Arm Limited

Microsoft Technology Licensing, LLC

Jose Alberto Joao

Alejandro Rico Carro

Joshua Randall

Onur Mutlu

Ziqiang Huang

Christopher Neal Hinds

Bruce James Mathewson

Saurabh Pijuskumar Sinha

Alexander Cole Shulyak

Alasdair Grant

Supreet Jeloka

Dam Sunwoo

Jesse Garrett Beu

Richard William Earnshaw

Krishnendra Nathella

Jaekyu Lee

Tiago Rogerio Muck

A method for controlling hardware resource configuration for a processing system comprises obtaining performance monitoring data indicative of processing performance associated with workloads to be executed on the processing system, providing a trained machine learning model with input data depending on the performance monitoring data; and based on an inference made from the input data by the trained machine learning model, setting control information for configuring the processing system to control an amount of hardware resource allocated for use by at least one processor core. A corresponding method of training the model is provided. This is particularly useful for controlling inter-core borrowing of resource between processor cores in a multi-core processing system, where resource is borrowed between respective cores, e.g. cores on different layers of a 3D integrated circuit.

Hardware resource configuration for processing system

An apparatus and method are provided for managing prefetch transactions. The apparatus has an interconnect for providing communication paths between elements coupled to the interconnect. The elements coupled to the interconnect comprise at least a requester element to initiate transactions, and a plurality of completer elements each of which is arranged to respond to a transaction received by that completer element. Congestion tracking circuitry maintains, in association with the requester element, a congestion indication for each of a plurality of routes through the interconnect used to propagate transactions initiated by that requester element. Each route comprises one or more communication paths, and the route employed to propagate a given transaction is dependent on a target completer element for that transaction. Prefetch throttling circuitry then identifies, in response to an indication of a given prefetch transaction that the requester element wishes to initiate, the target completer element amongst the plurality of completer elements to which that given prefetch transaction would be issued. It then determines whether to issue the given prefetch transaction in dependence on the congestion indication for the route that has been determined.

Apparatus and method for managing prefetch transactions

A method and apparatus is disclosed for transferring data from a first processor core to a second processor core. The first processor core executes a stash instruction having a first operand associated with a data address of the data. A second processor core is determined to be a stash target for a stash message, based on the data address or a second operand. A stash message is sent to the second processor core, notifying the second processor core of the written data. Responsive to receiving the stash message, the second processor core can opt to store the data in its cache. The data may be included in the stash message or retrieved in response to a read request by the second processing core. The second processor core may be determined by prediction based, at least in part, on monitored data transactions.

Core-to-core cache stashing and target discovery

Non-volatile storage circuitry is provided as primary storage accessible to processing circuitry, e.g. as registers, a cache, scratchpad memory, TLB or on-chip RAM. Power control circuitry powers down a given region of the non-volatile storage circuitry when information stored in said given region is not being used. This provides opportunities for more frequent power savings than would be possible if primary storage was implemented using volatile storage.

Non-volatile storage circuitry accessible as primary storage for processing circuitry

A method and apparatus for application thread prioritization to mitigate the effects of operating system noise is disclosed. The method generally includes executing in parallel a plurality of application threads of a parallel application. An interrupt condition of an application thread of the plurality of application threads is detected. A priority of the interrupted application thread is changed relative to priorities of one or more other application threads of the plurality of application threads, and control is returned to the interrupted application thread after the interrupt condition. The interrupted application thread then resumes execution in accordance with the changed priority.

Method and apparatus for application thread prioritization

Aspects of the present disclosure relate to an apparatus comprising a requester master processing device having an associated private cache storage to store data for access by the requester master processing device. The requester master processing device is arranged to issue a request to modify data that is associated with a given memory address and stored in a private cache storage associated with a recipient master processing device. The private cache storage associated with the recipient master processing device is arranged to store data for access by the recipient master processing device. The apparatus further comprises the recipient master processing device having its private cache storage. One of the recipient master processing device and its associated private cache storage is arranged to perform the requested modification of the data while the data is stored in the cache storage associated with the recipient master processing device.

Apparatus and method of modification of stored data

An apparatus comprises processing circuitry for executing instructions of two or more threads of processing, hardware registers to store context data for the two or more threads concurrently, and commit circuitry to commit results of executed instructions of the threads, where for each thread the commit circuitry commits the instructions of that thread in program order. At least one defer buffer is provided to buffer at least one blocked instruction for which execution by the processing circuitry is complete but execution of an earlier instruction of the same thread in the program order is incomplete. This can help to resolve inter-thread blocking and hence improve performance.

Defer buffer

An apparatus has processing circuitry to execute instructions from multiple threads and hardware registers to store context data for the multiple threads concurrently. At a given time a certain number of software-scheduled threads may be scheduled for execution by software executed by the processing circuitry. Hardware thread scheduling circuitry is provided to select one or more active threads to be executed from among the software-scheduled threads. The hardware thread scheduling circuitry adjusts the number of active threads in dependence on at least one performance metric indicating performance of the threads.

Hardware thread scheduling

A frequency-based prediction of indirect jumps executing in a computing environment is provided. Illustratively, a computing environment comprises a prediction engine that processes data representative of indirect jumps performed by the exemplary computing environment according to a selected frequency-based prediction paradigm. Operatively, the exemplary prediction engine can keep track of targets, in a table, taken for each indirect jump and program context (e.g., branch history and/or path information) of an exemplary computing program. Further, the prediction engine can also store a frequency counter associated with each target in the exemplary table. Illustratively, the frequency counter can record the number of times a target was taken in the recent past executions of an observed one or more indirect jump. The prediction engine can supply the target address of an indirect jump based on the values of the frequency counters of each stored target address.

Target-frequency based indirect jump prediction for high-performance processors

Systems and methods are provided to detect instances where dynamic predication of indirect jumps (DIP) is considered to be ineffective utilizing data collected on the recent effectiveness of dynamic predication on recently executed indirect jump instructions. Illustratively, a computing environment comprises a DIP monitoring engine cooperating with a DIP monitoring table that aggregates and processes data representative of the effectiveness of DIP on recently executed jump instructions. Illustratively, the exemplary DIP monitoring engine collects and processes historical data on DIP instances, where, illustratively, a monitored instance can be categorized according to one or more selected classifications. A comparison can be performed for currently monitored indirect jump instructions using the collected historical data (and classifications) to determine whether DIP should be invoked by the computing environment or whether to invoke other indirect jump prediction paradigms.

Growing community of inventors

Austin, TX, United States of America

Jose Alberto Joao