BFLOAT16 fused multiply instructions
Techniques for performing BF16 FMA in response to an instruction are described. In some examples, an instruction has fields for an opcode, an identification of location of a packed data source/destination operand (a first source), an identification of a location of a second packed data source operand, an identification of a location of a third packed data source operand, and an identification of location of a packed data source/destination operand, wherein the opcode is to indicate operand ordering and that execution circuitry is to, per data element position, perform a BF16 value fused multiply-accumulate operation using the first, second, and third source operands and store a result in a corresponding data element position of the source/destination operand.
1. An apparatus comprising:
decoder circuitry to decode a single instruction, the single instruction to include fields for an opcode, an identification of location of a packed data source/destination operand (a first source), an identification of a location of a second packed data source operand, an identification of a location of a third packed data source operand, and an identification of location of a packed data source/destination operand, wherein the opcode is to indicate operand ordering and that execution circuitry is to, per data element position, perform a BF16 value fused multiply-accumulate operation using the first, second, and third source operands and store a result in a corresponding data element position of the source/destination operand; and
execution circuitry to execute the decoded single instruction according to the opcode.
2. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element position multiplication of the BF16 value from the first source operand with the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 values in the second source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
3. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand with the BF16 value in the first source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 value in the third source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
4. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 value in the first source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
5. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand to the BF16 value in the first source operand to generate an infinite precision intermediate result, an addition of a the BF16 value in the third source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
6. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of a the BF16 value in the second source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
7. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition the BF16 value in the first source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
8. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the second source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
9. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand to the BF16 value in the first source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the third source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
10. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the first source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
11. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the second source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
12. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand to the BF16 value in the first source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the third source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
13. The apparatus of claim 1 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the first source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
14. The apparatus of claim 1 , wherein the identified first and second source operands are registers.
15. The apparatus of claim 1 , further comprising:
memory to store the single instruction.
16. A method comprising:
decoding a single instruction, the single instruction to include fields for an opcode, an identification of location of a packed data source/destination operand (a first source), an identification of a location of a second packed data source operand, an identification of a location of a third packed data source operand, and an identification of location of a packed data source/destination operand, wherein the opcode is to indicate operand ordering and that execution circuitry is to, per data element position, perform a BF16 value fused multiply-accumulate operation using the first, second, and third source operands and store a result in a corresponding data element position of the source/destination operand; and
executing the decoded single instruction according to the opcode.
17. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element position multiplication of the BF16 value from the first source operand with the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 values in the second source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
18. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand with the BF16 value in the first source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 value in the third source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
19. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of the infinite precision intermediate result negated to the BF16 value in the first source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
20. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition of a the BF16 value in the second source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
21. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the second source operand to generate an infinite precision intermediate result, an addition of a the BF16 value in the third source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
22. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, an addition the BF16 value in the first source operand to generate an infinite precision addition result, and a round of the infinite precision addition result.
23. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the second source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
24. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand to the BF16 value in the first source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the third source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
25. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the infinite precision intermediate result negated to the BF16 value in the first source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
26. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the first source operand to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the second source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
27. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source operand to the BF16 value in the first source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the third source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
28. The method of claim 16 , wherein the opcode is to indicate the fused multiply accumulation operation is per data element multiplication of the BF16 value from the second source to the BF16 value in the third source operand to generate an infinite precision intermediate result, a subtraction of the BF16 value in the first source operand to generate an infinite precision subtraction result, and a round of the infinite precision addition result.
29. The method of claim 16 , wherein the identified first and second source operands are registers.
30. The method of claim 16 , further comprising:
translating the single instruction to at least one instruction of a different instruction set architecture, wherein executing the decoded single instruction according to the opcode comprises executing the at least one instruction of the different instruction set architecture.