Assignment 6: Floating Point
CSE 30: Computer Organization and Systems, Fall 2023 Instructors: Paul Cao and Bryan Chin
Due: Tuesday Nov 14, 2023
Please read over the entire assignment before starting to get a sense of what you will need to get done in the next week. REMEMBER: Everyone procrastinates but it is important to know that you are procrastinating and still leave yourself enough time to finish. Start early, start often. You MUST run the assignment on the pi-cluster. You HAVE to SSH: You will not be able to compile or run the assignment otherwise.
ACADEMIC INTEGRITY REMINDER: you should do this assignment on your own. If you work with others, be sure to cite them in your submission. Never copy from others.
Please read the FAQ and search the existing questions on Edstem before asking for help.
This reduces the load on the teaching staff, and clutter/duplicate questions on Edstem.
Version updates:
● 1.0 [Nov 8] Final Draft
● 1.1 [Nov 8] Fix midpoint due date Sunday -> Friday
● 1.2 [Nov 9] Fix: somehexnums.txt 0x8000 to 0x4000 since it is 15 bit representation.
● 1.3 [Nov 9] Clarify: style won’t be regraded during resubmission
● 1.4 [Nov 10] Prerelease: Midpoint answers are visible before due date.
● 1.5 [Nov 12] Fix git clone link, fix # of bits in mantissas in page 4 table to be 8 bits
Table of Contents
1. Learning Goals
2. Assignment Overview
3. Getting Started
4. How the Program Works
5. Program Implementation
a. Functions to Implement
b. Developing Your Code
c. Testing Your Code
6. Submission and Grading a. Submitting
b. Grading Breakdown [50 pts]
Learning Goals
● Programming in ARM assembly
○ Bit masking
○ Function call
○ Branching
● Working with floating point numbers
● Coordinating a program with both ARM assembly and C code (you aren’t writing in C)
Assignment Overview
At the peak of time where pirates and bounty hunters are in the air. Porco Rosso makes his rounds in the vast ocean to capture any air pirates that disturb the peace near Adriano hotel.
On the radio, Porco Rosso tunes in to listen to his next job, however he discovers an issue. The coordinates given out are in 15-bit floating-point format. He doesn’t know how to convert from this format and only knows the standardized IEEE 754 format. Gina only has devices that are written in ARM so Porco plans to rely on your assembly skills to create the conversion function. Help him write and test code to convert the coordinates into IEEE format!
A note about representing number literals
In the number 8’b1101_0011:
● 8isthenumberofbinarybitsinthisnumber,inbase-10.
● bmeansbinary.Otherformatsaredfordecimal,oforoctal,andhforhexadecimal.
● To conserve space, we may also write the bits in hexadecimal, 0xd3 is equivalent to
8’b1101_0011.
CS Help, Email: tutorcs@163.com
● _isaspacercharacterthatisonlytheretomakeiteasiertoread.Ithasnonumerical meaning. A ‘_’ is usually placed every four digits.
You can read more about where this number literal representation comes from here. (Note: Anything past the first slide is irrelevant to this course, but will be useful in CSE 140 & 141.)
The 15-bit FP Format
The 15-bit floating-point format is similar to, but not the same as, the one we studied in class. It has a sign bit, 6 bits of biased exponent, a bias value of 31 (base-10), and 8 bits of mantissa. Note that we include special cases to represent infinities and subnormal numbers.
The following figure illustrates the bit assignments:
FP Format (15-bit)
sign (1 bit)
(6 bits, bias = 31)
mantissa (8 bits)
Points to note:
1. There is an implied “1.” in front of the mantissa, unless it is a subnormal number.
2. Subnormal numbers have an exponent field of 6’b000000 which represents 2−30 and implies a “0.” in front of the mantissa.
3. “Infinite” numbers have an exponent field equal to 6’b111111 with ANY value in the mantissa.
The following table shows how to interpret the exponent fields and mantissa fields.
