LOW ENERGY CONSUMPTION MANTISSA MULTIPLICATION FOR FLOATING POINT MULTIPLY-ADD OPERATIONS

• US 20180095728A1
• Filed: 10/01/2016
• Published: 04/05/2018
• Est. Priority Date: 10/01/2016
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Hasenplaugh, William C., Fleming, JR., Kermin E., Fossum, Tryggve, Steely, JR., Simon C.
• Granted Patent: US 10,402,168 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 1/3237 by disabling clock generation or distribution
• CPC Class Codes: G06F 7/4876 Multiplying
• 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: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

A floating point multiply-add unit having inputs coupled to receive a floating point multiplier data element, a floating point multiplicand data element, and a floating point addend data element. The multiply-add unit including a mantissa multiplier to multiply a mantissa of the multiplier data element and a mantissa of the multiplicand data element to calculate a mantissa product. The mantissa multiplier including a most significant bit portion to calculate most significant bits of the mantissa product, and a least significant bit portion to calculate least significant bits of the mantissa product. The mantissa multiplier has a plurality of different possible sizes of the least significant bit portion. Energy consumption reduction logic to selectively reduce energy consumption of the least significant bit portion, but not the most significant bit portion, to cause the least significant bit portion to not calculate the least significant bits of the mantissa product.

Claims

1. A processor comprising:

a floating point multiply-add unit having inputs coupled to receive a floating point multiplier data element, a floating point multiplicand data element, and a floating point addend data element, the floating point multiply-add unit including;

a mantissa multiplier to multiply a mantissa of the floating point multiplier data element and a mantissa of the floating point multiplicand data element to calculate a mantissa product, the mantissa multiplier including;

a most significant bit portion to calculate most significant bits of the mantissa product;

a least significant bit portion to calculate least significant bits of the mantissa product, wherein the mantissa multiplier has a plurality of different possible sizes of the least significant bit portion; and

energy consumption reduction logic to selectively reduce energy consumption of the least significant bit portion, but not the most significant bit portion, to cause the least significant bit portion to not calculate the least significant bits of the mantissa product.

2. The processor of claim 1, further comprising logic to select one of the plurality of different possible sizes of the least significant bit portion based, at least in part, on the floating point addend data element.

3. The processor of claim 2, wherein the logic is to select the one size based, at least in part, on an exponent of the floating point addend data element.

4. The processor of claim 3, wherein the logic is to select the one size based, at least in part, on the exponent of the floating point addend data element, and a sum of an exponent of the floating point multiplier data element and an exponent of the floating point multiplicand data element.

5. The processor of claim 1, wherein the mantissa multiplier further comprises an intermediate portion, which is coupled between the most and least significant bit portions, to calculate intermediate bits of the mantissa product that are to be between the most significant bits and the least significant bits.

6. The processor of claim 5, wherein the energy consumption reduction logic comprises logic to determine, based on the intermediate bits, whether the least significant bits could have affected the most significant bits of the mantissa product.

7. The processor of claim 5, wherein the intermediate portion is in an initial portion of a pipeline and the least significant bit portion is in a subsequent portion of the pipeline.

8. The processor of claim 5, wherein the intermediate portion and the least significant bit portion are in a same portion of the pipeline.

9. The processor of claim 1, wherein the most significant bit portion is to calculate all most significant bits of the mantissa product that are to be added to a mantissa of the floating point addend data element.

10. The processor of claim 1, wherein the energy consumption reduction logic comprises clock gate logic, which is coupled to the least significant bit portion, to selectively gate a clock signal to the least significant bit portion, but not the most significant bit portion.

11. The processor of claim 1, wherein the mantissa multiplier has at least three different possible sizes of the least significant bit portion.

12. The processor of claim 1, wherein the mantissa multiplier comprises a 53-bit by 53-bit multiplier.

13. The processor of claim 1, wherein the mantissa multiplier comprises a 24-bit by 24-bit multiplier.

14. The processor of claim 1, wherein the mantissa multiplier comprises a 11-bit by 11-bit multiplier.

15. A method in a processor comprising:

starting a floating point multiply-add operation on a floating point multiplier data element, a floating point multiplicand data element, and a floating point addend data element; and

multiplying a mantissa of the floating point multiplier data element and a mantissa of the floating point multiplicand data element to calculate a mantissa product, without calculating a least significant bit portion of the mantissa product that has a size that is based, at least in part, on the floating point addend data element.

16. The method of claim 15, further comprising selecting the size based, at least in part, on an exponent of the floating point addend data element, and a sum of an exponent of the floating point multiplier data element and an exponent of the floating point multiplicand data element.

17. The method of claim 15, further comprising selecting the size from a plurality of different possible sizes based, at least in part, on the floating point addend data element.

18. The method of claim 15, further comprising:

calculating an intermediate portion of the mantissa product between the least significant bit portion of the mantissa product and a most significant bit portion of the mantissa product; and

determining based on the intermediate portion whether least significant bit portion of the mantissa product could have affected the most significant bit portion of the mantissa product.

19. The method of claim 15, further comprising selectively reducing energy consumption of a portion of a mantissa multiplier that corresponds to the least significant portion of the mantissa product, without reducing energy consumption of a portion of the mantissa multiplier that corresponds to a most significant portion of the mantissa product.

20. The method of claim 19, wherein said selectively reducing the energy consumption comprises clock gating the portion of the mantissa multiplier that corresponds to the least significant portion of the mantissa product.

21. A computer system comprising:

an interconnect;

a memory coupled with the interconnect; and

a processor coupled with the interconnect, the processor comprising;

a floating point multiply-add unit having inputs coupled to receive a floating point multiplier data element, a floating point multiplicand data element, and a floating point addend data element, the floating point multiply-add unit including;

a mantissa multiplier to multiply a mantissa of the floating point multiplier data element and a mantissa of the floating point multiplicand data element to calculate a mantissa product, the mantissa multiplier including;

a most significant bit portion to calculate most significant bits of the mantissa product;

a least significant bit portion to calculate least significant bits of the mantissa product, wherein the mantissa multiplier has a plurality of different possible sizes of the least significant bit portion; and

energy consumption reduction logic to selectively reduce energy consumption of the least significant bit portion, but not the most significant bit portion, to cause the least significant bit portion to not calculate the least significant bits of the mantissa product.

22. The computer system of claim 21, further comprising logic to select one of the plurality of different possible sizes of the least significant bit portion based, at least in part, on the floating point addend data element.

23. The computer system of claim 21, wherein the mantissa multiplier further comprises an intermediate portion, which is coupled between the most and least significant bit portions, to calculate intermediate bits of the mantissa product that are to be between the most significant bits and the least significant bits, and wherein the energy consumption reduction logic comprises logic to determine, based on the intermediate bits, whether the least significant bits could have affected the most significant bits of the mantissa product.

24. The computer system of claim 21, wherein the intermediate portion is in an initial portion of a pipeline and the least significant bit portion is in a subsequent portion of the pipeline.

25. The computer system of claim 21, wherein the energy consumption reduction logic comprises clock gate logic, which is coupled to the least significant bit portion, to selectively gate a clock signal to the least significant bit portion, but not the most significant bit portion.