# US 2019/0317741 A1 · Methods and apparatus to improve utilization of a heterogeneous system executing software

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

- Number: US 2019/0317741 A1
- Document ID: US20190317741A1
- Status: Published application
- Kind: A1
- Application number: 16/455,628
- Filed: 2019-06-27
- Published: 2019-10-17
- Earliest claimed: 2019-06-27
- Assignees: Intel Corporation
- Applicants: Intel Corporation
- Inventors: Adam Herr; Sridhar Sharma; Mikael Bourges-Sevenier; Justin Gottschlich
- Web page: https://patentgrep.com/patent/US20190317741A1

## CPC

- G06F 8/41: Compilation
- G06F 9/5027: {the resource being a machine, e.g. CPUs, Servers, Terminals}
- G06F 9/5044: {considering hardware capabilities}
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/0985: Hyperparameter optimisation; Meta-learning; Learning-to-learn
- G06F 9/48: Program initiating; Program switching, e.g. by interrupt
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/045: Combinations of networks
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Methods, apparatus, systems and articles of manufacture are disclosed improve utilization of a heterogeneous system executing software. The disclosed methods, apparatus, systems and articles of manufacture include an apparatus comprising a variant manager to determine whether an algorithm is a candidate for sub-algorithmic partitioning (SAP) based on at least one of a first size of input data to the algorithm and a second size of output data from the algorithm; a partitioner to partition the algorithm into at least a first tile and a second tile; and a compiler to compile a first variant based on the first tile and a second variant based on the second tile into an executable file, the first variant to be executed on a first processing element of the heterogeneous system, the second variant to be executed on a second processing element of the heterogeneous system.

## Claims

### Claim 1

An apparatus to improve utilization of a heterogeneous system executing software, the apparatus comprising:
a variant manager to determine whether an algorithm is a candidate for sub-algorithmic partitioning (SAP) based on at least one of a first size of input data to the algorithm and a second size of output data from the algorithm;
a partitioner to partition the algorithm into at least a first tile and a second tile; and
a compiler to compile a first variant based on the first tile and a second variant based on the second tile into an executable file, the first variant to be executed on a first processing element of the heterogeneous system, the second variant to be executed on a second processing element of the heterogeneous system.

### Claim 2 (depends on 1)

The apparatus of claim 1, wherein when at least one of the first size and the second size meet a threshold value, the variant manager is to determine that the algorithm is a candidate for SAP.

### Claim 3 (depends on 1)

The apparatus of claim 1, wherein the partitioner is to partition the algorithm into at least the first tile and the second tile based on a deep neural network.

### Claim 4 (depends on 1)

The apparatus of claim 1, wherein the partitioner is to generate the first tile and the second tile based on a first partitioning strategy, the partitioner to generate a third tile and a fourth tile based on a second partitioning strategy different than the first partitioning strategy.

### Claim 5 (depends on 4)

The apparatus of claim 4, wherein the first tile includes a first buffer associated with at least a set of input data to the algorithm; and wherein when the third tile includes a second buffer with at least the set of input data to the algorithm, the first variant is to be executed on the first processing element to implement the third tile.

### Claim 6 (depends on 1)

The apparatus of claim 1, wherein the first tile includes a first buffer associated with at least a first set of input data to the algorithm and the second tile includes a second buffer associated with at least a second set of input data to the algorithm, the second set of input data larger than the first set of input data.

### Claim 7 (depends on 1)

The apparatus of claim 1, wherein the variant manager is to determine that the heterogeneous system includes the first processing element and the second processing element.

### Claim 8

A non-transitory computer readable storage medium comprising instructions which, when executed, cause at least one processor to at least:
determine whether an algorithm is a candidate for sub-algorithmic partitioning (SAP) based on at least one of a first size of input data to the algorithm and a second size of output data from the algorithm;
partition the algorithm into at least a first tile and a second tile; and
compile a first variant based on the first tile and a second variant based on the second tile into an executable file, the first variant to be executed on a first processing element of a heterogeneous system, the second variant to be executed on a second processing element of the heterogeneous system.

### Claim 9 (depends on 8)

The non-transitory computer readable storage medium of claim 8, wherein the instructions, when executed, cause the at least one processor to, when at least one of the first size and the second size meet a threshold value, determine that the algorithm is a candidate for SAP.

