International Business Machines Corporation

John David Irish

William John Starke

Lakshminarayana Baba Arimilli

Jeffrey A Stuecheli

Kerry Christopher Imming

Michael Steven Siegel

Andrew Henry Wottreng

Chad Balling McBride

Bernard Charles Drerup

Kenneth Michael Valk

Ibrahim Abdel-Rahman Ouda

Guy L Guthrie

Bartholomew Blaner

Tolga Ozguner

Jason Andrew Thompson

James A Steenburgh

David John Krolak

Paul Allen Ganfield

Jeffrey Douglas Brown

Kent Harold Haselhorst

Derek E Williams

John Michael Borkenhagen

Randal Craig Swanberg

Bradly George Frey

Sheldon Bernard Levenstein

Charles Francis Marino

Joseph Franklin Logan

Curtis Carl Wollbrink

Steven R Kunkel

Mark David Bellows

David Alan Norgaard

Donald Lee Freerksen

Lyle Edwin Grosbach

Daniel Mark Dreps

Charles Ray Johns

Russell Dean Hoover

Scott Douglas Clark

Glen Howard Handlogten

Charles Luther Johnson

Stephen P Morgan

Mark S Fredrickson

Yiftach Benjamini

Lonny Lambrecht

Gary Michael Lippert

Ryan Abel Heckendorf

Jack Chris Randolph

Peder James Paulson

Farnaz Mounes-Toussi

Gerald G Fagerness

Spencer K Millican

Michael D Varga

Frank Michael Pittelli

Ibrahim A Ouda

David W Sollender

Michael A Muston

An integrated circuit includes a first communication interface for communicatively coupling the integrated circuit with a coherent data processing system, a second communication interface for communicatively coupling the integrated circuit with an accelerator unit including an accelerator functional unit and an effective address-based accelerator cache for buffering copies of data from the system memory of the coherent data processing system, and a real address-based directory inclusive of contents of the accelerator cache. The real address-based directory assigns entries based on real addresses utilized to identify storage locations in the system memory. The integrated circuit includes request logic that, responsive to receipt on the first communication interface of a translation entry invalidation request, issues to the accelerator unit via the second communication interface an invalidation request that identifies an entry in the effective address-based accelerator cache to be invalidated utilizing a host tag identifying a storage location in the real address-based directory.

Translation invalidation in a translation cache serving an accelerator

An integrated circuit includes a first communication interface for communicatively coupling the integrated circuit with a coherent data processing system, a second communication interface for communicatively coupling the integrated circuit with an accelerator unit including an effective address-based accelerator cache for buffering copies of data from a system memory, and a real address-based directory inclusive of contents of the accelerator cache. The real address-based directory assigns entries based on real addresses utilized to identify storage locations in the system memory. The integrated circuit further includes request logic that communicates memory access requests and request responses with the accelerator unit. The request logic, responsive to receipt from the accelerator unit of a read-type request specifying an aliased second effective address of a target cache line, provides a request response including a host tag indicating that the accelerator unit has associated a different first effective address with the target cache line.

Integrated circuit and data processing system supporting address aliasing in an accelerator

An integrated circuit for a coherent data processing system includes a first communication interface for communicatively coupling the integrated circuit with the coherent data processing system, a second communication interface for communicatively coupling the integrated circuit with an accelerator unit including an effective address-based accelerator cache for buffering copies of data from a system memory of the coherent data processing system, and a real address-based directory inclusive of contents of the accelerator cache. The real address-based directory assigns entries based on real addresses utilized to identify storage locations in the system memory. The integrated circuit further includes request logic that communicates memory access requests and request responses with the accelerator unit via the second communication interface. A request response identifies a target of a corresponding memory access request utilizing a host tag specifying an entry associated with the target in the real address-based directory.

Integrated circuit and data processing system supporting attachment of a real address-agnostic accelerator

An integrated circuit includes a first communication interface for communicatively coupling the integrated circuit with a coherent data processing system, a second communication interface for communicatively coupling the integrated circuit with an accelerator unit including an effective address-based accelerator cache for buffering copies of data from a system memory, and a real address-based directory inclusive of contents of the accelerator cache. The real address-based directory assigns entries based on real addresses utilized to identify storage locations in the system memory. The integrated circuit further includes directory control logic that configures at least a number of congruence classes utilized in the real address-based directory based on configuration parameters specified on behalf of or by the accelerator unit.

