1) Floating point numbers represent decimal numbers like 3.75 using a binary mantissa and exponent similar to scientific notation.
2) The IEEE 754 standard represents floating point numbers using 32 bits divided among a sign bit, mantissa, and excess-127 exponent.
3) While this representation allows representing many real numbers with finite bits, there are also limitations such as the inability to precisely represent numbers like 1/3 and rounding errors with operations.
1) Floating point numbers represent decimal numbers like 3.75 using a binary mantissa and exponent similar to scientific notation.
2) The IEEE 754 standard represents floating point numbers using 32 bits divided among a sign bit, mantissa, and excess-127 exponent.
3) While this representation allows representing many real numbers with finite bits, there are also limitations such as the inability to precisely represent numbers like 1/3 and rounding errors with operations.
1) Floating point numbers represent decimal numbers like 3.75 using a binary mantissa and exponent similar to scientific notation.
2) The IEEE 754 standard represents floating point numbers using 32 bits divided among a sign bit, mantissa, and excess-127 exponent.
3) While this representation allows representing many real numbers with finite bits, there are also limitations such as the inability to precisely represent numbers like 1/3 and rounding errors with operations.
1) Floating point numbers represent decimal numbers like 3.75 using a binary mantissa and exponent similar to scientific notation.
2) The IEEE 754 standard represents floating point numbers using 32 bits divided among a sign bit, mantissa, and excess-127 exponent.
3) While this representation allows representing many real numbers with finite bits, there are also limitations such as the inability to precisely represent numbers like 1/3 and rounding errors with operations.
CS031 September 12, 2011 Motivation We’ve seen how unsigned and signed integers are represented by a computer.
We’d like to represent decimal numbers like
3.7510 as well. By learning how these numbers are represented in hardware, we can understand and avoid pitfalls of using them in our code. Fixed-Point Binary Representation Fractional numbers are represented in binary much like integers are, but negative exponents and a decimal point are used. 3.7510 = 2+1+0.5+0.25 = 21+20+2-1+2-2 = 11.112 Not all numbers have a finite representation: 0.110 = 0.0625+0.03125+0.0078125+… = 2-4+2-5+2-8+2-9+… = 0.00011001100110011…2 Computer Representation Goals Fixed-point representation assumes no space limitations, so it’s infeasible here. Assume we have 32 bits per number. We want to represent as many decimal numbers as we can, don’t want to sacrifice range. Adding two numbers should be as similar to adding signed integers as possible. Comparing two numbers should be straightforward and intuitive in this representation. The Challenges How many distinct numbers can we represent with 32 bits? Answer: 232 We must decide which numbers to represent. Suppose we want to represent both 3.7510 = 11.112 and 7.510 = 111.12. How will the computer distinguish between them in our representation? Excursus – Scientific Notation 913.8 = 91.38 x 101 = 9.138 x 102 = 0.9138 x 103 We call the final 3 representations scientific notation. It’s standard to use the format 9.138 x 102 (exactly one non-zero digit before the decimal point) for a unique representation. The decimal point is always in the same place in scientific notation, and we can distinguish between 9138 and 913.8. Floating Point Representation Floating point is much like scientific notation. Consider 1.23 x 101 Terminology: 1.23 is the mantissa 10 is the radix, or base 1 is the exponent We can use a radix of 2 with a binary mantissa for binary notation: 1.1112 x 22 represents 111.12 = 7.510 Floating Point Normalization Like in scientific notation, we want exactly one non-zero digit before the decimal point. We use normalization to ensure that our numbers meet this requirement. 11011.12 x 2-2 normalizes to 1.101112 x 22. 0.0112 x 23 normalizes to 1.12 x 21. Normalization ensures a unique representation for our numbers. Floating Point Representation We can represent many numbers this way: Range is limited by the exponent Precision is limited by the mantissa Addition (in binary): 1.011 x 22 + 1.101 x 2-1 = 1011 x 2-1 + 1.101 x 2-1 Convert to a common exponent
= 1100.101 x 2-1 Add the mantissas
= 1.100101 x 22 Normalize the result, if necessary
Floating Point Representation Comparison: 1) Compare the signs of the mantissa -1.1 x 25 < 1.01 x 22 2) If same sign, compare the exponents 1.001 x 25 > 1.01 x 24 -1.11 x 23 > -1 x 24 3) If same exponent, compare mantissa 1.111 x 25 > 1.11 x 25 -1.01 x 2-1 > -1.101 x 2-1 1.001 x 23 = 1.001 x 23 IEEE Representation To represent floating point numbers in a computer, distribute the 32 bits between the sign, mantissa, and exponent.
