Model Composer model and testbench
from the Model Composer Project 2 on Matrix Multiplication.) There is the
Vitis HLS block in Model Composer in the “HDL/User-Defined Functions”
palette. There are some extra tutorials at:
https://www.youtube.com/watch?v=25oO4LSgA2U (Note that we have
the built-in version of HLS not the stand-alone version.) Also see:
https://www.youtube.com/watch?v=HumzeOylfyE In addition, please look
at the ug1399-vitis-hls.pdf on Canvas in ELEC 522/ Files /
CAD_Tool_Documentation / Vitis_HLS
Test and verification:
1. First write C testbench code to verify the function of your code. Use Vitis
HLS simulation tools to verify at this level (Refer to Vitis HLS tutorials).
2. In Model Composer, wrap your HDL code from Vitis HLS with a “Vitis HLS
block” and then put it into a Model Composer model to test the complete
matrix multiplication as in Project 2.
3. With Model Composer, generate for ZedBoard to do Hardware in the loop
Co-Simulation to verify performance of your Vitis HLS code on the
ZedBoard in the lab.
Requirements:
1. The system should handle 4x4 matrix multiplication. You could assume
simple data types, such as 16-bit two’s complement fixed-point numbers
as inputs and you can constrain the output also to 16-bit.
2. You could use up to 16 multipliers. You can also use less than 16
multipliers if you think that there is a "sweet point" for the latency-area
tradeoff. You need to describe your design considerations when deciding
how many multipliers to use and explain why you chose that number.
Remember that there is no standard answer for this project.
3. The data should stream into and out of your Vitis HLS module and not be
block copied if possible. This may involve some additional work on the I/O
interface. Please look at the Vitis HLS tutorials for info on PIPELINE and
INTERFACE on github at: https://github.com/Xilinx/Vitis-HLS-IntroductoryExamples . There is also a FIFO capability.
4. Try to maximize the throughput as much as possible as you can. The
methods include but are not limited to maximizing the clock frequency,
optimizing loop performance, parallel processing…
5. Generate from Model Composer first for HDL Netlist to get timing and
resource estimate and report the results (before you generate for
ZedBoard Co-Sim.) Compare the resources used (DSP48, LUT, FF
registers) between project 2 and 3, the maximum clock frequency between
projects 2 and 3, and the number of clock cycle delay (latency) and matrix
multiplication throughput between projects 2 and 3.
6. Support continuous matrix multiplication, that is, enable your design to
stream multiple problems through the array. Please indicate how much
delay is needed between each problem.
7. This should be a multi-mode system that supports both matrix-matrix
multiplication and matrix-vector multiplication as in Project 2. Your design
should have good reconfigurability and scalability.
Documentation:
In your documentation, please describe how you changed your C code and the
architecture constraints to optimize your design. Describe your final choice of the
constraint configurations. Explain the advantages of your final design over the
other trial designs that you made. Explain the timing scheduling results from the
Vivado reports. If you think you have done something clever that makes verifying
your C code and HDL code faster and easier, please also describe how you do it
in your report. Give the timing and resource information from the synthesis and
place and route implementation report. Calculate maximum throughput of your
system. Compare the results of this project with those from the Model Composer
matrix multiplication project 2 and discuss the differences between the two
design flows that possibly cause the different results.
Please include your “C” code files, your Model Composer .slx model file and any
other testbench files needed to run and verify your design. Please also include a
file which includes your simulation results, either as a text file or a waveform.
Please include step by step instructions on building and simulating you design so
that the graders can test your design. Also include a screen capture of the
results of your Hardware in the loop Co-Simulation testing results.