Integer Representation and Operation

Size and range of types

`char - short - int- long

Overflow

  • Unsigned : x + y < x / x - y > x
  • Signed: p + p = n / n + n = p
  • When comparing signed with unsigned, signed is changed to unsigned and compared.

Shift

  • logical: shift in 0’s
  • arithmetic: replicate MSB
  • Diving numbers in the 2’s complement system causes rounding to the next smallest integer, not towards 0 as desired.

Biasing in division by shifting

  • Add 2^k-1 if x < 0
  • Then shift

Binary Multiplication

  • Multiplying two (n)-bit numbers yields at most a (2n) bit product.
  • When signed → must sign extend partial products (out to 2n bits).

Binary Division

  • Dividing two (n)-bit numbers may yield an (n)-bit quotient and (n)-bit remainder.


Floating Point

Floating Point, Base 10

1.2345 * 10^{exp}
  • Bias = 4
  • Stored as 12345[exp] (ex.123459 = 1.2345*10^5)
  • Not associative

Fixed Point, Base 2

  • Radix point assumed to be in a fixed location for all numbers.
  • Floating points allows the radix point to be in a different location for each value.

Floating Point, Base 2

± b.bbb * 2^{± exp}

[sign] b.[frac] * 2^[exp]

  • Normalized FP format: ± 1.bbbbbb * 2^{± exp}
  • Floating-point numbers are always normalized.
  • The 1. is not stored but assumed since we always will store normalized numbers.

IEEE 754 Floating Point Formats

Excess-N Exponent Representation

  • Instead of 2’s complement
  • So that comparisons x < y are simple.

    Single Precision (32-bit)

  • float in C
  • 1 sign bit
  • 8 exponent bits
    • range of exponent = -126 to +127
    • value = stored - 127
  • 23 fraction bits
  • Equivalent decimal range: 7 digits * 10^{ 38}
  • s(1) exp(8) fraction(23)

Double Precision (64-bit)

  • double in C
  • 1 sign bit
  • 11 exponent bits
    • value = stored - 1023
  • 52 fraction bits
  • Equivalent decimal range: 16 digits * 10^{± 308}
  • s(1) exp(11) fraction(52)

Special Values

  • float doesn’t wrap around like int

Denormalized

  • 0 00000001 0000..0 is (1.0) * 2^-126 == 2^-126 (norm)
  • 0 00000000 1000..0 is (0.1) * 2^-126 == 2^-127 (denorm)
  • 0 00000000 0100..0 is (0.01) * 2^-126 == 2^-128 (denorm)
  • Q. What exponent value is used by denormalized 32-bit floating-point numbers?
    • A: -126

12-bit “IEEE Short” Format

  • 1 sign bit, 5 exponent bits (excess 15), 6 fraction bits

Rounding

Round to Nearest, Half to Even

  • 10...0 : round to even
  • 1x...x : round up
  • 0x...x : round down

Round towards 0 (chopping)

Rounding Implementation -check

  • Guard bits: bits immediately after LSB of fraction
  • Round bit: bit to the right of the guard bits
  • Sticky bit: Logical OR of all other bits after Guard & R bits.

FP Addition/Subtraction

  • Not associative!!!
  • Add similar, small magnitude numbers first

FP Multiplication/Division

  • Not associative - order matters!!!
  • Doesn’t distribute over addition

Updated: