Floating point processing unit with forced arithmetic results

• US 5,748,516 A
• Filed: 09/26/1995
• Issued: 05/05/1998
• Est. Priority Date: 09/26/1995
• Status: Expired due to Term

• Current Assignee: GlobalFoundries, Inc.
• Original Assignee: Advanced Micro Devices, Inc.
• Inventors: Goveas, Kelvin D., Goddard, Michael D.
• Primary Examiner(s): Malzahn, David H.
• Application Number:
  US08/533,812
• Time in Patent Office: 952 Days
• Field of Search: 364/748, 364/736.5, 364/748.03
• US Class Current: 708/497
• CPC Class Codes: G06F 7/483 Computations with numbers represented by a non-linear combination of denominational numbers, e.g. rational numbers, logarithmic number system or floating-point numbers G06F7/4806, G06F7/4824, G06F7/49, G06F7/491, G06F7/544 take precedence
• CPC Class Codes: G06F 7/49905 Exception handling
• CPC Class Codes: G06F 7/4991 Overflow or underflow
• CPC Class Codes: G06F 7/49926 Division by zero
• CPC Class Codes: G06F 7/49957 Implementation of IEEE-754 Standard

Abstract

Logic for selectively forcing arithmetic results allows a floating point unit to bypass the normal flow through arithmetic units and pipelines depending on the particular floating point operation and operand conditions. Certain forced results (e.g., forced zeros, infinities, and those corresponding to certain invalid operand conditions) may bypass arithmetic units or pipelines and rounding circuitry entirely. On the other hand, other operand dependent results (e.g., the result of X+0 and results of operations involving a NaN operand or operands) may only partially bypass the normal flow. By providing logic for selectively forcing results, arithmetic pipelines may be freed for subsequent instructions in the instruction stream. Logic for selectively forcing arithmetic results may be particularly attractive in a superscalar processor. In a superscalar processor which includes a floating point unit with forced arithmetic results, microcode to handle special cases, pipeline bypass, and early result generation can be avoided because architectural approaches for handling out-of-order results allow dependencies to be resolved irrespective of result reordering. Therefore, the early and out-of-order generation of forced results may be handled by a reorder buffer.

Claims

1. A floating point processing unit comprising:

an arithmetic unit coupled between a floating point operand source and a round bus, said arithmetic unit being responsive to a normal result indication;

operand passing logic coupled between said floating point operand source and said round bus, said operand passing logic being responsive to a pass operand indication;

rounding logic coupled between said round bus and a result bus;

a forced result store coupled at an output port of said rounding logic to supply forced results without rounding in response to a forced result indication; and

operand characterization and control logic coupled to a floating point instruction source, to said arithmetic unit, to said operand passing logic, and to said forced result logic, wherein said control logic supplies one of said normal result indication, said pass operand indication, and said forced result indication, said pass operand indication and said forced result indication being supplied in response to respective first and second sets of predetermined combinations of floating point instructions and operand characterizations.

2. A floating point processing unit as recited in claim 1 wherein said operand characterization indications characterize first and second operands individually.

3. A floating point processing unit as recited in claim 2 wherein said operand characterization indications are indicative of operand conditions including operand equal to zero, operand equal to infinity, and operand is a NaN.

4. A floating point processing unit as recited in claim 2 wherein said operand characterization indications are indicative of operand conditions including operand is a denormal and operand is unsupported.

5. A floating point processing unit as recited in claim 1 wherein said operand characterization indications characterize first and second operands with respect to each other.

6. A floating point processing unit as recited in claim 5 wherein said operand characterization indications include an indication of the larger one of the significand of said first and said second operands.

7. A floating point processing unit as recited in claim 1, wherein said operand passing logic includes an arithmetic unit bypass path and said rounding logic rounds operands passed via said arithmetic unit bypass path in accordance with a result format and a rounding mode.

8. A floating point processing unit as recited in claim 1, wherein said first set of predetermined combinations includes the combination of an addition instruction and a zero operand indication.

9. A floating point processing unit as recited in claim 1, wherein said first set of predetermined combinations includes combinations of a valid one of said floating point instructions and a NaN operand indication.

10. A floating point processing unit as recited in claim 1, wherein said first set includes of predetermined combinations the combination of a partial remainder instruction and a second operand equal to ±

infinity indication.

11. A floating point processing unit as recited in claim 1, wherein said second set of predetermined combinations includes combinations of said floating point instructions and operand characterization indications corresponding to ones of +infinity, -infinity, infinity, +0, -0, and 0.

12. A floating point processing unit as recited in claim 1, wherein said second set of predetermined combinations includes combinations of said floating point instructions and operand characterization indications corresponding to invalid operations.

13. A floating point processing unit as recited in claim 1, wherein said rounding logic comprises:

exponent logic coupled to said operand passing logic and said arithmetic unit, said exponent logic supplying candidate exponent values corresponding to arithmetic results rounded in accordance with ones of a predetermined set of rounding modes;

a predetermined exponent store of exponent values corresponding to ones of a predetermined set of forced results;

significand logic coupled to said operand passing logic and said arithmetic unit, said significand logic supplying candidate significand values corresponding to arithmetic results rounded in accordance with ones of said predetermined set of rounding modes,

a predetermined significand store of significand values corresponding to ones of said predetermined set of forced results;

selection logic coupled to said exponent logic, to said predetermined exponent store, to said significand logic, and to said predetermined significand store, said selection logic responsive to said force predetermined result indication, wherein said force predetermined result indication is selective for significand and exponent values corresponding to selected ones of said predetermined set of forced results.

14. In a processor having a floating point instruction source, a floating point operand source, and a result bus, a floating point processing unit comprising:

an arithmetic unit coupled to said floating point operand source and to said result bus, said arithmetic unit being responsive to a normal result indication;

operand characterization logic coupled to said floating point operand source;

control logic coupled to said operand characterization logic, to said floating point instruction source, to said arithmetic unit, wherein said control logic supplies one of said normal result indication, a pass operand indication, and a force predetermined result indication, said pass operand and force predetermined result indications being supplied in response to predetermined combinations of floating point instructions and operand characterizations;

operand passing logic including an arithmetic unit bypass path coupled between said floating point operand source and said result bus, said operand passing logic also coupled to said control logic and responsive to said pass operand indication;

rounding logic coupled between said operand passing logic and said result bus and between said arithmetic unit and said result bus, said rounding logic rounding operands passed via said arithmetic unit bypass path in accordance with a result format and a rounding mode; and

predetermined result forcing logic coupled to said result bus and to said control logic said predetermined result forcing logic being responsive to said force predetermined result indication and supplying without operation of said rounding logic predetermined results in accordance with said result format and said rounding mode.

15. A floating point processing unit as recited in claim 14 wherein outputs of said rounding logic and said predetermined result forcing logic are multiplexed onto said result bus.

16. A floating point processing unit comprising:

at least one floating point arithmetic data path coupled between first and second operand busses and a rounding unit;

predetection logic coupled in parallel with said at least one floating point arithmetic data path between said first and second operand busses and said rounding unit, said predetection logic coupled to pass an operand from a selected one of said first and said second operand busses in response to detection of a first combination of operands and operation thereupon, to supply forced result indications in response to detection of a second combination of operands and operation thereupon, and otherwise to neither pass an operand nor supply a forced result indication, a normal result instead being supplied by said at least one floating point arithmetic data path; and

said rounding unit including logic selective for stored representations of exponent and fraction portions of a particular forced arithmetic result in response to a corresponding of said forced result indications and otherwise selective for rounded result data rounded in accordance with a current rounding mode, said rounded result data being one of a rounded version of said passed operand and a rounded version of said normal result.

17. A method of supplying floating point arithmetic results, said method comprising:

predetecting, in parallel with arithmetic result computation by at least one arithmetic pipeline, first and second combinations of operands and operation thereupon;

in response to said predetecting of said first combination of operands and operation thereupon, passing an operand from a selected one of first and said second operand busses, and thereafter rounding said passed operand in accordance with a prevailing rounding mode;

in response to said predetecting of said second combination of operands and operation thereupon, supplying, from a precomputed rounded result store, a result rounded in accordance with said prevailing rounding mode; and

otherwise, supplying said arithmetic result from said at least one arithmetic data pipeline and thereafter rounding said arithmetic result in accordance with a prevailing rounding mode.