Apple Inc.

Arm Limited

Texas Instruments Corporation

Gerard Richard Williams

Jeffry E Gonion

Stephan G Meier

John H Mylius

James B Keller

Brian P Lilly

Pradeep Kanapathipillai

Conrado Blasco-Allue

Ali Sazegari

Wei-Han Lien

David James Williamson

Inder M Sodhi

Per Hakan Hammarlund

Joseph T DiBene, Ii

Hari Kannan

Harshavardhan Kaushikkar

Keith Alan Cox

Steven John Fishwick

Sandeep Kumar Gupta

Lior Zimet

Srinivasa Rangan Sridharan

Gaurav Garg

Eric Bainville

Lital Levy-Rubin

Shyam Sundar

Sagi Lahav

Suparn Vats

Charles E Tucker

Mahnaz P Sadoughi-Yarandi

David Walter Flynn

Ian D Kountanis

Chris Cheng-Chieh Lee

James Nolan Hardage

Richard F Russo

James Vash

Tyler J Huberty

Prashant Jain

Richard Slobodnik

Jeff Gonion

Perumal R Subramoniam

Eran Tamari

Samer Nassar

Perumal R Subramonium

Mark Allen Silla

Tan Ba Tran

Richard Roy Grisenthwaite

Manu Gulati

Sukalpa Biswas

Joshua P De Cesare

Ramesh B Gunna

Timothy R Paaske

Gurjeet S Saund

Michael J Smith

Josh P De Cesare

R Stephen Polzin

Shinye Shiu

Sergio Kolor

Bernard Joseph Semeria

Mridul Agarwal

Stephen John Hill

Andrew J Beaumont-Smith

Conrado Blasco

Tal Kuzi

Shih-Chieh Wen

Rohit Kumar

Kulin N Kothari

Jacques Anthony Vidrine

David Terrence Matheny

Teresa Louise McLaurin

Barry Duane Williamson

Tristan F Schaap

Muralidharan Santharaman Chinnakonda

Pavlos Konas

Peter G Soderquist

Gregory Daniel Hughes

Suresh B Periyacheri

Dany Davidov

Nir Leshem

Shyam Sundar Balasubramanian

Fang Liu

Mark Pilip

Arthur R Piejko

Michael L Brauer

Kim Rasmussen

Conrade Blasco-Allue

In an embodiment, a system may support programmable hashing of address bits at a plurality of levels of granularity to map memory addresses to memory controllers and ultimately at least to memory devices. The hashing may be programmed to distribute pages of memory across the memory controllers, and consecutive blocks of the page may be mapped to physically distant memory controllers. In an embodiment, address bits may be dropped from each level of granularity, forming a compacted pipe address to save power within the memory controller. In an embodiment, a memory folding scheme may be employed to reduce the number of active memory devices and/or memory controllers in the system when the full complement of memory is not needed.

Hashing with soft memory folding

Techniques are disclosed relating to an I/O agent circuit. The I/O agent circuit may include one or more queues and a transaction pipeline. The I/O agent circuit may issue, to the transaction pipeline from a queue of the one or more queues, a transaction of a series of transactions enqueued in a particular order. The I/O agent circuit may generate, at the transaction pipeline, a determination to return the transaction to the queue based on a detection of one or more conditions being satisfied. Based on the determination, the I/O agent circuit may reject, at the transaction pipeline, up to a threshold number of transactions that issued from the queue after the transaction issued. The I/O agent circuit may insert the transaction at a head of the queue such that the transaction is enqueued at the queue sequentially first for the series of transactions according to the particular order.

Ensuring transactional ordering in I/O agent

In an embodiment, a system includes multiple power management mechanism operating in different time domains (e.g., with different bandwidths) and control circuitry that is configured to coordinate operation of the mechanisms. If one mechanism is adding energy to the system, for example, the control circuitry may inform another mechanism that the energy is coming so that the other mechanism may not take as drastic an action as it would if no energy were coming. If a light workload is detected by circuitry near the load, and there is plenty of energy in the system, the control circuitry may cause the power management unit (PMU) to generate less energy or even temporarily turn off. A variety of mechanisms for the coordinated, coherent use of power are described.