### Claim 10 (depends on 8)

The non-transitory computer readable storage medium of claim 8, wherein the instructions, when executed, cause the at least one processor to partition the algorithm into at least the first tile and the second tile based on a deep neural network.

### Claim 11 (depends on 8)

The non-transitory computer readable storage medium of claim 8, wherein the instructions, when executed, cause the at least one processor to generate the first tile and the second tile based on a first partitioning strategy and a third tile and a fourth tile based on a second partitioning strategy different than the first partitioning strategy.

### Claim 12 (depends on 11)

The non-transitory computer readable storage medium of claim 11, wherein the first tile includes a first buffer associated with at least a set of input data to the algorithm; and wherein when the third tile includes a second buffer with at least the set of input data to the algorithm, the first variant is to be executed on the first processing element to implement the third tile.

### Claim 13 (depends on 8)

The non-transitory computer readable storage medium of claim 8, wherein the first tile includes a first buffer associated with at least a first set of input data to the algorithm and the second tile includes a second buffer associated with at least a second set of input data to the algorithm, the second set of input data larger than the first set of input data.

### Claim 14 (depends on 8)

The non-transitory computer readable storage medium of claim 8, wherein the instructions, when executed, cause the at least one processor to determine that the heterogeneous system includes the first processing element and the second processing element.

### Claim 15

An apparatus to improve utilization of a heterogeneous system executing software, the apparatus comprising:
means for managing, the means for managing to determine whether an algorithm is a candidate for sub-algorithmic partitioning (SAP) based on at least one of a first size of input data to the algorithm and a second size of output data from the algorithm;
means for partitioning, the means for partitioning to partition the algorithm into at least a first tile and a second tile; and
means for compiling, the means for compiling to compile a first variant based on the first tile and a second variant based on the second tile into an executable file, the first variant to be executed on a first processing element of the heterogeneous system, the second variant to be executed on a second processing element of the heterogeneous system.

### Claim 16 (depends on 15)

The apparatus of claim 15, wherein when at least one of the first size and the second size meet a threshold value, the means for managing is to determine that the algorithm is a candidate for SAP.

### Claim 17 (depends on 15)

The apparatus of claim 15, wherein the means for partitioning is to partition the algorithm into at least the first tile and the second tile based on a deep neural network.

### Claim 18 (depends on 15)

The apparatus of claim 15, wherein the means for partitioning is to generate the first tile and the second tile based on a first partitioning strategy and a third tile and a fourth tile based on a second partitioning strategy different than the first partitioning strategy.

### Claim 19 (depends on 18)

The apparatus of claim 18, wherein the first tile includes a first buffer associated with at least a set of input data to the algorithm; and wherein when the third tile includes a second buffer with at least the set of input data to the algorithm, the first variant is to be executed on the first processing element to implement the third tile.

### Claim 20 (depends on 15)

The apparatus of claim 15, wherein the first tile includes a first buffer associated with at least a first set of input data to the algorithm and the second tile includes a second buffer associated with at least a second set of input data to the algorithm, the second set of input data larger than the first set of input data.

### Claim 21 (depends on 15)

The apparatus of claim 15, wherein the means for managing is to determine that the heterogeneous system includes the first processing element and the second processing element.

### Claim 22

A method to improve utilization of a heterogeneous system executing software, the method comprising:
determining whether an algorithm is a candidate for sub-algorithmic partitioning (SAP) based on at least one of a first size of input data to the algorithm and a second size of output data from the algorithm;
partitioning the algorithm into at least a first tile and a second tile; and
compiling a first variant based on the first tile and a second variant based on the second tile into an executable file, the first variant to be executed on a first processing element of a heterogeneous system, the second variant to be executed on a second processing element of the heterogeneous system.

### Claim 23 (depends on 22)

The method of claim 22, further including determining that the algorithm is a candidate for SAP when at least one of the first size and the second size meet a threshold value.

### Claim 24 (depends on 22)

The method of claim 22, further including partitioning the algorithm into at least the first tile and the second tile based on a deep neural network.

### Claim 25 (depends on 23)

The method of claim 23, further including generating the first tile and the second tile based on a first partitioning strategy and a third tile and a fourth tile based on a second partitioning strategy different than the first partitioning strategy.
