Graphcore Limited

Graphcore Limtied

Richard Luke Southwell Osborne

Stephen Felix

Daniel John Pelham Wilkinson

Alan Graham Alexander

Simon Christian Knowles

David Lacey

Matthew David Fyles

Jonathan Mangnall

Graham Bernard Cunningham

Brian Manula

Ola Tørudbakken

Harald Høeg

Ian James Quinn

Lars Paul Huse

Edward Andrews

Hachem Yassine

Sam Chesney

Ola TØRUNDBAKKEN

Harald HØEG

The invention relates to a computer program comprising a sequence of instructions for execution on a processing unit having instruction storage for holding the computer program, an execution unit for executing the computer program and data storage for holding data, the computer program comprising one or more computer executable instruction which, when executed, implements: a send function which causes a data packet destined for a recipient processing unit to be transmitted on a set of connection wires connected to the processing unit, the data packet having no destination identifier but being transmitted at a predetermined transmit time; and a switch control function which causes the processing unit to control switching circuitry to connect a set of connection wires of the processing unit to a switching fabric to receive a data packet at a predetermined receive time.

Instruction set

For certain applications, parts of the application data held in memory of a processing device (e.g. that are produced as a result of operations performed by the execution unit) are arranged in regular repeating patterns in the memory, and therefore, the execution unit may set up a suitable striding pattern for use by a send engine. The send engine accesses the memory at locations in accordance with the configured striding pattern so as to access a plurality of items of data that are arranged together in a regular pattern. In a similar manner as done for sends, the execution may set up a striding pattern for use by a receive engine. The receive engine, upon receiving a plurality of items of data, causes those items of data to be stored at locations in the memory, as determined in accordance with the configured striding pattern.

Processing device using variable stride pattern

A method for debugging a processor which is executing vertices of a software application is described. Each vertex is assigned to a programming thread of the processor. The processor has debug hardware for raising exceptions in certain break conditions. The method comprises inspecting a vertex identifier, comparing the vertex identifier and raising an instruction exception event for the programming thread if the vertex identifier assigned to the thread matches the vertex break identifier in the debug hardware. Exceptions are raised based on identified vertices, rather than just individual instructions or instruction addresses.

Method of debugging a processor that executes vertices of an application, each vertex being assigned to a programming thread of the processor

Two or more die are stacked together in a stacked integrated circuit device. Each of the processors on these die is able to communicate with other processors on its die by sending data over the switching fabric of its respective die. The mechanism for sending data between processors on the same die (i.e. intradie communication) is reused for sending data between processors on different die (i.e. interdie communication). The reuse of the mechanism is enabled by assigning each processor a vertical neighbour on its opposing die. Each processor has an interdie connection that connects it to the output exchange bus of its neighbour. A processor is able to borrow the output exchange bus of its neighbour by sending data along the output exchange bus of its neighbour.

Exchange between stacked die

In a computer comprising multiple processing units, a method of exchanging read only elements between the processing units is described. The read only elements may be code or data, such as vector or matrix data for an AI graph. A master processing unit is identified. At compile time, at least one shareable read only element is allocated to the master processing unit. The at least one shareable read only element is stored in the local memory of the master processing unit. At compile time a transmitting exchange code sequence designated to be executed at the execution stage of the master processing unit is also allocated to the master processing unit. At a time point determined at compile time, the transmitting exchange code sequence causes the processing unit to identify the shareable read only element and to generate a message to be transmitted for reception by another processing unit, the message comprising the shareable read only data element.

Sharing data structures

A processor in a network has a plurality of processing units arranged on a chip. An on-chip interconnect enables data to be exchanged between the processing units. A plurality of external interfaces are configured to communicate data off chip in the form of packets, each packet having a destination address identifying a destination of the packet. The external interfaces are connected to respective additional connected processors. A routing bus routes packets between the processing units and the external interfaces. A routing register defines a routing domain for the processor, the routing domain comprising one or more of the additional processor, and at least a subset of further additional processors of the network, wherein the additional processors of the subset are directly or indirectly connected to the processor. The routing domain can be modified by changing the contents of the routing register as a sliding window domain.

