Bipartite look-up table with output values having minimized absolute error

• US 6,223,192 B1
• Filed: 06/16/1998
• Issued: 04/24/2001
• Est. Priority Date: 10/23/1997
• Status: Expired due to Fees

• Current Assignee: GlobalFoundries, Inc.
• Original Assignee: Advanced Micro Devices, Inc.
• Inventors: Juffa, Norbert, Oberman, Stuart F.
• Primary Examiner(s): NGO, CHUONG D
• Application Number:
  US09/098,482
• Time in Patent Office: 1,043 Days
• Field of Search: 708/270, 708/271, 708/272, 708/273, 708/274, 708/275, 708/276, 708/277, 708/290, 708/582, 708/504, 708/605, 708/650, 708/653, 708/654
• US Class Current: 708/270
• CPC Class Codes: G06F 17/16 Matrix or vector computation , e.g. matrix-matrix or matrix-vector multiplication, matrix factorization matrix transposition G06F7/78
• CPC Class Codes: G06F 2207/3828 Multigauge devices, i.e. capable of handling packed numbers without unpacking them
• CPC Class Codes: G06F 7/4991 Overflow or underflow
• CPC Class Codes: G06F 7/49936 Normalisation mentioned as feature only
• CPC Class Codes: G06F 7/49963 Rounding to nearest G06F7/49957 takes precedence
• CPC Class Codes: G06F 7/49994 Sign extension
• CPC Class Codes: G06F 7/53 in parallel-parallel fashion, i.e. both operands being entered in parallel G06F7/533 takes precedence
• CPC Class Codes: G06F 7/5338 each bitgroup having two new bits, e.g. 2nd order MBA
• CPC Class Codes: G06F 7/5443 Sum of products for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00
• CPC Class Codes: G06F 9/30014 with variable precision
• CPC Class Codes: G06F 9/30021 Compare instructions, e.g. Greater-Than, Equal-To, MINMAX
• CPC Class Codes: G06F 9/30036 Instructions to perform operations on packed data, e.g. vector, tile or matrix operations
• CPC Class Codes: G06F 9/30038 using a mask
• CPC Class Codes: G06F 9/3017 Runtime instruction translation, e.g. macros
• CPC Class Codes: G06F 9/3804 for branches, e.g. hedging, branch folding
• CPC Class Codes: G06F 9/3885 using a plurality of independent parallel functional units
• CPC Class Codes: H03M 7/24 Conversion to or from floating-point codes

Abstract

A method for generating entries for a bipartite look-up table having base and difference table portions. In one embodiment, these entries are usable to form output values for a mathematical function, f(x), in response to receiving corresponding input values within a predetermined input range. The method first comprises partitioning the input range into I intervals, J subintervals/interval, and K sub-subintervals/subinterval. For a given interval M, the method includes generating K difference table entries and J base table entries. Each of the K difference table entries corresponds to a particular group of sub-subintervals within interval M, each of which has the same relative position within their respective subintervals. Each difference table entry is computed by averaging difference values for the sub-subintervals included in a corresponding group N. Each difference value which makes up this average is equal to f(X1)−f(X2), where X1 is the midpoint of the sub-subinterval within group N, and X2 is the midpoint of a predetermined reference sub-subinterval within the same subinterval as X1. Each of these midpoints is calculated such that maximum absolute error is minimized for all possible input values in the sub-subinterval. Each of the J base table entries, on the other hand, corresponds to a subinterval within interval M. Each entry is equal to f(X2)+adjust, where X2 is the midpoint of the reference sub-subinterval of the subinterval corresponding to the base table entry. The adjust value is calculated so that error introduced by the averaging of the difference table entries is evenly distributed over the entire subinterval.

Claims

1. A computer-readable medium encoded with a data structure, wherein the data structure comprises a bipartite look-up table usable for determining an initial estimated value for a function within a predefined input range, wherein the predefined input range is partitioned into a first number of intervals, wherein each interval is partitioned into a second number of equal subintervals, wherein each subinterval is partitioned into a third number of equal sub-subintervals, wherein said bipartite look-up table is formed by:

generating a base table entry for each subinterval of each interval;

computing one or more difference table entries for each interval, wherein computing an I^th difference table entry for a first interval comprises;

(a) computing a first midpoint in an I^th sub-subinterval of each subinterval of the first interval, and evaluating the function at the first midpoint to determine a first function value;

(b) computing a reference midpoint in a reference sub-subinterval of each subinterval of the first interval;

(c) evaluating the function at the reference midpoint to determine a reference function value;

(d) subtracting the first function value and the reference function value to determine a difference value for each subinterval of the first interval; and

(e) averaging three or more of the difference values corresponding to three or more of the subintervals of the first interval to determine an I^th average value, wherein said I^th difference table entry is substantially equal to said I^th averaged value.

2. The data structure of claim 1, wherein the function is the reciprocal function.

3. The data structure of claim 1, wherein the function is the reciprocal square-root function.

4. The data structure of claim 1, wherein the function is a monotonic function.

5. The data structure of claim 1, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint of the I^th sub-subinterval is substantially equal to (2*A*B)/(A+B).

6. The data structure of claim 1, wherein the reference sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the reference midpoint of the reference sub-subinterval is substantially equal to (2*A*B)/(A+B).

7. The data structure of claim 1, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint of the I^th sub-subinterval is substantially equal to (4*A*B)/(A+B).

8. The data structure of claim 1, wherein the reference sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the reference midpoint of the reference sub-subinterval is substantially equal to (4*A*B)/(A+B).

9. The data structure of claim 1, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint is chosen so that the first function value equals the midpoint of f(A) and f(B), wherein f(A) equals the function evaluated at lower bound A, and f(B) equals the function evaluated at the upper bound B.