Exponent/mantissa
represents
111111/mmmmmmm
111110/mmmmmmm
2^31 x 1.mmmmmmm
normal number
111100/mmmmmmm
2^29 x 1.mmmmmmm
normal number
111000/mmmmmmm
2^25 x 1.mmmmmmm
normal number
100000/mmmmmmm
2^1 x 1.mmmmmmm
normal number
011111/mmmmmmm
2^0 x 1.mmmmmmm
normal number
001111/mmmmmmm
2^-16 x 1.mmmmmmm
normal number
000011/mmmmmmm
2^-28 x 1.mmmmmmm
normal number
000001/mmmmmmm
2^-30 x 1.mmmmmmm
normal number
000000/mmmmmmm
2^-30 x 0.mmmmmmm
subnormal number (no leading 1)
“““““““““““““““““““““““““““““““““““““““““““““““““““““`
Exponent bits are shown in purple below to help you distinguish it from the sign bit and mantissa.
Encoding in 15-bits
15’b000_0000_0000_0000 (15 bits of 0 in binary)
15’b100_0000_0000_0000
`15-Bit Representation
Binary Representation
Base-10 Representation
15’b011_1111_xxxx_xxxx
15’b111_1111_xxxx_xxxx
Most positive #
15’b011_1110_1111_1111
2^31 * 9’b1.11111111
4286578688
Smallest positive #
(subnormal)
15’b000_0000_0000_0001
2^-30 * 9’b0.00000001
2^-38 ≅ 3.637978807e-12
Most negative #
15’b111_1110_1111_1111
-2^31 * 9’b1.11111111
-4286578688
Smallest negative # (subnormal)
15’b100_0000_0000_0001
-2^-30 * 9’b0.00000001
-2^-38 ≅ -3.637978807e-1 2
IEEE-754 Single Precision Format
The bias for the IEEE Format is 127 (base-10) and the format uses an implied “1.” for normal numbers, as usual. The smallest possible exponent is -126 represented by 8’b0000_0001 for normal numbers, whereas 8’b0000_0000 represents subnormal numbers. For subnormal numbers, we prepend the mantissa with “0.” instead of “1.” similar to how subnormal numbers are evaluated in our 15-bit FP format.
Infinities
In IEEE single precision, any exponent of all 1’s (8’b1111_1111) represents a number too large to represent. For example, 0xff80_0000 is a number with a negative sign bit, all 1’s for
the exponent and all 0’s for the mantissa. This represents negative infinity (-Inf). Similarly, 0x7f80_0000 represents positive infinity (+Inf). Note that the mantissa bits are all 0. Non-0 mantissa bits represent another kind of IEEE special number (NaN, “not a number”) which is not required in this assignment since our 15-bit floating point format does not use NaN.
Summary of Select Conversions
IEEE-754 Single Precision Format
sign (1 bit)
(8 bits, bias = 127)
mantissa (23 bits)
IEEE-754 Single
15’b000_0000_0000_0000
0x00000000
15’b100_0000_0000_0000
0x80000000
15’b000_0000_0000_0001
0x2c800000
15’b100_0000_0000_0001
0xac800000
4286578688
15’b011_1110_1111_1111
0x4f7f8000
-4286578688
15’b111_1110_1111_1111
0xcf7f8000
15’b011_1111_xxxx_xxxx
0x7f800000
15’b111_1111_xxxx_xxxx
0xff800000
Getting Started Developing Your Program
For this class, you MUST compile and run your programs on the pi-cluster.
Need help or instructions? See the Edstem FAQ. (Do NOT wait until the end to try this. There
will be limited or no ETS support on the weekends!)
We’ve provided you with the starter code at the following link:
https://github.com/cse30-fa23/hw6-starter
1. Download the files in the repository. a. You can either use
git clone https://github.com/cse30-fa23/hw6-starter.git directly from pi-cluster, or download the repo locally and scp the folder to pi-cluster if that doesn’t work.
2. Fill out the fields in the README before turning in.
Running Your Program
We’ve provided you with a Makefile so compiling your program should be easy! Additionally, the reference solution binary will be placed on Saturday 11/11 morning at:
Code Help, Add WeChat: cstutorcs
/home/linux/ieng6/cs30fa23/public/bin/fpconvert-a6-ref
Makefile: The Makefile provided will create a fpconvert executable from the source files provided. Compile it by typing make into the terminal. Run make clean to remove files generated by make.
