Exponent Processing Systems and Methods

• US 20090037504A1
• Filed: 08/02/2007
• Published: 02/05/2009
• Est. Priority Date: 08/02/2007
• Status: Active Grant

• Current Assignee: VIA Technologies Incorporated (VIA Technologies)
• Original Assignee: VIA Technologies Incorporated (VIA Technologies)
• Inventors: Hussain, Zahid
• Granted Patent: US 7,912,883 B2
• Time in Patent Office: Days
• US Class Current: 708/277
• CPC Class Codes: G06F 1/0307 Logarithmic or exponential functions G06F1/0314, G06F1/035 take precedence
• CPC Class Codes: G06F 7/556 Logarithmic or exponential functions

Abstract

Embodiments of exponent processing systems and methods are disclosed. One method embodiment, among others, comprises performing a first table lookup using a first address to provide a first value corresponding to the first component part, setting an integer exponent to provide an integer-based value corresponding to the integer component, performing a second table lookup using a second and third address to provide a second value and a third value corresponding to the second component part and the third component part, respectively, expanding and normalizing the second and third values to provide expanded and normalized second and third values, combining the expanded and normalized second and third values to produce a first product, and computing the exponential function by combining the first value, the integer-based value, and the first product.

Claims

1. An exponent processing method for evaluating a base 2 exponential function 2^x, where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y0×

2^y1×

2^y2, and where 2^y0, 2^y1, and 2^y2 correspond to first, second, and third component parts, respectively, of the fractional component, the method comprising;

performing a first table lookup using a first address to provide a first value corresponding to the first component part;

setting an integer exponent to provide an integer-based value corresponding to the integer component;

performing a second table lookup using a second and third address to provide a second value and a third value corresponding to the second component part and the third component part, respectively;

expanding and normalizing the second and third values to provide expanded and normalized second and third values;

combining the expanded and normalized second and third values to produce a first product; and

computing the exponential function by combining the first value, the integer-based value, and the first product.

2. The method of claim 1, further comprising extracting a fraction part and an integer part from a shift register storing a mantissa value, the mantissa value derived from a source register.

3. The method of claim 2, wherein extracting comprises performing a shift operation on the mantissa value stored in the shift register.

4. The method of claim 3, wherein performing the shift further comprises performing a 1s-complement operation on the mantissa value stored in the shift register.

5. The method of claim 4, wherein performing the 1s-complement operation further comprises using a biased exponent bit derived from the source register.

6. The method of claim 2, further comprising either negating the fraction part and storing in a fractional register when the source register comprises a sign bit indicating a negative value, or storing the fraction part, un-negated, in the fractional register when the source register comprises a sign bit indicating a positive value.

7. The method of claim 6, wherein performing the first and second table lookups comprises using first, second, and third segments of the fractional register as the first, second, and third addresses.

8. The method of claim 2, wherein setting the integer exponent comprises either negating the integer part and biasing the negated value and storing the biased, negated value in an integer register when the source register comprises a sign bit indicating a negative value, or biasing an un-negated integer part and storing the biased, un-negated part in the integer register when the source register comprises a sign bit indicating a positive value.

9. The method of claim 8, wherein setting the integer exponent further comprises designating the integer exponent with a value in the integer register.

10. The method of claim 9, wherein setting the integer exponent further comprises handling exceptions pertaining to overflow and underflow conditions, wherein handling exceptions comprises determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

11. The method of claim 10, wherein setting the integer exponent further comprises resetting the integer exponent to a computed value responsive to determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

12. The method of claim 1, wherein expanding and normalizing comprises formatting in IEEE-754 floating point format.

13. The method of claim 1, further comprising storing the expanded and normalized second and third values in a first register and a second register.

14. The method of claim 13, wherein expanding and normalizing comprises adding leading zeroes and biasing an exponent portion in the first and second registers.

15. The method of claim 1, wherein combining the expanded and normalized values comprises multiplying the expanded and normalized second value with the expanded and normalized third value in a fused multiply and add (FMAD) unit, and storing the first product resulting from the multiplication in a temporary register.

16. The method of claim 1, wherein performing a first table lookup and a second table lookup comprises performing the first and second table lookups in a single table corresponding to a plurality of N single-instruction, multiple data (SIMD) units by limiting an issue rate to one every N cycles.

17. The method of claim 1, wherein performing the second table lookup comprises performing the table lookups in two tables in parallel.

