Floating point pipeline method and circuit for fast inverse square root calculations
View Patent ↗Methods and systems are provided for fast computation of reciprocal square root for floating-point numbers. A piece-wise linear approximation of the result mantissa is computed in two cycles and used as the input to an iteration sequence that converges cubically. Three iterations produce a result with accuracy sufficient for computer graphic applications. The initial estimate and input operand are scaled to minimize final adjustments to the result mantissa and final exponent adjustments required by the algorithm are performed concurrently with any adjustment required by rounding. A pipelined implementation of the algorithm produces a result with a latency of 24 and a repeat rate of 21 clock cycles.
1. A computer implemented method for use in a floating point pipeline for estimating a reciprocal square root of a mantissa component in a floating point number, comprising the steps of:
providing a constant number, K, between [1,2); and
computing an initial estimate, X i , of the inverse square root of a mantissa component, R, in two clock cycles wherein the initial estimate is computed using the constant number K, and iteratively using the expression: X i−1 =(3/2 −R/ 2×X i 2)×X i ;
wherein K is the hexadecimal value of 1.D413CC; and
wherein the initial estimate, X i is computed using <LS23>,<LS23> representing the least significant 23 bits of 1.D413CC-R.
2. The method of claim 1 , wherein the initial estimate, X i , is selected as a piecewise linear approximation to a curve 1/√{square root over (R)}.
3. The method of claim 2 , further comprising the step of iteratively refining a piecewise linear approximation to produce a result within two least significant bits (LSB) of the result obtained.
4. The method of claim 3 , wherein the result is obtained by the steps of:
calculating and rounding a square root; and
calculating and rounding the reciprocal of the square root.
5. The method of claim 1 , wherein the mantissa component, R, is scaled to produce a scaled mantissa component, Rs, and wherein the initial estimate, X i , is also scaled, to produce a product of R×X i 2 with an exponent value of 0.
6. The method of claim 5 , wherein the scaled mantissa component, R S , comprises the step of shifting the mantissa component, R, by an n-number of times.
7. A computer implemented method for use in a floating point pipeline for estimating a reciprocal square root of a mantissa component in a floating point number, comprising the steps of:
providing a constant number, K, between [1,2); and
computing an initial estimate, X i , of the inverse square root of a mantissa component, R, in two clock cycles wherein the initial estimate is computed using the constant number K;
wherein the constant number K comprises a hexadecimal value of 1.D413CC; and
wherein the initial estimate, X i is computed using <LS23>, <LS23> representing the least significant 23 bits of 1.D413CC-R.