How the Program Works
Your program will take a filename as an argument and read it in. This file is a txt file storing the 15 bit FP numbers. The main function (implemented for you in main.c) will parse the input file and call the fpconvert function which you will implement in assembly on each 15-bit FP number to convert it into IEEE floating point format, and print the result to stdout.
Once you compiled the program with make, you can run it as follows: ./fpconvert somehexnums.txt
where somehexnums.txt is the name of the input txt file that holds the hex numbers you want to convert.
Input Guarantees
● fpconvert will be only given valid 15-bit wide numbers.
Program Implementation Files to Edit
You need to edit fpconvert.S and convert_inf.S
Functions to Implement
You will need to implement 1 function within fpconvert.S:
● fpconvert(n):Thisisthefunctionthatwilldomostofthefloating-pointconversion.
○ Argument: n the 15-bit FP number to convert
○ Returns: n’s equivalent IEEE 754 single precision representation.
○ If n is ±infinity, you MUST call convert_infinity(n) to do the conversion
You need to implement 1 function within convert_inf.S
● convert_infinity(n):
○ Argument: n the 15-bit FP number to convert (should only be ±infinity)
○ Returns: the FP number’s equivalent IEEE 754 single precision representation.
● 32-bit ARM stores arguments passed into the function in registers r0-r3; n only symbolizes that the function takes in one argument. You cannot directly use n in your assembly program to refer to the first argument.
● As registers are all 32-bits wide, our 15-bit floating point format will always only occupy the least significant 15 bits, the upper 17 bits will be padded with 0’s.
● Return value should be stored in r0. Calling a Function in ARM
To call a function in ARM, you must use the bl “branch and link” instruction. It is not sufficient or correct to use a regular branch instruction. Without branch-and-link, the return operations in the epilogue of the function will not work and return as expected.
Developing Your Code
Development Process
To make development easier, you should first implement the conversion of normal numbers. Test your code on a range of normal numbers (smallest, largest). For the smallest numbers, you should familiarize yourself with their scientific notation representations. You can also check the IEEE column of the output to see if it matches the expected IEEE version. Additionally, be sure to check the special cases of +0.0, -0.0, +Inf, and -Inf.
After thoroughly testing the functionality of your code, you should consider subnormal numbers. Subnormal numbers are represented when the exponent field is 6’b000000.
After implementing the conversion of subnormal numbers, your code should be able to produce all of the values in the Summary of Select Conversions table.
Development Tips
Before you write assembly code, think about the algorithm.
● How are the 15-bit format and the 32-bit IEEE format similar and different?
● How do I break down the 15-bit format into the 3 individual fields?
● How does each field convert from the 15-bit format to the 32-bit IEEE format?
You should find the bitwise instructions useful for this assignment. In particular, you will want to make use of bitmasks.
While an immediate can only be 8 bits wide, you can use left and right shifts to move the mask into the right position. For example, if you need the bitmask 0xFF00, you can shift the immediate 0xFF left by 8 bits.
Testing your Code
To run your code you need a txt file that holds the hex numbers that you want to convert, separated by a new line.
Example text input file, named somehexnums.txt:
NOTE: you should make sure each hexadecimal number only has four digits, otherwise you may get unexpected results.
Checking For Exact Output Match
A common technique is to redirect the outputs to files and compare against the reference solution1:
./your-program args > output; our-reference args > ref
diff -s output ref
This command will output lines that differ with a < or > in front to tell which file the line came
Debugging Tips
The public autograder will only printf test some features. DO NOT RELY ON THE AUTOGRADER. (Many CSE 30 students have been burned by this.) Test your code using your own test cases!
GDB treats ARM assembly labels as functions except those that begin with the prefix “.L”. If you want to use GDB to debug your ARM code, you will need to prefix your labels with “.L”.
1 You might want to check out vimdiff on the pi-cluster (https://www.tutorialspoint.com/vim/vim_diff.htm).
Thus, ARM code for the given C if statement would look like the code snippet on the right, rather than the snippet on the left.
if (r5 == 99) { r3 = r3 + 2;
r3 = r3 + 3;
r4 = r4 – 1;
GDB will not recognize labels:
GDB will recognize labels:
cmp r5, 99 bne else
add r3, r3, 2 B end_if
add r3, r3, 3
sub r4, r4, 1
cmp r5, 99 bne .Lelse add r3, r3, 2 b .Lend_if
add r3, r3, 3
sub r4, r4, 1
Allowed ARM Instructions
You are only allowed to use the instructions provided in the ARM ISA Green Card. Failure to comply will result in a score of 0 on the assignment.