For all numbers but 0, the mantissa will start
with a 1. Why? Answer: The mantissa is in binary and must start with a non-zero digit. To gain an extra bit of precision, we won’t include that 1 (zero will get a special representation). IEEE Representation
Sign bit: 0 for positive numbers, 1 for
negative numbers Mantissa: unsigned (like Lecture 1) for ease of adding and comparison But what about the exponent? IEEE Representation The exponent can be negative, so unsigned representation won’t work. We could use two’s complement. But which exponent looks bigger: 11111110 or 00000111? Comparison is not trivial in two’s complement. Excursus – Excess Notation Another way to represent signed integers, also called bias notation. In excess 127 notation, an 8-bit string whose unsigned integer value is M represents M – 127. Ex: 00000000 represents -127. 11111111 represents 128 What does 01111111 represent? 127 – 127 = 0. IEEE Representation Larger numbers look larger in excess 127 notation, so we’ll use it for the exponent. We can now represent numbers. Ex: -3.7510 = -1 * (21+20+2-1+2-2) = -11.112 = -1.1112 x 21 Sign bit: 1 (negative) Exponent: 10000000 (128-127=1) Mantissa: 11100000000000000000000 Answer: 1100 0000 0111 0000 0000 0000 0000 0000 IEEE Representation What about converting to decimal? 0011 1110 1010 1000 0000 0000 0000 0000 Sign bit: 0 (positive) Exponent: 01111101 (125-127=-2) Mantissa: 01010000000000000000000 Answer: 1.01012 x 2-2 = 0.0101012 = 0.25+0.0625+0.015625 = 0.328125 Properties of IEEE Infinitely many reals, finitely many IEEE floating point numbers. Cannot represent some numbers, like 1/3 or 1/10. More numbers are closer to zero than to 2127. (Why?) Special representation needed for zero. Properties of IEEE Representation of zero: 0000 0000 0000 0000 0000 0000 0000 0000 Positive/Negative Infinity: ?111 1111 1000 0000 0000 0000 0000 0000 NaN: (with non-zero mantissa) 0111 1111 1??? ???? ???? ???? ???? ???? Denormalized number: a number of the form ?000 0000 0??? ???? ???? ???? ???? ???? with a non-zero mantissa does not assume a hidden 1. Properties of IEEE Biggest positive number (roughly 2128): 0111 1111 0111 1111 1111 1111 1111 1111 Smallest positive number (2-126): 0000 0000 1000 0000 0000 0000 0000 0000 Negative number furthest from zero (roughly -2128): 1111 1111 0111 1111 1111 1111 1111 1111 Negative number closest to zero (-2-126): 1000 0000 0100 0000 0000 0000 0000 0000 Problems with IEEE Not all ranges can be represented.
(Figure not to scale)
Problems with IEEE Addition is not necessarily associative: Let x = -1.1 x 2120, y = 1.1 x 2120, z = 1.0 (x+y)+z = (-1.1 x 2120 + 1.1 x 2120) + 1.0 = 0 + 1.0 = 1.0 x+(y+z) = -1.1 x 2120 + (1.1 x 2120 + 1.0) = -1.1 x 2120 + 1.1 x 2120 = 0.0 Rounding errors: 0.1+0.1+0.1 != 0.3 2100 + 2-100 = 2100 Where Do We Go Next? • We have covered IEEE-754 Single Precision Floating Point representation, which is what C++ uses for the float type and close to what Java uses. Other IEEE standards exist: Java’s double type uses 64-bit, excess 1023, 11 bits for exponent. • Careful analysis can identify places where errors may occur. • We can avoid writing code that could result in floating point errors.
