Lab 3: Floating Point Muliplier

Lab 3: Floating Point Muliplier
Introduction
In this lab, you will build a floating point multiplier. In Part I, you will design a simplified IEEE-754 32-bit
floating point multiplier, which will take two 32-bit floating points numbers as input and produce a 32-bit
floating point number as a result. The multiplier in Part I will not have to deal with special cases (e.g., NaN,
Overflow, Underflow etc.) In Part II, you will extend the multiplier from Part I to support these special cases.
Lastly, in Part III, you will modify you multiplier to work with arbitrarily sized floating point numbers.
Part I: Simple Floating Point Multiplier
The floating point multiplication algorithm you will use for part I will be a simplified version of the one
covered in lecture. You will not have to worry about special cases such as infinity, not a number (i.e., NaN),
overflow and underflow for this Part. Your circuit must take two IEEE 754 single precision 32-bit floating
point numbers and produce their product as a 32-bit floating point result in a single cycle.
Normalize
X
S Exponent Mantissa
31 30 23 22 0
Y
S Exponent Mantissa
31 30 23 22 0
XOR
Add Exponent of X & Y
Multiply Mantissa of X & Y
Round
Rounded
value > 1
Add 1 to exponent
and shift mantissa
right
S Exponent Mantissa
31 30 23 22 0
Yes
No
Figure 1: Simple floating point multiplication flow diagram
Figure 1 shows the multiplication algorithm for two inputs X & Y. The algorithm works as follows:
ˆ First, XOR the sign bits of X & Y to get the sign of the result.
1
ECE342: Computer Hardware Lab 3: Floating point multiplier
ˆ Next, add the exponents of X & Y.
ˆ Multiply the (23-bit) mantissas of X & Y. Along with the hidden 1 in each mantissa, this results in a
48-bit product.
ˆ Round the 48-bit product by truncating the least significant 23-bits. You should be left with a 25-bit
number; 2 hidden bits and 23 mantissa bits.
ˆ Normalize the mantissa by shifting it right and adding one to the exponent if the result is greater 1.
Note that shifting right is equal to a division by two. This means the exponent must be increased by
1. For example, 23.05 x 103 − > 2.305 x 104
.
ˆ Truncate the 2 most significant bits of the result mantissa as they are hidden bits.
Example
We illustrate this process by means of an example. Note that values within parenthesis () represent ‘hidden
bits’, which are values that are used in calculation but not stored along with the number. Consider the
multiplication of X = -18.0 & Y = 9.5. In IEEE-754 representation, these values are:
X = 1 10000011 00100000000000000000000 Y = 0 10000010 00110000000000000000000
First, we XOR the sign of both numbers to get the sign of the result. Next, we extract the mantissas and
add a 1 for normalization to the most significant bit making the mantissa 24-bits instead of 23-bits.
Xmantissa = (1).00100000000000000000000 Ymantissa = (1).00110000000000000000000
Multiplying these mantissas yields a 48-bit result with 46-bits to the right of the binary point. (The series
of ... represent all 0s.)
Xmantissa × Ymantissa = 01.0101011000000000...000
We see that the MSB of this product is 0. This means that the mantissa is already in normal form. However,
if the MSB had been 1, we would then need to shift the mantissa right and increment the exponent by 1 (as
described above). In this case we just need to truncate the first two bits so that we get the 46 bits after the
decimal point.
P roductnormalised = (01).01010110000000000...000
Next, we round the value after the binary point by truncating the 23 LSBs, down from 46 bits to 23 bits.
P roductrounded = (01).01010110000000000000000
Lastly, we add the exponents of the two numbers. But keep in mind the exponents are stored with a bias of
127. Thus, when we add both exponents, we are adding 2*127 to the final number. Thus we must subtract
127 to get the correct final exponent.
P roductexponent = Xexponent + Yexponent − 127
Shown in binary, the sum of the exponents is calculated as follows:
1000 0011 (131 = 127 + 4)
+ 1000 0010 (130 = 127 + 3)
——————————
0000 0101 (7)
+ 1000 0001 (-127)
———–
1000 0110 (134 = 127 + 7)
This gives us the result of -171 (i.e., -18.0 × 9.5), which in IEEE-754 FP format is. Once again, the hidden
bit is shown for clarity but is not part of the 32-bit number itself.
Result = 1 10000110 (1) 01010110000000000000000
2
ECE342: Computer Hardware Lab 3: Floating point multiplier
Part II: IEEE-754 Special Cases
Feature Condition Expected Output
Zero E = 0, M = 0 S = 0, E = 0, M = 0
Not-a-Number E = EB, M != 0 S = 0, E = EB, M = 0
Infinity E = EB , M = 0 S = 0, E = EB, M = 0
Underflow E < 0 S = 0, E = 0, M = 0
Overflow E > EB S = 0, E = EB, M = 0
Table 1: List of features for floating point numbers, condition of the features and the expected output
In this part, you will extend your algorithm to cover special cases such as infinity, not a number (NaN),
zero, overflow and underflow. Table 1 lists special cases to check for and set flags for them accordingly. In
Table 1, E stands for exponent value, M stands for mantissa value and EB stands for exponent bias. For the
IEEE-754 32-bit representation, E = 8 and EB = 28-1 - 1 = 127. Similarly, for this format:
ˆ Overflow occurs when the sum of the exponents exceeds 127 (i.e., the largest value which is defined in
bias-127 exponent representation). When this occurs, the exponent is set to 128 (i.e, the E field is set to
255) and the mantissa is set to zero (indicating + or - infinity).
