FLOATING POINT ARITHMETIC

TOPICS Binary representation of floating point Numbers Computer representation of floating point numbers Floating point instructions

BINARY REPRESENTATION OF FLOATING POINT NUMBERS Converting decimal fractions into binary representation. Consider a decimal fraction of the form: 0.d1d2...dn We want to convert this to a binary fraction of the form: 0.b1b2...bn (using binary digits instead of decimal digits)

Algorithm for conversion Let X be a decimal fraction: 0.d 1 d 2..d n i = 1 Repeat until X = 0 or i = required no. of binary fractional digits { Y = X * 2 X = fractional part of Y B i = integer part of Y i = i + 1 }

EXAMPLE 1 Convert 0.75 to binary X = 0.75 (initial value) X* 2 = Set b1 = 1, X = 0.5 X* 2 = 1.0. Set b2 = 1, X = 0.0 The binary representation for 0.75 is thus 0.b1b2 = 0.11b

Let's consider what that means... In the binary representation 0.b 1 b 2...b m b 1 represents 2 -1 (i.e., 1/2) b 2 represents 2 -2 (i.e., 1/4)... b m represents 2 -m (1/(2 m )) So, 0.11 binary represents = 1/2 + 1/4 = 3/4 = 0.75

EXAMPLE 2 Convert the decimal value 4.9 into binary Part 1: convert the integer part into binary: 4 = 100b

Part 2. Convert the fractional part into binary using multiplication by 2: X =.9*2 = 1.8. Set b 1 = 1, X = 0.8 X*2 = 1.6. Set b 2 = 1, X = 0.6 X*2 = 1.2. Set b 3 = 1, X = 0.2 X*2 = 0.4. Set b 4 = 0, X = 0.4 X*2 = 0.8. Set b 5 = 0, X = 0.8, which repeats from the first line above.

Since X is now repeating the value 0.8, we know the representation will repeat. The binary representation of 4.9 is thus:

COMPUTER REPRESENTATION OF FLOATING POINT NUMBERS In the CPU, a 32-bit floating point number is represented using IEEE standard format as follows: S | EXPONENT | MANTISSA where S is one bit, the EXPONENT is 8 bits, and the MANTISSA is 23 bits.

The mantissa represents the leading significant bits in the number. The exponent is used to adjust the position of the binary point (as opposed to a "decimal" point)

The mantissa is said to be normalized when it is expressed as a value between 1 and 2. I.e., the mantissa would be in the form 1.xxxx.

The leading integer of the binary representation is not stored. Since it is always a 1, it can be easily restored.

The "S" bit is used as a sign bit and indicates whether the value represented is positive or negative.

If a number is smaller than 1, normalizing the mantissa will produce a negative exponent. But 127 is added to all exponents in the floating point representation, allowing all exponents to be represented by a positive number.

Example 1. Represent the decimal value 2.5 in 32-bit floating point format. 2.5 = 10.1b In normalized form, this is: 1.01 * 2 1 The mantissa: M = (23 bits without the leading 1) The exponent: E = = 128 = b The sign: S = 0 (the value stored is positive) So, 2.5 =

Example 2: Represent the number b in floating point form b = * 2 -4 Mantissa: M = (23 bits with the integral 1 not represented) Exponent: E = = b S = 1 Result:

Exercise 1: represent in floating point format. Exercise 2: represent 4.9 in floating point format.

FLOATING POINT INSTRUCTIONS Floating point Architecture: 8 80-bit stack registers ST(0), ST(1),..,ST(7) (ST(0) can be abbreviated as ST) To use the floating point stack, we: Push data from memory onto the stack Process data Pop data from stack to memory.

Some floating point instructions: INSTRUCTIONDESCRIPTION Push and pops FLD, FSTPPush and pop floating point data FILD, FISTPPush and pop integer data Arithmetic In all 4 cases below, the stack is popped subsequently. So the ST(1) shown becomes ST(0) FADDST(1) = ST(0) + ST(1) FSUBST(1) = ST(0) - ST(1) FMULST(1) = ST(0) * ST(1) FDIVST(1) = ST(0) / ST(1)

Trigonometry FSINST(0) = sine of ST(0) radians FCOSST(0) = cosine of ST(0) radians FTANST(0) = tan of ST(0) radians FLDPI Push value of  onto stack

Example 1. X DW 3.4 Y DW 2 'This is an integer, while 2.0 is flt. pt. Z DW ? To evaluate Z = X + Y FLD X 'ST(0) = X FILD Y 'ST(0) = Y, ST(1) = Y FADD 'ST(0) = X + Y FSTP Z 'Z = X + Y

Example 2. To evaluate X * Y - U / V X DD 3.9 Y DD 2.8 U DD 7.3 V DD 4.62 'code follows FLD U'st(0) = U FLD V'st(0) = V, st(1) = U FDIV 'st(0) = U / V FLD X'st(0) = X, st(1) = U / V FLD Y'st(0) = Y, st(1) = X, st(2) = U / V FMUL'st(0) = Y * X, st(1) = U / V FSUB'st(0) = Y * X - U / V FSTP Z'Z = result, st(0) = empty

Util.lib contains the following subroutines for inputting and outputting floating point numbers: GetFP This inputs a no. such as 33 or 3.56 from the keyboard, and pushes it, in binary floating point form, onto the stack. PutFP This pops the floating point number from the top of the stack, and outputs it in ascii.

Note The example following assumes that you have a copy of mymacros.txt in your masm615/programs directory, and that, if the program in the example is named circle.asm, you compile the program using:mymacros.txt ml circle.asm util.lib

Example 3. Calculating Area of a Circle title calculation area of circle with inputted radius extrn getfp:near, putfp:near include mymacros.txt.model small.stack 100h.data radius dd ?.386.code circleproc startup

display "Enter the radius: " call getfp 'could have used the macro: inftp fstp radius fld radius ; instead of the above 3 instructions, we could have ; used fld st(0) or simply fld st fmul fldpi fmul display "The area of the circle is " call putfp 'could have used the macro: outftp endup circleendp end circle

ADDITIONAL FLOATING POINT INSTRUCTIONS In declaring a floating point variable use “real4” rather than “dd”, e.g.: num real4 7.3 count real4 ?

fabsreplaces st (i.e. st(0)) by its absolute value fsqrt replaces st by its square root frndint rounds st up or down to the nearest integer

We have already covered fadd used without arguments. Given,e.g.: m real4 3.6 n dd 21 fadd m will set st(0) = st(0) + m fiadd m will do the same after converting the integer n to floating point fadd st(3) will set st(0) = st(0) + st(3)

The same variations apply to: fmul fdiv fld e.g: fld st will push the value of st onto the stack, so now the top two members of the stack will have this value.