# US 2021/0279038 A1 · Using fuzzy-jbit location of floating-point multiply-accumulate results

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0279038 A1
- Document ID: US20210279038A1
- Status: Published application
- Kind: A1
- Application number: 17/330,064
- Filed: 2021-05-25
- Published: 2021-09-09
- Earliest claimed: 2019-03-29
- Assignees: Intel Corporation
- Applicants: Intel Corporation
- Inventors: Amit GRADSTEIN; Simon RUBANOVICH; Zeev SPERBER
- Web page: https://patentgrep.com/patent/US20210279038A1

## CPC

- G06F 7/5443: {Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00)}
- 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)}
- G06F 7/4836: {Computations with rational numbers}
- G06F 9/30014: {with variable precision}
- G06F 9/30025: {Format conversion instructions, e.g. Floating-Point to Integer, decimal conversion}
- G06F 9/30036: {Instructions to perform operations on packed data, e.g. vector, tile or matrix operations}
- G06F 9/30038: {using a mask}
- G06F 9/30094: {Condition code generation, e.g. Carry, Zero flag}
- G06F 9/3016: {Decoding the operand specifier, e.g. specifier format}
- G06F 7/49936: {Normalisation mentioned as feature only}
- G06F 2207/4824: Neural networks
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Disclosed embodiments relate to performing floating-point (FP) arithmetic. In one example, a processor is to decode an instruction specifying locations of first, second, and third floating-point (FP) operands and an opcode calling for accumulating a FP product of the first and second FP operands with the third FP operand, and execution circuitry to, in a first cycle, generate the FP product having a Fuzzy-Jbit format comprising a sign bit, a 9-bit exponent, and a 25-bit mantissa having two possible positions for a JBit and, in a second cycle, to accumulate the FP product with the third FP operand, while concurrently, based on Jbit positions of the FP product and the third FP operand, determining an exponent adjustment and a mantissa shift control of a result of the accumulation, wherein performing the exponent adjustment concurrently enhances an ability to perform the accumulation in one cycle.

## Claims

### Claim 1

A processor comprising:
decode circuitry to decode an instruction specifying locations of first, second, and third floating-point (FP) operands and an opcode calling for accumulating a FP product of the first and second FP operands with the third FP operand; and
execution circuitry to execute the decoded instruction to:
generate the FP product having a format comprising a sign bit, an exponent, and a mantissa having two adjacent bits that allow two possible positions for an implicit bit,
accumulate the FP product with the third FP operand, and
determine an exponent adjustment and a mantissa shift control of a result of the accumulation based on implicit bit positions of the FP product and the third FP operand.

### Claim 2 (depends on 1)

The processor of claim 1, wherein the implicit bit is an implicit 24th bit of the mantissa for a first of the two possible positions for the implicit bit, and the implicit bit is an implicit 25th bit of the mantissa for a second of the two possible positions for the implicit bit.

### Claim 3 (depends on 1)

The processor of claim 1, wherein the first, second, and third FP operands comprise packed-data vectors having multiple floating-point elements, and wherein the execution circuitry is to execute the decoded instruction on each of the multiple elements in parallel.

### Claim 4 (depends on 1)

The processor of claim 1, wherein the first, second, and third FP operands comprise matrices having M rows by N columns of floating-point elements, and wherein the execution circuitry is to execute the decoded instruction on a subset of the M×N elements in parallel.

### Claim 5 (depends on 1)

The processor of claim 1, wherein determining the implicit bit positions of the product and the third FP operand comprises identifying one or more possible implicit bit positions between an actual exponent and an actual fraction.

### Claim 6 (depends on 1)

The processor of claim 1, wherein a carryout, if any, of the accumulation falls into the possible implicit bit positions of the result and does not require adjusting the exponent of the result.

### Claim 7 (depends on 1)

The processor of claim 1, wherein the mantissa shift control of the result is selectable between a left-shift by one, a right-shift by one, and a no shift.

### Claim 8 (depends on 1)

The processor of claim 1, wherein the result, maintaining its format, is routed for use by a subsequent floating-point operation.