Style Requirements
Reading raw assembly is hard and debugging will be nigh impossible if you don’t put comments! To encourage you to make your life easier, style will be worth 2 points in this assignment on a holistic readable/unreadable basis. You will get full style points as long as your code is reasonably commented to be readable (so that someone who doesn’t know ARM can still roughly understand what it’s doing), so don’t worry if you can’t get all the details right. However, you will get no style points if it’s not (e.g. very inconsistent indentation, sparse or unreadable comments). In addition, staff won’t be able to provide any assistance other than styling advice unless code is readable. For reference, here is the Style Guideline for ARM assembly. We strongly recommend you to use comments after each instruction to help describe what step occurs like what is done in the style guide. Note: style will not be graded for resubmission.
Midpoint (5 Points)
This part of the assignment is due earlier than the full assignment, on
程序代写 CS代考 加QQ: 749389476
Friday 11/10 at 11:59 pm PST. There are no late submissions on the Midpoint.
Complete the Gradescope assignment “HW6: Checkpoint”, an Online Assignment that is done entirely through Gradescope. This assignment consists of a few quiz questions and a free-response question where you will document your algorithm in plain English or C code.
Discuss your implementations of the following functions: fpconvert and convert_infinity. Your fpconvert should call convert_infinity when appropriate.
Submission and Grading
Submitting
1. Submit your files to Gradescope under the assignment titled “HW6 (Coding): Floating
Point”. You will submit ONLY the following files: fpconvert.S
convert_inf.S README.md
Submission will open by Saturday morning. You should test your code extensively on pi-cluster before submitting to gradescope.
You can upload multiple files to Gradescope by holding CTRL (⌘ on a Mac) while you are clicking the files. You can also hold SHIFT to select all files between a start point and an endpoint.
Alternatively, you can place all files in a folder and upload the folder to the assignment. Gradescope will upload all files in the folder. You can also zip all of the files and upload the .zip to the assignment. Ensure that the files you submit are not in a nested folder.
2. After submitting, the autograder will run a few tests:
a. Check that all required files were submitted.
b. Check that fpconvert.S and convert_inf.S compiles.
c. Runs some sanity tests on the resulting fpconvert executable.
Grading Breakdown [5 + 45 points]
Make sure to check the autograder output after submitting! We will be running additional tests after the deadline passes to determine your final grade. Also, throughout this course, make sure to write your own test cases. It is bad practice to rely on the minimal public autograder tests as this is an insufficient test of your program.
To encourage you to write your own tests, we are not providing any public tests that have not already been detailed in this writeup.
The assignment will be graded out of 50 points and will be allocated as follows:
● Midpoint writeup: 5 points. This part of the assignment is due earlier than the full assignment, on Friday 11/10. Complete the Gradescope assignment “Homework 6: Checkpoint”.
● Code compiles with no warnings: 1 point
● Style: 2 points
● Public tests with the provided examples.
● Private tests with hidden test cases.
NOTE: The tests expect an EXACT output match with the reference binary. There will be NO partial credit for any differences in the output. Test your code – do NOT rely on the autograder for program validation.
Make sure your assignment compiles correctly through the provided Makefile on the pi-cluster without warnings. Any assignment that does not compile will receive 0 credit.
[Optional] Bells and Whistles2 (epsilon points)
Write a new function add_fp(a, b) that takes in 2 numbers a and b that are in the 15-bit floating point format. It should add these 2 numbers together and return the value. However, what makes this complicated is that a and b may not have the same exponent! You’ll need to make the exponents the same first before you can add them.
This part of the assignment will NOT be graded – and does not need to be submitted. It is completely up to you to try writing a program which achieves the above output.
The Bells and Whistles component may be submitted to a separate Gradescope assignment: Homework 6 Optional: Bells and Whistles.
2 In our experience, students like extra credit. However, we find that extra credit isn’t used by students who need it the most. Thus, we have an extra credit where the number of points assigned is epsilon, where epsilon is a very small number [0, 1).