Qualcomm Incorporated

Google Inc.

Fitbit, Inc.

Aravind Natarajan

Conor Joseph Heneghan

Han Zhao

Lindsey Michelle Sunden

Arun Raman

Dario Suarez Gracia

Gheorghe Calin Cascaval

Tushar Kumar

Belen Lafon

Xi Zhang

Pablo Montesinos Ortego

Hao-Wei Su

Aniket Sanjay Deshpande

Behnam Robatmili

Shaizeen Aga

Ravish Hastantram

Embodiments include computing devices, apparatus, and methods implemented by the apparatus for implementing speculative loop iteration partitioning (SLIP) for heterogeneous processing devices. A computing device may receive iteration information for a first partition of iterations of a repetitive process and select a SLIP heuristic based on available SLIP information and iteration information for the first partition. The computing device may determine a split value for the first partition using the SLIP heuristic, and partition the first partition using the split value to produce a plurality of next partitions.

Speculative loop iteration partitioning for heterogeneous execution

Embodiments include computing devices, apparatus, and methods implemented by the apparatus for implementing data flow management on a computing device. Embodiment methods may include initializing a buffer partition of a first memory of a first heterogeneous processing device for an output of execution of a first iteration of a first operation by the first heterogeneous processing device on which a first iteration of a second operation assigned for execution by a second heterogeneous processing device depends. Embodiment methods may include identifying a memory management operation for transmitting the output by the first heterogeneous processing device from the buffer partition as an input to the second heterogeneous processing device. Embodiment methods may include allocating a second memory for storing data for an iteration executed by a third heterogeneous processing device to minimize a number of memory management operations for the second allocated memory.

Managing data flow in heterogeneous computing

Embodiments include computing devices, systems, and methods identifying enhanced synchronization operation outcomes. A computing device may receive a first resource access request for a first resource of a computing device including a first requester identifier from a first computing element of the computing device. The computing device may also receive a second resource access request for the first resource including a second requester identifier from a second computing element of the computing device. The computing device may grant the first computing element access to the first resource based on the first resource access request, and return a response to the second computing element including the first requester identifier as a winner computing element identifier.

Identifying enhanced synchronization operation outcomes to improve runtime operations

Methods, devices, and non-transitory processor-readable storage media for a computing device to merge concurrent writes from a plurality of processing units to a buffer associated with an application. An embodiment method executed by a processor may include identifying a plurality of concurrent requests to access the buffer that are sparse, disjoint, and write-only, configuring a write-set for each of the plurality of processing units, executing the plurality of concurrent requests to access the buffer using the write-sets, determining whether each of the plurality of concurrent requests to access the buffer is complete, obtaining a buffer index and data via the write-set of each of the plurality of processing units, and writing to the buffer using the received buffer index and data via the write-set of each of the plurality of processing units in response to determining that each of the plurality of concurrent requests to access the buffer is complete.

Random-access disjoint concurrent sparse writes to heterogeneous buffers

Various embodiments include methods for data management in a computing device utilizing a plurality of processing units. Embodiment methods may include generating a data transfer heuristic model based on measurements from a plurality of sample data transfers between a plurality of data storage units. The generated data transfer heuristic model may be used to calculate data transfer costs for each of a plurality of tasks. The calculated data transfer costs may be used to schedule execution of the plurality of tasks in an execution order on selected ones of the plurality of processing units. The data transfer heuristic model may be updated based on measurements of data transfers occurring during the executions of the plurality of tasks (e.g., time, power consumption, etc.). Code executing on the processing units may indicate to a runtime when certain data blocks are no longer needed and thus may be evicted and/or pre-fetched for others.

Data management for multiple processing units using data transfer costs

Embodiments include computing devices, systems, and methods for task-based handling of repetitive processes in parallel. At least one processor of the computing device, or a specialized hardware controller, may be configured to partition iterations of a repetitive process and assign the partitions to initialized tasks to be executed in parallel by a plurality of processor cores. Upon completing a task, remaining divisible partitions of the repetitive process of ongoing tasks may be subpartitioned and assigned to the ongoing task, and the completed task or a newly initialized task. Information about the iteration space for a repetitive process may be stored in a descriptor table, and status information for all partitions of a repetitive process stored in a status table. Each processor core may have an associated local table that tracks iteration execution of each task, and is synchronized with the status table.

Method for exploiting parallelism in task-based systems using an iteration space splitter

Growing community of inventors

Sunnyvale, CA, United States of America

Aravind Natarajan