ˆ Underflow occurs when the sum of the exponents is less than -126, (i.e, the most negative value which is
defined in bias-127 exponent representation). When this occurs, the exponent is set to -127 (E = 0). If M
= 0, the number is exactly zero. NOTE: Underflow occurs when subtracting the bias.
In your code for part II, the occurrence of these special cases is indicated by means of dedicated outputs.
You must check for these features on both the inputs and the outputs of your multiplier and set the result
to the expected value accordingly. For example, if either of the inputs is 0, you should set the output to 0 as
shown in the table.
Part III: N-bit Floating Point Mulitplier
Thus far, your design has needed to work with a standard 32-bit IEEE-754 floating point format. However,
this is not the only format of floating point in use today. The IEEE-754 standard also details ‘short’ 16-bit
FP values, as well as 64-bit and 128-bit values as well. There are also non-IEEE formats such as the bfloat16
format, shown in Figure 2.
Figure 2: Example of non IEEE-754 FP representation: BFloat16
In part III you must extend your multiplier from part II to be parameterized so that your hardware can work
with any arbitrary size of floating point number. Your multiplier in Part III must still handle special cases
as shown in Table 1.
Shown below is the module inputs and outputs for Part III, with the parameters needed to configure your
multiplier. You can set these parameters via the testbench as shown in the ‘Creating generic hardware’
Tutorial on Quercus. The default values for EXP and MAN are set to the IEEE-754 standard. You must
calculate the BITS (i.e., the total number of bits in the number) as well as the BIAS (the value to be
subtracted from the exponent). You must also set the various ‘flags’ such as inf, NaN etc. if any of those
conditions are true for the inputs provided.
3
ECE342: Computer Hardware Lab 3: Floating point multiplier
1 module part3 #(
2 parameter EXP = 8 , // Number of bits in the Exponent
3 parameter MAN = 23 , // Number of bits in the Mantissa
4 parameter BITS = /* TODO */ , // Total number of bits in the floating point number
5 parameter BIAS = /* TODO */ // Value of the bias , based on the exponent .
6 )(
7 input [ BITS - 1:0] X,
8 input [ BITS - 1:0] Y,
9 output reg[ BITS - 1:0] result
10 output inf , nan , zero , overflow , underflow ,
11 );
12
13 endmodule
Example
To demonstrate how Part III should work, we illustrate with an example of multiplying two Bfloat16 values.
For Bfloat16, EXP = 8 and MAN = 7 (as shown in Figure 2). Once again, note that values within parenthesis
() represent ‘hidden bits’, which are values that are used in calculation but not stored along with the number.
We once again consider the multiplication of X = -18.0 & Y = 9.5, but this time using the Bfloat16 format.
Thus, in BFloat16 representation, these values are:
X = 1 10000011 (1) 0010000 Y = 0 10000010 (1) 0011000
First, we XOR the sign of both numbers to get the sign of the result. Next, we extract the mantissas and
add a 1 for normalization to the most significant bit making the mantissa 8-bits instead of 7-bits:
Xmantissa = 1.0010000 Ymantissa = 1.0011000
Multiplying these mantissas yields a 16-bit result with 14-bits to the right of the binary point:
Xmantissa × Ymantissa = 01.010101100000000
We see that the MSB of this product is 0 indicating that the mantissa is already in normal form. In this case
we just need to truncate the first two bits so that we get the 14 bits after the decimal point.
P roductnormalised = (01).01010110000000
Next, we round the value after the binary point, down from 16 bits to 7 bits. Next, we round the value after
the binary point by truncating the 7 LSBs, down from 14 bits to 7 bits.
P roductrounded = (01).0101011
Lastly, we add the exponents of the two numbers. Similar to IEEE-754, Bfloat16 uses 8-bits for the exponent
so we must subtract 127 to get the correct final exponent.
P roductexponent = Xexponent + Yexponent − 127
The sum of the exponents is:
1000 0011 (131 = 127 + 4)
+ 1000 0010 (130 = 127 + 3)
——————————
0000 0101 (7)
+ 1000 0001 (-127)
———–
1000 0110 (134 = 127 + 7)
This gives us the result of -171 (i.e., -18.0 × 9.5), which in Bfloat16 format is:
Result = 1 10000110 (1) 0101011
4
ECE342: Computer Hardware Lab 3: Floating point multiplier
Preparation
ˆ Familiarize yourself with the floating point number representation, particularly the IEEE-754 format.
You can do so by referring to ieee754-1985-STANDARD cleanedup.pdf document on Quercus
ˆ Name the finished code part1.sv, part2.sv and part3.sv respectively.
Submission
Your should submit part1.sv, part2.sv and part3.sv using the instructions provided in the submission document.
5