Routing in a network of processors

A computer program comprising a sequence of instructions for execution on a processing unit having instruction storage for holding the computer program, an execution unit for executing the computer program and data storage for holding data, the computer program comprising: a switch control instruction which when executed causes the processing unit to control switching circuitry to connect a set of connection wires of the processing unit to a switching fabric to receive a data packet at a predetermined received time, the switch control instruction comprising a delay control field which holds a value defining a delay between issuance of the instruction in the sequence of instructions and its execution by the execution unit.

Controlling timing in computer processing

The invention relates to a computer implemented method of generating multiple programs to deliver a computerised function, each program to be executed in a processing unit of a computer comprising a plurality of processing units each having instruction storage for holding a local program, an execution unit for executing the local program and data storage for holding data, a switching fabric connected to an output interface of each processing unit and connectable to an input interface of each processing unit by switching circuitry controllable by each processing unit, and a synchronisation module operable to generate a synchronisation signal, the method comprising: generating a local program for each processing unit comprising a sequence of executable instructions; determining for each processing unit a relative time of execution of instructions of each local program whereby a local program allocated to one processing unit is scheduled to execute with a predetermined delay relative to a synchronisation signal a send instruction to transmit at least one data packet at a predetermined transmit time, relative to the synchronisation signal, destined for a recipient processing unit but having no destination identifier, and a local program allocated to the recipient processing unit is scheduled to execute at a predetermined switch time a switch control instruction to control the switching circuitry to connect its processing unit wire to the switching fabric to receive the data packet at a receive time.

Compiler method

A gateway for interfacing a host with a subsystem for acting as a work accelerator to the host. The gateway enables the transfer of batches of data to the subsystem at precompiled data exchange synchronisation points. The gateway comprises a streaming engine having a data mover engine and a memory management engine, the data mover engine and memory management engine being configured to execute instructions in coordination from work descriptors. The memory management engine is configured to execute instructions from the work descriptor to transfer data between external storage and the local memory associated with the gateway. The data mover engine is configured to execute instructions from the work descriptor to transfer data between the local memory associated with the gateway and the subsystem.

Streaming engine

The present disclosure relates to a method of scheduling messages to be exchanged between tiles in a computer where there is a fixed transmission time between sending and receiving tiles. According to the method a total size of message data to be sent or received by each tile is determined. One of the tiles is selected based at least on the size of the message data to schedule a first message. The first message to be scheduled is selected from the set of messages on that tile. In order to schedule the message the other end points of this selected message are determined, and then respective time slots are allocated at the sending and receiving tiles for that message. The size of the selected message is then deducted from each of the tiles acting as end points for the message, and then the sequence is carried out again until all messages have been scheduled. This technique optimises message exchange in an exchange phase of a BSP system.

Scheduling messages

A processing system comprising: a subsystem for acting as a work accelerator to a host processor, the subsystem comprising an arrangement of tiles; and an interconnect for communicating between the tiles and connecting the subsystem to the host. The interconnect comprises synchronization logic to coordinate barrier synchronizations between a group of the tiles. The synchronization logic comprises a host sync proxy module, comprising a counter written with a number of credits by the host processor, and being configured to automatically decrement the number of credits each time one of the barrier synchronizations requiring host involvement is performed. When the number of credits in the counter is exhausted, the barrier is not released until a further write from the host to the host sync proxy module, but when the number is credits in the counter is not exhausted the barrier is released without a separate write from the host.

Synchronization with a host processor

A method of operating a system comprising multiple processor tiles divided into a plurality of domains wherein within each domain the tiles are connected to one another via a respective instance of a time-deterministic interconnect and between domains the tiles are connected to one another via a non-time-deterministic interconnect. The method comprises: performing a compute stage, then performing a respective internal barrier synchronization within each domain, then performing an internal exchange phase within each domain, then performing an external barrier synchronization to synchronize between different domains, then performing an external exchange phase between the domains.

Growing community of inventors

Bristol, United Kingdom

Richard Luke Southwell Osborne