18. The method of claim 1, further comprising storing the second and third values in a single register.

19. An exponent processing method for evaluating a base 2 exponential function 2^x using a bipartite table method where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y1×

2^y2, where 2^y1 and 2^y2 correspond to first and second component parts, respectively, of the fractional component, the method comprising;

performing a first table lookup using a first address to provide a first value corresponding to the first component part;

setting an integer exponent to provide an integer-based value corresponding to the integer component;

performing a second table lookup using a second address to provide a second value corresponding to the second component part;

expanding and normalizing the second value to provide an expanded and normalized second value; and

computing the exponential function by combining the first value, the expanded and normalized second value, and the integer-based value.

20. The method of claim 19, further comprising extracting a fraction part and an integer part from a shift register storing a mantissa value, the mantissa value derived from a source register.

21. The method of claim 20, wherein extracting comprises performing a shift operation on the mantissa value stored in the shift register.

22. The method of claim 21, wherein performing the shift further comprises performing a 1s-complement operation on the mantissa value stored in the shift register.

23. The method of claim 22, wherein performing the 1s-complement operation comprises using a biased exponent bit derived from the source register.

24. The method of claim 20, further comprising either negating the fraction part and storing in a fractional register when the source register comprises a sign bit indicating a negative value, or storing the fraction part, un-negated, in the fractional register when the source register comprises a sign bit indicating a positive value.

25. The method of claim 24, wherein performing the first and second table lookups comprises using first and second segments of the fractional register as the first and second addresses.

26. The method of claim 20, wherein setting the integer exponent comprises either negating the integer part and biasing the negated value and storing the biased, negated value in an integer register when the source register comprises a sign bit indicating a negative value, or biasing an un-negated integer part and storing the biased, un-negated part in the integer register when the source register comprises a sign bit indicating a positive value.

27. The method of claim 26, wherein setting the integer exponent further comprises designating the integer exponent with a value in the integer register.

28. The method of claim 27, wherein setting the integer exponent further comprises handling exceptions pertaining to overflow and underflow conditions, wherein handling exceptions comprises determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

29. The method of claim 28, wherein setting the integer exponent further comprises resetting the integer exponent to a computed value responsive to determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

30. The method of claim 19, wherein expanding and normalizing comprises formatting in IEEE-754 floating point format.

31. The method of claim 19, further comprising storing the expanded and normalized second value in a register.

32. The method of claim 31, wherein expanding and normalizing comprises adding leading zeroes and biasing an exponent portion in the register.

33. The method of claim 19, wherein computing by combining comprises multiplying the expanded and normalized second value, the first value, and the second value in a fused multiply and add (FMAD) unit.

34. The method of claim 19, wherein performing a first table lookup and a second table lookup comprises performing the first and second table lookups in a single table corresponding to a plurality of N single-instruction, multiple data (SIMD) units by limiting an issue rate to one every N cycles.

35. An exponent processing method for evaluating a base 2 exponential function 2^x, where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y0, and where 2^y0 corresponds to a first component part of the fractional component, the method comprising;

performing a first table lookup using a first address derived from a first portion of a mantissa derived from a source register to provide a value corresponding to the first component part;

based on another portion of the mantissa, setting an integer exponent to provide an integer-based value corresponding to the integer component; and

designating the integer-based value and the integer exponent as a solution to the exponential function.

36. An exponent processing system for evaluating a base 2 exponential function 2^x, where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y0×

2^y1×

2^y2, and where 2^y0, 2^y1, and 2^y2 correspond to first, second, and third component parts, respectively, of the fractional component, the system comprising;

register logic configured to derive a fraction part and integer part from a source mantissa;

memory configured to provide, based on the fraction part, a first value corresponding to the first component, a second value corresponding to the second component part, and a third value corresponding to the third component part;

setting logic configured to set an integer exponent based on the integer part to provide an integer-based value corresponding to the integer component;

formatting logic configured to expand and normalize the second and third values to provide expanded and normalized second and third values; and

combining logic configured to combine the expanded and normalized second and third values to produce a first product and compute the exponential function by combining the first value, the integer-based value, and the first product.

37. The system of claim 36, further comprising a source register configured to store the source mantissa and an exponent bit.

38. The system of claim 37, further comprising subtractor logic configured to bias the exponent bit, wherein the register logic is configured to receive the biased exponent bit and perform a shift operation and 1s-complement operation on the source mantissa.

39. The system of claim 36, wherein the register logic is further configured to either negate the fraction part and store the negated fraction part when the source register comprises a sign bit indicating a negative value, or store the fraction part, un-negated, when the source register comprises a sign bit indicating a positive value.

40. The system of claim 36, wherein the memory comprises a first table, a second table, and a third table.

41. The system of claim 36, wherein the setting logic comprises an integer register, the setting logic configured to set the integer exponent by either negating the integer part and biasing the negated value and storing the biased, negated value in the integer register when the source register comprises a sign bit indicating a negative value, or by biasing an un-negated integer part and storing the biased, un-negated part in the integer register when the source register comprises a sign bit indicating a positive value.