10. The data structure of claim 1, wherein the reference sub-subinterval in each subinterval of the first interval is the last sub-subinterval of the subinterval.

11. The data structure of claim 1, wherein said generating a base table entry for each subinterval of the first interval comprises:

(e) subtracting the reference function value of the reference sub-subinterval of the subinterval from the first function value of an initial sub-subinterval of the subinterval to determine an actual difference;

(f) subtracting the I^th average value corresponding to an initial difference table entry for the first interval from the actual difference to determine a maximum error;

(g) dividing the maximum error by two to determine an adjustment value;

(h) adding the adjustment value to the reference function value of the reference sub-subinterval of the subinterval to generate a sum value Q.

12. The data structure of claim 11, wherein said generating a base table entry for each subinterval of the first interval further comprises:

computing a first intermediate value according to (2{circumflex over ( )}(PR+1)*Q+0.5, wherein PR is a positive integer, wherein Q is the sum value, wherein A denotes exponentiation;

truncating the first intermediate value to determine a first integer value;

subtracting 2{circumflex over ( )}PR from the first integer value to determine the base table entry.

13. The data structure of claim 1, wherein the computer-readable medium is a read-only memory.

14. The data structure of claim 1, wherein (d) comprises summing the difference value corresponding to all of the subintervals of the first interval resulting in a final sum, and dividing the final sum by the second number.

15. The data structure of claim 1, wherein said first midpoint in the I^th sub-subinterval of each subinterval of the first interval is chosen so that the corresponding first function value minimizes absolute error with respect to the function for input values within the I^th sub-subinterval.

16. A processor having a memory which contains the data structure of claim 1.

17. A method for making a microprocessor, the method comprising:

forming a memory which contains the data structure of claim 1.

18. A bipartite lookup table for generating approximations to a function for input values in a predefined input range, wherein the predefined input range is partitioned into a first number of intervals, wherein each interval is partitioned into a second number of equal subintervals, wherein each subinterval is partitioned into a third number of equal sub-subintervals, the bipartite look-up table comprising:

a difference table configured to store the third number of difference table entries for each interval, wherein an I^th difference table entry for a first interval is determined by;

(a) computing a first midpoint in an I^th sub-subinterval of each subinterval of the first interval, and evaluating the function at the first midpoint to determine a first function value;

(b) computing a reference midpoint in a reference sub-subinterval of each subinterval of the first interval, and evaluating the function at the reference midpoint to determine a reference function value;

(c) subtracting the first function value and the reference function value to determine a difference value for each subinterval of the first interval;

(d) averaging three or more of the difference values corresponding to three or more of the subintervals of the first interval to determine an I^th average value;

a base table configured to store a base table entry for each subinterval of each interval;

an address control unit configured to receive an input value, and to extract high, middle and low order bit segments from the input value, to generate a base table index by concatenating the high and middle order bit segments, to generate a difference table index by concatenating the high and low order bit segments;

wherein the base table is further configured to receive the base table index and to provide a first base table entry in response to the base table index, wherein the difference table is further configured to receive the difference table index and to provide a first difference table entry in response to the difference table index.

19. The bipartite lookup table of claim 18 further comprising an adder configured to add the first base table entry and the first difference table entry to determine an output value.

20. The bipartite lookup table of claim 18, wherein the third number of difference table entries for each interval are determined offline by a computer executing (a), (b), (c) and (d) repeatedly for each subinterval of each interval of the predetermined input range.

21. The bipartite lookup table of claim 18, wherein the function is the reciprocal function.

22. The bipartite lookup table of claim 18, wherein the function is the reciprocal square-root function.

23. The bipartite lookup table of claim 18, wherein the function is a monotonic function.

24. The bipartite lookup table of claim 18, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint of the I^th sub-subinterval is substantially equal to (2*A*B)/(A+B).

25. The bipartite lookup table of claim 18, wherein the reference sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the reference midpoint of the reference sub-subinterval is substantially equal to (2*A*B)/(A+B).

26. The bipartite lookup table of claim 18, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint of the I^th sub-subinterval is substantially equal to (4*A*B)/(A+B).

27. The bipartite lookup table of claim 18, wherein the reference sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the reference midpoint of the reference sub-subinterval is substantially equal to (4*A*B)/(4A+4B).

28. The bipartite lookup table of claim 18, wherein the I^th sub-subinterval of each subinterval of the first interval has a lower bound A and an upper bound B, wherein the first midpoint is chosen so that the first function value equals the midpoint of f(A) and f(B), wherein f(A) equals the function evaluated at lower bound A, and f(B) equals the function evaluated at the upper bound B.

29. The bipartite lookup table of claim 18, wherein the reference sub-subinterval in each subinterval of the first interval is the last sub-subinterval of the subinterval.

30. The bipartite lookup table of claim 18, wherein the base table entry for each subinterval of the first interval is generated by:

(e) subtracting the reference function value of the reference sub-subinterval of the subinterval from the first function value of an initial sub-subinterval of the subinterval to determine an actual difference;

(f) subtracting the I^th average value corresponding to an initial difference table entry for the first interval from the actual difference to determine a maximum error;

(g) dividing the maximum error by two to determine an adjustment value;

(h) adding the adjustment value to the reference function value of the reference sub-subinterval of the subinterval to generate a sum value Q.

31. The bipartite lookup table of claim 30, wherein the base table entry for each subinterval of each interval is determined offline by a computer executing (e), (f), (g) and (h).

32. The bipartite lookup table of claim 18, wherein said first midpoint in the I^th sub-subinterval of each subinterval of the first interval is chosen so that the corresponding first function value minimizes absolute error with respect to the function for input values within the I^th sub-subinterval.