Generalized acceleration of matrix multiply accumulate operations

Patent metadata

Field	Value
Publication number	US-11797301-B2
Application number	US-202117141082-A
Country	US
Kind code	B2
Filing date	Jan 4, 2021
Priority date	May 8, 2017
Publication date	Oct 24, 2023
Grant date	Oct 24, 2023

Abstract

A method, computer readable medium, and processor are disclosed for performing matrix multiply and accumulate (MMA) operations. The processor includes a datapath configured to execute the MMA operation to generate a plurality of elements of a result matrix at an output of the datapath. Each element of the result matrix is generated by calculating at least one dot product of corresponding pairs of vectors associated with matrix operands specified in an instruction for the MMA operation. A dot product operation includes the steps of: generating a plurality of partial products by multiplying each element of a first vector with a corresponding element of a second vector; aligning the plurality of partial products based on the exponents associated with each element of the first vector and each element of the second vector; and accumulating the plurality of aligned partial products into a result queue utilizing at least one adder.

First claim

What is claimed is:

1. A multi-threaded processor, comprising:
a decoder to decode a floating point matrix multiply and accumulate (MMA) instruction;
a buffer to store 16-bit floating point matrix data specified by operands of the MMA instruction;
a scheduler to schedule the MMA instruction;
a fused multiply accumulate (FMA) unit to perform a dot product of corresponding pairs of the 16-bit floating point matrix data;
logic to shift a plurality of partial products of the dot product;
an arithmetic logic unit (ALU) to add the plurality of partial products to be accumulated into a register; and
memory to store a result of the MMA instruction.

2. The multi-threaded processor of claim 1 , wherein the logic to shift the plurality of partial products of the dot product aligns the plurality of partial products based on exponents associated with each element of the corresponding pairs.

3. The multi-threaded processor of claim 1 , further comprising a tree of adders, where the tree of adders comprises at least a 3:2 carry sum adder (CSA).

4. The multi-threaded processor of claim 1 , further comprising a dispatch unit to transmit the MMA instruction to the FMA unit.

5. The multi-threaded processor of claim 1 , further comprising a register file to provide the register.

6. The multi-threaded processor of claim 1 , further comprising a special function unit (SFU).

7. The multi-threaded processor of claim 1 , further comprising an interconnect to connect the ALU and the register.

8. A system comprising the multi-threaded processor of claim 1 , wherein the system further comprises:
a system bus to connect the multi-threaded processor to one or more peripheral devices; and
one or more dynamic random access memory (DRAM) devices.

9. A single instruction multiple data (SIMD) multi-threaded processor, comprising:
a plurality of cores to perform a floating point matrix multiply and accumulate (MMA) instruction, where each of the plurality of cores comprises:
a front end to fetch the MMA instruction;
an instruction cache to store the MMA instruction;
an L1 cache to store data;
an L2 cache to store data;
a plurality of ports to read from and write to a memory;
one or more load/store units to read and write the memory;
an interconnect to couple the memory and the plurality of cores;
a decoder to decode the MMA instruction;
a buffer to store 16-bit floating point matrix data specified by operands of the MMA instruction;
a scheduler to schedule the MMA instruction;
a fused multiply accumulate (FMA) unit to perform a dot product of corresponding pairs of the 16-bit floating point matrix data;
logic to shift a plurality of partial products of the dot product;
an arithmetic logic unit (ALU) to add the plurality of partial products to be accumulated into a register; and
wherein the memory is to store a result of the MMA instruction.

10. The SIMD multi-threaded processor of claim 9 , wherein the L1 cache comprises at least 24 kilobytes (KB) of storage.

11. The SIMD multi-threaded processor of claim 9 , wherein the memory comprises at least 64 kilobytes (KB) of storage.

12. The SIMD multi-threaded processor of claim 9 , wherein the interconnect connects the one or more load/store units to the register.

13. The SIMD multi-threaded processor of claim 9 , wherein the scheduler dispatches the MMA instruction to one or more cores of the plurality of cores.

14. The SIMD multi-threaded processor of claim 9 , wherein the logic to shift the plurality of partial products of the dot product aligns the plurality of partial products based on exponents associated with each element of the corresponding pairs.

15. A computer-implemented method, comprising:
decoding, by a decoder, a floating point matrix multiply and accumulate (MMA) instruction;
storing, by a buffer, 16-bit floating point matrix data specified by operands of the MMA instruction;
scheduling, by a scheduler, the MMA instruction;
performing, by a fused multiply accumulate (FMA) unit, a dot product of corresponding pairs of the 16-bit floating point matrix data;
shifting, by logic to shift, a plurality of partial products of the dot product;
adding, by an arithmetic logic unit (ALU), the plurality of partial products to be accumulated into a register; and
storing, by a memory, a result of the MMA instruction.

16. The computer-implemented method of claim 15 , wherein the ALU comprises at least one adder.

17. The computer-implemented method of claim 15 , wherein shifting the plurality of partial products further comprises aligning the plurality of partial products based on exponents associated with each element of the corresponding pairs.

18. The computer-implemented method of claim 15 , wherein the scheduler comprises a dispatch unit to dispatch the MMA instruction.

19. The computer-implemented method of claim 15 , further comprising accumulating the plurality of partial products into the register using an interconnect.

20. The computer-implemented method of claim 15 , wherein a register file provides the register.

Assignees

• Nvidia Corp

Inventors

• Boswell Brent Ralph
• Siu Ming Y
• Choquette Jack H
• Alben Jonah M
• Oberman Stuart

Classifications

• G06F9/3888

  controlled by a single instruction for multiple threads [SIMT] in parallel · CPC title

• G06F9/3851

  from multiple instruction streams, e.g. multistreaming · CPC title

• G06F9/30036

  Instructions to perform operations on packed data, e.g. vector, tile or matrix operations · CPC title

• G06F9/30014 Primary

  with variable precision · CPC title

• G06F9/3001

  Arithmetic instructions · CPC title