Binary16 Normalization

Floating-Point Normalization

The process of repositioning an intermediate significand into the required floating-point form while compensating the exponent so the represented value remains unchanged.

The significand and exponent must be adjusted together.

What fundamental rule preserves a floating-point value during normalization?

Desired Leading Position

The bit position where the leading 1 must appear for an intermediate FP16 result to have normal form.

Extra Upper Bit

An additional normalizer input bit used to capture a carry from addition or an oversized significand from multiplication.

The significand is too large for the desired normal form.

What does a 1 in the normalizer's extra upper bit indicate?

Shift the significand right by one and increase the exponent by one.

How does the FP16 normalizer correct an intermediate significand whose extra upper bit is 1?

Because shifting the significand right divides it by two, while increasing the exponent multiplies the scale by two.

Why does a one-bit right normalization require the exponent to increase by one?

Leading Zero

A zero occurring before the first 1 when a binary value is examined from its most-significant end.

Leading-Zero Count

The number of consecutive zero bits encountered from the most-significant end before the first 1 is reached.

FP16LeadingZeroCounter

The Proteus combinational unit that determines how many leading zeros occur in an extended significand and separately indicates an all-zero input.

From the most-significant bit toward the least-significant bit.

In what direction does FP16LeadingZeroCounter examine its input?

0

For an 8-bit input 10000000, what is the leading-zero count?

1

For an 8-bit input 01000000, what is the leading-zero count?

2

For an 8-bit input 00110000, what is the leading-zero count?

7

For an 8-bit input 00000001, what is the leading-zero count?

8, with IsZero = 1

For an 8-bit input 00000000, what leading-zero count and zero indication are produced?

The leading-zero count directly indicates how far the first 1 would need to move left toward the desired leading position.

Why is a leading-zero counter useful during floating-point normalization?

Left Normalization

The process of shifting an undersized nonzero significand left until its leading 1 reaches the normal position or the minimum normal exponent boundary is reached.

Decrease the exponent by one.

What exponent adjustment accompanies each one-bit left normalization shift?

Because shifting the significand left multiplies it by two, so reducing the exponent by one compensates for that factor of two.

Why does each left normalization shift require an exponent decrement?

-14

What minimum unbiased exponent prevents the FP16 normalizer from continuing ordinary left normalization indefinitely?

Maximum Left Shift

The greatest number of left shifts that may be performed before the exponent would fall below binary16's minimum normal unbiased exponent of -14.

CurrentExponent + 14

When the current exponent is at least -14, how does the Proteus normalizer determine the maximum permitted left-shift distance?

The smaller of the leading-zero count and the maximum permitted left shift.

How does the normalizer choose the actual left-normalization distance?

It remains subnormal.

What happens if the exponent reaches -14 before the leading 1 reaches the required normal position?

Shift the significand right by -14 - CurrentExponent and hold the effective exponent at -14.

How does the normalizer handle an intermediate exponent already below -14?

2 positions

An intermediate result has exponent -16. How far must its significand be shifted right to enter the binary16 subnormal exponent range?

Exponent Overflow

A normalization condition in which the final normalized unbiased exponent remains greater than binary16's maximum finite exponent of +15.