Ho Chi Minh City University of Technology

Faculty of Computer Science and Engineering

Course: Computer Architecture Lab (CO2008)

Assignment
CONVOLUTION OPERATION

Ho Chi Minh City, September 2024

 Ho Chi Minh City University of Technology
Faculty of Computer Science and Engineering

Contents

1 Outcomes 2

2 Introduction 2

3 Requirements 4
3.1 Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.3 Pre-determined variables . . . . . . . . . . . . . . . . . . . . . 4
3.4 Test cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.5 Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

4 Submission 5

5 Plagiarism 5

6 Rubric for evaluation 5

Assignment for Computer Architecture

Page 1/5
Academic year 2024 - 2025
 Ho Chi Minh City University of Technology
Faculty of Computer Science and Engineering

1 Outcomes
After finishing this assignment, students can proficiently use:

• MARS MIPS simulator.

• Arithmetic & data transfer instructions.

• Conditional branch and unconditional jump instructions.

• Procedures.

2 Introduction
In this modern day and age, artificial intelligence has been developing
exponentially fast. Rapid improvements in technique and data lead to a
remarkable increase in calculations. Its sub-categories, machine learning and
deep learning, have reached a point where billions of calculations must be
done to compute the millions of parameters in a model.
Convolutional neural network (CNN) is one of the most widely used type
of network used in deep learning. Its usage can be seen in a wide range
of applications, particularly in fields that involve image and video analy-
sis. However advanced a CNN model maybe, it always stems from matrix
computations and most notably, convolution operations.
The convolution operation involves sliding a filter, also known as a kernel,
across the input image. This filter is a small matrix of weights that is applied
to a local region of the input matrix, producing a single value in the output
feature map. The process is repeated across the entire matrix, allowing the
network to learn various features of the matrix. In the case of image analysis,
these features can include edges, textures, and patterns. The mathematical
operation performed is essentially a dot product between the filter and the
receptive field of the input image.
The process involves taking a small matrix, called a kernel or filter, and
sliding it over an input matrix, such as an image. At each position, the
dot product of the kernel and the overlapping region of the input matrix
is computed, and the result is stored in the output matrix. This process
is repeated for all positions of the kernel over the input matrix, effectively
blending the kernel with the input to highlight specific features like edges or
textures. In advance,

Assignment for Computer Architecture

Page 2/5
Academic year 2024 - 2025
 Ho Chi Minh City University of Technology
Faculty of Computer Science and Engineering

Figure 1: An example of convolution operation

In Figure 1, a 5x5 input matrix and a 3x3 kernel filter have been provided.
We first match the kernel onto a corner of the input, in this case the small
3x3 matrix covered in a large red square. We then calculate the sum of
element-wise products of both matrices to get a single value. This value will
be the first value of the output matrix. Proceed to slide the kernel from left
to right, then up to down to cover every square of the input to get the final
result. Notice that the stride here is 1 as the kernel move 1 column and 1
row at a time.
One tricky issue when applying convolutional layers is that we tend to
lose pixels on the perimeter of our image. One straightforward solution to
this problem is to add extra pixels of filler around the boundary of our input
image, thus increasing the effective size of the image. Typically, we set the
values of the extra pixels to zero.

Figure 2: An example of matrix with zero-padding

In Figure 2, applying the padding with value equal to 1 into 3x3 input
matrix will increase its size to 5x5.

Assignment for Computer Architecture

Page 3/5
Academic year 2024 - 2025
 Ho Chi Minh City University of Technology
Faculty of Computer Science and Engineering

3 Requirements
In this assignment, your task is to perform convolution operations using
MIPS assembly. The requirements are listed in the following sections.

3.1 Input
The program should be able to receive input from an external input file.
It should be named input matrix.txt. The content is numbers separated
by spaces. Both image matrix and kernel matrix are square matrix mxm and
nxn, respectively. Additionally, input matrix.txt contain 3 rows, the first
row will include 4 values which are:

• N: The size of the image matrix. (3 ≤ N ≤ 7)

• M: The size of the kernel matrix. (2 ≤ M ≤ 4)

• p: The value of padding. (0 ≤ p ≤ 4)

• s: The value of stride. (1 ≤ s ≤ 3)

Thus, the contents of image and kernel matrix will be in the second and third
row, in turn.
For simplicity, the stride will be the same for both dimensions while
symmetrical padding is applied.
All of the matrices content are floating-point numbers while the stride
and padding are integer numbers.

3.2 Output
A text file named output matrix.txt contains the convolution matrix
result, also numbers separated by spaces.

3.3 Pre-determined variables

Some variables must be defined as follows for grading:

• image (word): Where the image matrix is saved.

• kernel (word): Where the kernel matrix is saved.

• out (word): Where the output matrix is saved.

Assignment for Computer Architecture

Page 4/5
Academic year 2024 - 2025
 Ho Chi Minh City University of Technology
Faculty of Computer Science and Engineering

3.4 Test cases

You are provided a set of example test cases along with this file. It is also
crucial to remember that the result of such test cases will not be included.
You are advised to develop a separated program in high-level programming
languages you have learned so far to test the MIPS program yourself.

3.5 Report
A report about your work should be made for this assignment. Images
of test runs as well as description of the functions and logic are anticipated.
Flow charts and other visualization of your ideas are also encouraged. The
report must contain basic information such as name, ID, class, and subject.

4 Submission
Students are requested to submit the MIPS program(s)/source code (.asm
files) and the Assignment report to BK E-learning system (LMS) right after
the last lab session of your class. Assignment must be done individually.
Students who do not submit on time will get 0 for the assignment.
The report should not contain code. Instead, students should present the
algorithms as well as the idea in your implementation.
There will be NO deadline extensions.

5 Plagiarism
Similarity less than 30% in MIPS code is allowed. In other words, you
will get 0 for assignment if your answers are similar to another student’s more
than 30%. Note that, we will use the MOSS tool developed by Stanford for
checking similarity (https://theory.stanford.edu/~aiken/moss/).

6 Rubric for evaluation

Your grade will base on the number of test cases you got right.

Assignment for Computer Architecture

Page 5/5
Academic year 2024 - 2025