42. The system of claim 41, wherein the setting logic is configured to set the integer exponent by designating the integer exponent with a value in the integer register.

43. The system of claim 42, wherein the setting logic further comprises exception handling logic, the exception handling logic configured to handle exceptions pertaining to overflow and underflow conditions by determining whether the value in the integer register corresponds to an overflow condition or an underflow condition and resetting the integer exponent to a computed value responsive to determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

44. The system of claim 36, wherein the formatting logic comprises wire-OR logic configured to expand and normalize according to IEEE-754 floating point format.

45. The system of claim 44, wherein the formatting logic comprises second and third registers, the second and third registers configured to store the expanded and normalized second and third values by adding leading zeroes and biasing an exponent portion in the second and third registers.

46. The system of claim 36, wherein the combining logic comprises a fused multiply and add (FMAD) unit and a temporary register, the FMAD unit configured to multiply the expanded and normalized second value with the expanded and normalized third value, the temporary register configured to store the first product resulting from the multiplication.

47. The system of claim 36, wherein the memory comprises a single table configured for N single-instruction, multiple data (SIMD) operations.

48. The system of claim 36, further comprising a single register configured to store the second and third values.

49. An exponent processing system for evaluating a base 2 exponential function 2^x using a bipartite table method (BPTM) where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y1×

2^y2, where 2^y1 and 2^y2 correspond to first and second component parts, respectively, of the fractional component, the system comprising;

register logic configured to derive a fraction part and integer part from a source mantissa;

a BPTM memory configured to provide, based on the fraction part, a first value corresponding to the first component and a second value corresponding to the second component part;

setting logic configured to set an integer exponent based on the integer part to provide an integer-based value corresponding to the integer component;

formatting logic configured to expand and normalize the second value to provide an expanded and normalized second value; and

combining logic configured to compute the exponential function by combining the first value, the expanded and normalized second value, and the integer-based value.

50. The system of claim 49, further comprising a source register configured to store the source mantissa and an exponent bit.

51. The system of claim 50, further comprising subtractor logic configured to bias the exponent bit, wherein the register logic is configured to receive the biased exponent bit and perform a shift operation and 1s-complement operation on the source mantissa.

52. The system of claim 49, wherein the register logic is further configured to either negate the fraction part and store the negated fraction part when the source register comprises a sign bit indicating a negative value, or store the fraction part, un-negated, when the source register comprises a sign bit indicating a positive value.

53. The system of claim 49, wherein the BPTM memory comprises a first BPTM table and a second BPTM table.

54. The system of claim 49, wherein the setting logic comprises an integer register, the setting logic configured to set the integer exponent by either negating the integer part and biasing the negated value and storing the biased, negated value in the integer register when the source register comprises a sign bit indicating a negative value, or by biasing an un-negated integer part and storing the biased, un-negated part in the integer register when the source register comprises a sign bit indicating a positive value.

55. The system of claim 54, wherein the setting logic is configured to set the integer exponent by designating the integer exponent with a value in the integer register.

56. The system of claim 55, wherein the setting logic further comprises exception handling logic, the exception handling logic configured to handle exceptions pertaining to overflow and underflow conditions by determining whether the value in the integer register corresponds to an overflow condition or an underflow condition and resetting the integer exponent to a computed value responsive to determining whether the value in the integer register corresponds to an overflow condition or an underflow condition.

57. The system of claim 49, wherein the formatting logic comprises wire-OR logic configured to expand and normalize according to IEEE-754 floating point format.

58. The system of claim 57, wherein the formatting logic comprises a register, the register configured to store the expanded and normalized second value by adding leading zeroes and biasing an exponent portion in the register.

59. The system of claim 49, wherein the combining logic comprises a fused multiply and add (FMAD) unit.

60. The system of claim 49, wherein the BPTM memory comprises a single table configured for N single-instruction, multiple data (SIMD) operations.

61. An exponent processing system for evaluating a base 2 exponential function 2^x, where 2^x=2^I×

2^y, where 2^I corresponds to an integer component, where 2^y corresponds to a fractional component, where 2^y=2^y0, and where 2^y0 corresponds to a first component part of the fractional component, the system comprising;

register logic configured to derive a fraction part and integer part from a source mantissa;

memory configured to provide, based on the fraction part, a value corresponding to the first component part;

setting logic configured to set an integer exponent, based on the integer part, to provide an integer-based value corresponding to the integer component and designate the value and the integer-based value as a solution to the exponential function.