Courses & Projects by Rob Marano

Notes for Week 9 — Floating Point Numbers & Arithmetic

← back to syllabus ← back to notes

Reading Assignment

1. Read sections 3.1, 3.2, 3.3, and 3.4 from CODmips textbook’s Chapter 3.

These topics are covered in our CODmips textbook’s Chapter 3 presentation deck

Topics

1. Floating point number to represent the very large and the very small of numbers.
2. Arithmetic of floating point numbers.

Topic Deep Dive

Representing real numbers

Section 3.5 of CODmips textbook focuses on floating-point arithmetic, addressing the representation and manipulation of fractions and real numbers in computers. It explains why integers alone are insufficient for many calculations and introduces the IEEE 754 standard as the dominant format for representing floating-point numbers.

Representation of Floating-Point Numbers:

The IEEE 754 standard defines two common formats:

• Single Precision (32 bits): Consists of a 1-bit sign (S), an 8-bit exponent (E), and a 23-bit fraction (F).
• Double Precision (64 bits): Consists of a 1-bit sign (S), an 11-bit exponent (E), and a 52-bit fraction (F).
• The sign bit determines the number’s sign (0 for positive, 1 for negative).
• The exponent is stored in an excess or biased format.
  • For single precision, the bias is 127.
  • For double precision, it is 1023.
  • The actual exponent is calculated by subtracting the bias from the stored exponent. This allows for representing both positive and negative exponents without using a separate sign bit.
• The fraction (also called the mantissa) represents the significant bits of the number.
  • IEEE 754 uses a normalized form where there is an implied leading ‘1’ before the binary point (except for the number 0).
  • This “hidden bit” is not explicitly stored, providing one extra bit of precision.
• Normalization: Floating-point numbers are typically normalized so that there is a single non-zero digit to the left of the binary point.
  • In binary, this means the number is of the form 1.xxxx $\times 2^y$. The hidden bit is this leading ‘1’.
• Special Values: The IEEE 754 standard defines specific bit patterns for representing special values:
  • Zero: Represented with an exponent of all 0s and a fraction of all 0s.
    • Both positive and negative zero exist (differing by the sign bit).
  • Infinity: Represented with an exponent of all 1s and a fraction of all 0s.
    • Both positive and negative infinity exist.
  • NaN (Not a Number): Represented with an exponent of all 1s and a non-zero fraction.
    • NaNs are used to represent the results of invalid operations, e.g.,
      • division by zero
      • the square root of a negative number

Floating-Point Arithmetic

The section discusses the general steps involved in performing arithmetic operations on floating-point numbers:

Addition and Subtraction

Requires aligning the exponents of the two numbers so that the binary points match. Then, the fractions can be added or subtracted. The result may need to be normalized and rounded.

Multiplication

The fractions are multiplied, and the exponents are added. The result might need to be normalized and rounded.

Division

The fraction of the dividend is divided by the fraction of the divisor, and the exponent of the divisor is subtracted from the exponent of the dividend. The result might need to be normalized and rounded.

Precision and Rounding

Due to the finite number of bits used to represent floating-point numbers, rounding is often necessary when the result of an operation cannot be exactly represented. The IEEE 754 standard defines several rounding modes. The use of guard bits, round bits, and sticky bits during calculations helps to improve the accuracy of the rounding process.

Underflow and Overflow

Overflow occurs when the result of a floating-point operation is too large to be represented in the given format.

Underflow occurs when the result is too small (close to zero) to be represented accurately, potentially leading to a loss of precision or being rounded to zero.

Worked Example (based on Exercise 3.23):

Write down the binary representation of the decimal number 63.25 assuming the IEEE 754 single precision format.

Steps:

1. Convert the integer part to binary:${63}_{10}=32+16+8+4+2+1={111111}_2$

2. Convert the fractional part to binary: ${0.25}_{10}=1/4=2^{-2}={0.01}_2$
3. Combine the integer and fractional parts: ${63.25}_{10}={111111.01}_2$
4. Normalize the binary number: Move the binary point to the left until there is only one ‘1’ to its left. Count the number of places moved; this is the exponent. ${111111.01}_2=1.1111101\times 2^5$ The fraction part (after the leading ‘1’) is $1111101$.
5. Determine the sign bit: Since 63.25 is positive, the sign bit (S) is 0.
6. Calculate the biased exponent: For single precision, the bias is 127. The actual exponent is 5.
   • Biased exponent = $5+127={132}_{10}$.
   • Convert the biased exponent to binary: ${132}_{10}=128+4={10000100}_2$
7. Write down the fraction: The fraction part is $1111101$. Since the fraction field is 23 bits long, we need to pad it with zeros on the right: ${11111010000000000000000}_2$
8. Combine the sign bit, biased exponent, and fraction:

   • Sign (1 bit)

     Biased Exponent (8 bits)

     Fraction (23 bits)

   • 0

     10000100

     11111010000000000000000

   • Therefore, the IEEE 754 single-precision binary representation of 63.25 is: 0 10000100 11111010000000000000000.

Worked Example (based on Exercise 3.29):

Calculate the sum of $2.6125\times {10}^1$ and $4.150390625\times {10}^{-1}$ by hand, assuming A and B are stored in the 16-bit half precision described in Exercise 3.27 (1 sign bit, 5 exponent bits with an excess-16 bias, and 10 mantissa bits with a hidden 1). Assume 1 guard, 1 round bit, and 1 sticky bit, and round to the nearest even. Show all the steps.

High-Level Steps:

1. Convert the decimal numbers to binary floating-point representation.
2. Adjust the exponents so they are the same.
3. Add the mantissas.
4. Normalize the result and handle potential overflow/underflow.
5. Round the result to the nearest even.
6. Convert the binary result back to decimal.

Detailed Steps to work through:

(1) Converting the decimal numbers to the 16-bit half-precision floating-point format.

Convert $2.6125\times {10}^1$ to binary

Convert the decimal number to binary: 26.125 = 11010.001 ​ Normalize the binary number: Move the binary point to the left until there is only one ‘1’ to its left: $1.1010001\times 2^4$

Determine the sign bit: Since the number is positive, the sign bit is 0.

Determine the exponent: The exponent is 4. With an excess-16 bias for 5 exponent bits, the stored exponent is 4+16=20. Convert 20 to binary: 20 = 10100 ​ Determine the mantissa: The mantissa is the fractional part of the normalized binary number, which is 1010001. Since we have 10 mantissa bits, we pad with zeros: 1010001000.

Combine the parts: Sign bit: 0 Exponent: 10100 Mantissa: 1010001000

So, the 16-bit representation of A is: 0 10100 1010001000

Now, convert B = $4.150390625\times {10}^{-1}$ to binary

Convert the decimal number to binary: 0.4150390625 = 0.0110101

Normalize the binary number: Move the binary point to the right until there is a ‘1’ to its left: $1.10101\times 2^{-2}$

Determine the sign bit: Since the number is positive, the sign bit is 0.

Determine the exponent: The exponent is -2. With an excess-16 bias for 5 exponent bits, the stored exponent is −2+16=14. Convert 14 to binary: 14 = 01110

Determine the mantissa: The mantissa is the fractional part of the normalized binary number, which is 10101. Since we have 10 mantissa bits, we pad with zeros: 1010100000.

Combine the parts: Sign bit: 0 Exponent: 01110 Mantissa: 1010100000

So, the 16-bit representation of B is: 0 01110 1010100000

← back to syllabus ← back to notes

This site built with GitHub.