Integrated circuit and data processing system having a configurable cache directory for an accelerator

A claw-back request, received from an accelerator, is issued for an address line. While waiting for a response to the claw-back request, a cast-out push request with a matching address line is received. The cast-out push request is associated with a cache having a modified copy of the address line. A combined-response, associated with the cast-out push request, is received from a bus. Data associated with the modified copy of the address line is received from the cache. A claw-back response, with the data associated with the modified version of the address line, is issued to an accelerator.

Accelerator memory coherency with single state machine

Techniques are disclosed for faster loading of data for hardware accelerators. One technique includes after determining that an accelerator is not ready to perform a workload, identifying data associated with performing the workload and negotiating for the data on behalf of the accelerator. After the negotiation, a cache directory entry associated with the data is marked with a 'claimed' state indicating that the accelerator has obtained ownership of the data but does not have possession of the data. After an indication that the accelerator is ready to accept the data for the workload is received, the data is moved from a previous owner that has possession of the data to the accelerator. Another technique includes requesting a processing unit to perform a workload. If the processing unit is not ready to perform the workload, a translation cache used by the processing unit is warmed up by another unit.

Techniques for faster loading of data for accelerators

A serial communication system includes a transmitting circuit for serially transmitting data via a serial communication link including N channels where N is an integer greater than 1. The transmitting circuit includes an input buffer having storage for input data frames each including M bytes forming N segments of M/N contiguous bytes. The transmitting circuit additionally includes a reordering circuit coupled to the input buffer. The reordering circuit includes a reorder buffer including multiple entries. The reordering circuit buffers, in each of multiple entries of the reorder buffer, a byte in a common byte position in each of the N segments of an input data frame. The reordering circuit sequentially outputs the contents of the entries of the reorder buffer via the N channels of the serial communication link.

Pre-transmission data reordering for a serial interface

In at least some embodiments, a processor core executes a sending thread including a first push instruction and a second push instruction subsequent to the first push instruction in a program order. Each of the first and second push instructions requests that a respective message payload be pushed to a mailbox of a receiving thread. In response to executing the first and second push instructions, the processor core transmits respective first and second co-processor requests to a switch in the data processing system via an interconnect fabric of the data processing system. The processor core transmits the second co-processor request to the switch without regard to acceptance of the first co-processor request by the switch.

Non-serialized push instruction for pushing a message payload from a sending thread to a receiving thread

In a data processing system, a switch includes a receive data structure including receive entries each uniquely corresponding to a receive window, where each receive entry includes addressing information for one or more mailboxes into which messages can be injected, a send data structure including send entries each uniquely corresponding to a send window, where each send entry includes a receive window field that identifies one or more receive windows, and switch logic. The switch logic, responsive to a request to push a message to one or more receiving threads, accesses a send entry that corresponds to a send window of the sending thread, utilizes contents of the receive window field of the send entry to access one or more of the receive entries, and pushes the message to one or more mailboxes of one or more receiving threads utilizing the addressing information of the receive entry or entries.

Addressing for inter-thread push communication

A processor core of a data processing system receives a push instruction of a sending thread that requests that a message payload identified by at least one operand of the push instruction be pushed to a mailbox of a receiving thread. In response to receiving the push instruction, the processor core executes the push instruction of the sending thread. In response to executing the push instruction, the processor core initiates transmission of the message payload to the mailbox of the receiving thread. In one embodiment, the processor core initiates transmission of the message payload by transmitting a co-processor request to a switch of the data processing system via an interconnect fabric.

Push instruction for pushing a message payload from a sending thread to a receiving thread

In a data processing system, a switch of the data processing system receives a request to push a message referenced by an instruction of a sending thread to a receiving thread. In response to receiving the request, the switch determines whether the sending thread is authorized to push the message to the receiving thread by attempting to access an entry of a data structure of the switch utilizing a key derived from at least one identifier of the sending thread. In response to access to the entry being successful, content of the entry is utilized to determine an address of a mailbox of the receiving thread, and the switch pushes the message to the mailbox of the receiving thread. In response to access to the entry not being successful, the switch refrains from pushing the message to the mailbox of the receiving thread.

Hardware-assisted interthread push communication

Growing community of inventors

Rochester, MN, United States of America

John David Irish