Apparatus and method for reducing precision of data
AssigneeBroadcom Corporation
InventorsTracy C. Denk, Jeffrey S. Putnam
Life of this patent
Estimating the term…
Abstract
Methods and apparatus for reducing precision of an input signal, by comparing a portion of the input signal to a preselected threshold value, and determining a selectable bias responsive to the comparison. By combining a portion of the input signal with the selectable bias, a reduced precision signal, having minimized or eliminated rounding error, is generated. The selectable bias corresponds to a predetermined characteristic of one of bias, an error signal, the input datum, the reduced precision datum, and a combination thereof.
Claims36
A method for reducing a precision of an input datum having precision portion and a loss portion, comprising:
- a. comparing the loss portion to a preselected threshold value, ft;
- b. determining a selectable bias, α, responsive to the loss portion being in a defined relation to the preselected threshold value, ft;
- and
- c. combining the precision portion with α, creating a reduced precision datum thereby,
- wherein α corresponds to a predetermined characteristic of one of α, the input datum, the reduced precision datum, and a combination thereof.
A method for rounding a first datum, X, having precision of a digits, to a second datum, {circumflex over (X)}, having precision of b digits, wherein a>b, first b digits of X being a precision portion, and remaining a−b digits of X being a loss portion, the method comprising:
- a. evaluating the loss portion relative to a preselected rounding threshold value;
- b. if the loss portion is substantially equal to the preselected threshold, then defining {circumflex over (X)} according to the equation:
- {circumflex over (X)}=X+2−(b+1)α,
- where α is a selectable bias represented by a rounding digit;
- c. if the loss portion is not substantially equal to the preselected threshold, then defining {circumflex over (X)} according to the equation:
- {circumflex over (X)}=X+2−(b+1);
- and
- d. eliminating the loss portion of X, producing {circumflex over (X)} thereby.
A method for rounding a first two's complement fixed point datum, X, having an integer part of n bits, a fractional part of a bits the integer part, and sign bit, si, to a second two's complement fixed point datum, {circumflex over (X)}, having a fractional part of b bits following the radix point, where a and b are representative of the respective precisions of X and {circumflex over (X)}, and where a>b, comprising:
- a. evaluating the fractional part of X and defining Y as the most significant bit (MSB) of the a bits;
- b. if the first bit following the radix point of X is equal to a 1 bit trailed by (a−1) zero bits, then defining {circumflex over (X)} according to the equation:
- {circumflex over (X)}=n+si
- and
- c. otherwise, defining {circumflex over (X)} according to the equation:
- {circumflex over (X)}=n+y
A method for rounding signal values, comprising:
- a. detecting a predetermined state value wherein rounding is desired;
- and
- b. rounding the state value according to one of
- i. an alternating round-up/round-down method and
- ii. a sign addition round-up/round-down method.
An arithmetic device, comprising a bias generator producing a selectable bias α, responsive to a predetermined signal characteristic, the device receiving an input signal and coupling the selectable bias α thereto.
Dependencies are inferred from the claim text.