Placement and scheduling of radio signal processing dataflow operations

Inventors

O'Shea, Timothy James

Assignees

DeepSig Inc

Interested in licensing this patent?

MTEC can help explore whether this patent might be available for licensing for your application.

Publication Number

US-10841810-B2

Patent

Publication Date

2020-11-17

Expiration Date


Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for placement and scheduling of radio signal processing dataflow operations. An example method provides a primitive radio signal processing computational dataflow graph that comprises nodes representing operations and directed edges representing data flow. The nodes and directed edges of the primitive radio signal processing computational dataflow graph are partitioned to produce a set of software kernels that, when executed on processing units of a target hardware platform, achieve a specific optimization objective. Runtime resource scheduling, including data placement for individual software kernels in the set of software kernels are performed to efficiently execute operations on the processing units of the target hardware platform. The resources of the processing units in the target hardware platform are then allocated according to the defined runtime resource scheduling.

Core Innovation

The invention obtains information about a radio signal processing application to be executed using a radio communications modem and obtains hardware configuration information about the radio communications modem. It determines a computational dataflow graph corresponding to the radio signal processing application, where the computational dataflow graph comprises one or more tensor blocks representing one or more operations and directed edges representing data flow among the tensor blocks.

The invention partitions, using the hardware configuration information and one or more machine learning models, the tensor blocks and directed edges of the computational dataflow graph to produce a set of software kernels for execution by one or more processing units of the radio communications modem to achieve a specific optimization objective. It then determines runtime resource scheduling of the set of software kernels to execute operations on the one or more processing units, and allocates resources according to the runtime resource scheduling.

The invention is further described as allowing the radio signal processing application to be expressed in functional form as a primitive dataflow graph that can be flattened from functional tensor-block graphs, with operations including merging and splitting, and parallelization and pipelining across tensor dimensions. It also includes using machine learning models and performance-related information to drive iterative partitioning and to account for characteristics related to compute, memory access, and communication times.

Claims Coverage

The document includes three independent claims (a method, non-transitory computer readable media, and a system), each centered on obtaining application and modem hardware information, representing the application as a computational dataflow graph with tensor blocks and directed edges, partitioning the graph using machine learning models to create software kernels for a specific optimization objective, and then performing runtime resource scheduling and resource allocation for execution on processing units. Across the independent claims, the number of inventive features is at least four.

Dataflow graph representation of radio signal processing

Determining a computational dataflow graph corresponding to the radio signal processing application, wherein the computational dataflow graph comprises one or more tensor blocks that represent one or more operations of the radio signal processing application and directed edges representing data flow among the one or more tensor blocks.

Machine-learning partitioning into software kernels for optimization

Partitioning, using the hardware configuration information and one or more machine learning models, the tensor blocks and directed edges of the computational dataflow graph to produce a set of software kernels for execution by one or more processing units of the radio communications modem to achieve a specific optimization objective.

Runtime resource scheduling of software kernels

Determining runtime resource scheduling of the set of software kernels to execute operations on the one or more processing units of the radio communications modem.

Resource allocation according to runtime scheduling

Allocating resources of the one or more processing units of the radio communications modem according to the runtime resource scheduling.

Overall claim coverage centers on representing a radio signal processing application as a computational dataflow graph with tensor blocks and directed edges, partitioning that graph into software kernels using hardware configuration information and one or more machine learning models to achieve a specific optimization objective, then determining runtime resource scheduling and allocating processing-unit resources according to that scheduling.

Stated Advantages

Not explicitly described in patent.

Documented Applications

Not explicitly described in patent.

JOIN OUR MAILING LIST

Stay Connected with MTEC

Keep up with active and upcoming solicitations, MTEC news and other valuable information.