### Claim 9 (depends on 1)

The processor of claim 1, wherein the execution circuitry is to convert the result into Institute of Electronics and Electrical Engineers 754 standard format and write the converted result into a destination operand further specified by the instruction.

### Claim 10

A method to be performed by a processor, the method comprising:
decoding, using decode circuitry, an instruction specifying locations of first, second, and third floating-point (FP) operands and an opcode calling for accumulating a FP product of the first and second FP operands with the third FP operand; and
executing, using execution circuitry, the decoded instruction to:
generate the FP product having a format comprising a sign bit, an exponent, and a mantissa having two adjacent bits that allow two possible positions for an implicit bit,
accumulate the FP product with the third FP operand, and,
determine an exponent adjustment and a mantissa shift control of a result of the accumulation based on implicit bit positions of the FP product and the third FP operand.

### Claim 11 (depends on 10)

The method of claim 10, wherein the implicit bit is an implicit 24th bit of the mantissa for a first of the two possible positions for the implicit bit, and the implicit bit is an implicit 25th bit of the mantissa for a second of the two possible positions for the implicit bit.

### Claim 12 (depends on 10)

The method of claim 10, wherein the first, second, and third FP operands comprise packed-data vectors having multiple floating-point elements, and wherein the execution circuitry is to execute the decoded instruction on each of the multiple elements in parallel.

### Claim 13 (depends on 10)

The method of claim 10, wherein the first, second, and third FP operands comprise matrices having M rows by N columns of floating-point elements, and wherein the execution circuitry is to execute the decoded instruction on a subset of the M×N elements in parallel.

### Claim 14 (depends on 10)

The method of claim 10, wherein determining the implicit bit positions of the product and the third FP operand comprise identifying one or more possible implicit bit positions between an actual exponent and an actual fraction.

### Claim 15 (depends on 10)

The method of claim 10, wherein a carryout, if any, of the accumulation falls into the possible implicit bit positions of the result and does not require adjusting the exponent of the result.

### Claim 16 (depends on 10)

The method of claim 10, wherein the mantissa shift control of the result is selectable between a left-shift by one, a right-shift by one, and a no shift.

### Claim 17 (depends on 10)

The method of claim 10, wherein the result, maintaining its format, is routed for use by a subsequent floating-point operation.

### Claim 18 (depends on 10)

The method of claim 10, wherein the executing converts the result into Institute of Electronics and Electrical Engineers 754 standard format and writes the converted result into a destination operand further specified by the instruction.

### Claim 19

A non-transitory machine-readable medium containing code to which a processor is to respond by:
decoding, using decode circuitry, an instruction specifying locations of first, second, and third floating-point (FP) operands and an opcode calling for accumulating a FP product of the first and second FP operands with the third FP operand; and
executing, using execution circuitry, the decoded instruction to:
generate the FP product having a format comprising a sign bit, an exponent, and a mantissa having two adjacent bits that allow two possible positions for an implicit bit,
accumulate the FP product with the third FP operand, and
determine an exponent adjustment and a mantissa shift control of a result of the accumulation based on implicit bit positions of the FP product and the third FP operand.

### Claim 20 (depends on 19)

The non-transitory machine-readable medium of claim 19, wherein the implicit bit is an implicit 24th bit of the mantissa for a first of the two possible positions for the implicit bit, and the implicit bit is an implicit 25th bit of the mantissa for a second of the two possible positions for the implicit bit.

### Claim 21 (depends on 19)

The non-transitory machine-readable medium of claim 19, wherein the mantissa shift control of the result is selectable between a left-shift by one, a right-shift by one, and a no shift.

### Claim 22 (depends on 19)

The non-transitory machine-readable medium of claim 19, wherein the result, maintaining its format, is routed for use by a subsequent floating-point operation.

### Claim 23 (depends on 19)

The non-transitory machine-readable medium of claim 19, wherein the executing converts the result into Institute of Electronics and Electrical Engineers 754 standard format and writes the converted result into a destination operand further specified by the instruction.
