Instruction and logic for detecting numeric accumulation error

1. A processor, comprising:

circuitry to decode at least one instruction, the instruction to compute a floating point result associated with a floating point operation;

an execution unit including circuitry to:

execute the instruction to determine the floating point result;

compute an amount of precision lost in a mantissa of the floating point result relative to an actual result of the floating point operation, the amount of precision lost corresponding to a plurality of bits lost in the mantissa of the floating point result;

compare the amount of precision lost in the mantissa of the floating point result to a numeric accumulation error precision threshold;

determine whether a numeric accumulation error occurred based on the comparison between the amount of precision lost in the mantissa of the floating point result and the numeric accumulation error precision threshold; and

write a value to a flag for notification that the numeric accumulation error occurred, the value based on a determination that the numeric accumulation error occurred.

2. The processor of claim 1 , wherein:

the execution unit further includes circuitry to determine the numeric accumulation error precision threshold based on the instruction, a previous instruction, or a floating point control register; and

the comparison between the amount of precision lost and the numeric accumulation error precision threshold is based on the numeric accumulation error precision threshold that is determined.

3. The processor of claim 1 , wherein the execution unit further includes circuitry to:

determine whether a mask bit is set to prevent signaling of the numeric accumulation error; and

signal an exception based on a determination that the mask bit is not set and based on the determination that the numeric accumulation error occurred.

4. The processor of claim 1 , wherein:

the execution unit further includes circuitry to:

round the floating point result;

store the rounded floating point result; and

the amount of precision lost in the mantissa of the floating point result represents a percentage of bits in the mantissa of the floating point result that is lost, the percentage of bits that is lost includes:

a percentage of bits that is lost when the floating point result is rounded; or

a percentage of bits that is lost when the rounded floating point result is stored.

5. The processor of claim 1 , the execution unit further comprising circuitry to:

determine a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

control, using the numeric accumulation error non-zero precision flag, whether to ignore at least one trailing bit of the mantissa of the floating point result with a value of zero, wherein the computation of the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

6. The processor of claim 1 , the execution unit further includes circuitry to:

determine a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

control, using the numeric accumulation error non-zero precision flag, whether to ignore bits of the mantissa of the floating point result with a value of zero, wherein the computation of the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

7. The processor of claim 1 , wherein:

the floating point result is computed from source values;

the instruction is a fused multiply-add instruction; and

the execution of the instruction includes circuitry to:

compute a sum based on the source values; and

compute the floating point result based on the sum and at least one of the source values.

8. A method for detecting numeric accumulation error, comprising:

decoding at least one instruction, the instruction for computing a floating point result associated with a floating point operation;

executing the instruction to determine the floating point result;

computing the amount of precision lost in a mantissa of the floating point result relative to an actual result of the floating point operation, the amount of precision lost corresponding to a plurality of bits lost in the mantissa of the floating point result;

comparing the amount of precision lost in the mantissa of the floating point result to a numeric accumulation error precision threshold;

determining whether a numeric accumulation error occurred based on the comparison between the amount of precision lost in the mantissa of the floating point result and the numeric accumulation error precision threshold; and

writing a value to a flag for notification that the numeric accumulation error occurred, the value based on a determination that the numeric accumulation error occurred.

9. The method of claim 8 , further comprising:

determining the numeric accumulation error precision threshold based on the instruction, a previous instruction, or a floating point control register.

10. The method of claim 8 , further comprising:

determining whether a mask bit is set to prevent signaling of the numeric accumulation error; and

signaling an exception based on a determination that the mask bit is not set and based on the determination that the numeric accumulation error occurred.

11. The method of claim 8 , further comprising:

rounding the floating point result; and

storing the rounded floating point result;

wherein the amount of precision lost in the mantissa of the floating point result represents a percentage of bits in the mantissa of the floating point result that is lost, the percentage of bits that is lost including:

a percentage of bits that is lost when rounding the floating point result; or

a percentage of bits that is lost when storing the rounded floating point result.

12. The method of claim 8 , further comprising:

determining a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

controlling, using the numeric accumulation error non-zero precision flag, whether to ignore at least one trailing bit of the mantissa of the floating point result with a value of zero, wherein computing the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

13. The method of claim 8 , further comprising:

determining a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

controlling, using the numeric accumulation error non-zero precision flag, whether to ignore bits of the mantissa of the floating point result with a value of zero, wherein computing the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

14. An execution unit, comprising circuitry to:

execute at least one instruction to determine a floating point result associated with a floating point operation;

compute an amount of precision lost in a mantissa of the floating point result relative to an actual result of the floating point operation, the amount of precision lost corresponding to a plurality of bits lost in the mantissa of the floating point result;

compare the amount of precision lost in the mantissa of the floating point result to a numeric accumulation error precision threshold;

determine whether a numeric accumulation error occurred based on the comparison between the amount of precision lost in the mantissa of the floating point result and the numeric accumulation error precision threshold; and

write a value to a flag for notification that the numeric accumulation error occurred, the value based on a determination that the numeric accumulation error occurred.

15. The execution unit of claim 14 , the execution unit further comprising circuitry to determine the numeric accumulation error precision threshold based on the instruction, a previous instruction, or a floating point control register, wherein the comparison between the amount of precision lost and the numeric accumulation error precision threshold is based on the numeric accumulation error precision threshold that is determined.

16. The execution unit of claim 14 , further comprising circuitry to:

determine whether a mask bit is set to prevent signaling of the numeric accumulation error; and

signal an exception based on a determination that the mask bit is not set and based on the determination that the numeric accumulation error occurred.

17. The execution unit of claim 14 , further comprising circuitry to round the floating point result and store the rounded floating point result, wherein the amount of precision lost in the mantissa of the floating point result represents a percentage of bits in the mantissa of the floating point result that is lost, the percentage of bits that is lost includes:

a percentage of bits that is lost when the floating point result is rounded; or

a percentage of bits that is lost when the rounded floating point result is stored.

18. The execution unit of claim 14 , further comprising circuitry to:

determine a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

control, using the numeric accumulation error non-zero precision flag, whether to ignore at least one trailing bit of the mantissa of the floating point result with a value of zero, wherein the computation of the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

19. The execution unit of claim 14 , further comprising circuitry to:

determine a numeric accumulation error non-zero precision flag based on the instruction, a previous instruction, or a floating point control register; and

control, using the numeric accumulation error non-zero precision flag, whether to ignore bits of the mantissa of the floating point result with a value of zero, wherein the computation of the amount of precision lost in the mantissa of the floating point result is based on the numeric accumulation error non-zero precision flag.

20. The execution unit of claim 14 , wherein:

the floating point result is computed from source values;

the instruction is a fused multiply-add instruction; and

the execution of the instruction includes circuitry to:

compute a sum based on the source values; and

compute the floating point result based on the sum and at least one of the source values.