Systems and methods for coherent power management

An integrated circuit (IC) including a plurality of processor cores, a plurality of graphics processing units, a plurality of peripheral circuits, and a plurality of memory controllers is configured to support scaling of the system using a unified memory architecture. For example, the IC may include an interconnect fabric configured to provide communication between the one or more memory controller circuits and the processor cores, graphics processing units, and peripheral devices; and an off-chip interconnect coupled to the interconnect fabric and configured to couple the interconnect fabric to a corresponding interconnect fabric on another instance of the integrated circuit, wherein the interconnect fabric and the off-chip interconnect provide an interface that transparently connects the one or more memory controller circuits, the processor cores, graphics processing units, and peripheral devices in either a single instance of the integrated circuit or two or more instances of the integrated circuit.

Scalable system on a chip

Techniques are disclosed relating to an I/O agent circuit of a computer system. The I/O agent circuit may receive, from a peripheral component, a set of transaction requests to perform a set of read transactions that are directed to one or more of a plurality of cache lines. The I/O agent circuit may issue, to a first memory controller circuit configured to manage access to a first one of the plurality of cache lines, a request for exclusive read ownership of the first cache line such that data of the first cache line is not cached outside of the memory and the I/O agent circuit in a valid state. The I/O agent circuit may receive exclusive read ownership of the first cache line, including receiving the data of the first cache line. The I/O agent circuit may then perform the set of read transactions with respect to the data.

I/O agent

In an embodiment, a system includes multiple power management mechanism operating in different time domains (e.g. with different bandwidths) and control circuitry that is configured to coordinate operation of the mechanisms. If one mechanism is adding energy to the system, for example, the control circuitry may inform another mechanism that the energy is coming so that the other mechanism may not take as drastic an action as it would if no energy were coming. If a light workload is detected by circuitry near the load, and there is plenty of energy in the system, the control circuitry may cause the power management unit (PMU) to generate less energy or even temporarily turn off. A variety of mechanisms for the coordinated, coherent use of power are described.

In an embodiment, a computation engine is configured to perform vector multiplications, producing either vector results or outer product (matrix) results. The instructions provided to the computation engine specify a matrix mode or a vector mode for the instructions. The computation engine performs the specified operation. The computation engine may perform numerous computations in parallel, in an embodiment. In an embodiment, the instructions may also specify an offset with the input memories, providing additional flexibility in the location of operands. More particularly, the computation engine may be configured to perform numerous multiplication operations in parallel and to accumulate results in a result memory, performing multiply-accumulate operations for each matrix/vector element in the targeted locations of the output memory.

Computation engine that operates in matrix and vector modes

Systems, apparatuses, methods, and computer-readable mediums for preventing return oriented programming (ROP) attacks. A compiler may insert landing pads adjacent to valid return targets in an instruction sequence. When a return instruction is executed, the processor may treat the return as suspicious if the target of the return instruction does not have an adjacent landing pad. Additionally, each landing pad may be encoded with a color, and a colored launch pad may be inserted into the instruction stream next to each return instruction. When a return instruction is executed, the processor may determine if the target of the return has a landing pad with the same color as the launch pad of the return instruction. Return-target pairs with color mismatches may be treated as suspicious and the offending process may be killed.

Marking valid return targets

In an embodiment, a system may support a 'coast mode' in which the power management unit (PMU) that supplies the supply voltage to an integrated circuit is disabled temporarily for certain modes of the integrated circuit. The integrated circuit may continue to operate, consuming the energy stored in capacitance in and/or around the integrated circuit. When coast mode is initiated, a time interval for coasting may be determined. When the time interval expires, the PMU may re-enable the power supply voltage.

Reduced power operation using stored capacitor energy

In an embodiment, a computation engine may offload work from a processor (e.g. a CPU) and efficiently perform computations such as those used in LSTM and other workloads at high performance. In an embodiment, the computation engine may perform computations on input vectors from input memories in the computation engine, and may accumulate results in an output memory within the computation engine. The input memories may be loaded with initial vector data from memory, incurring the memory latency that may be associated with reading the operands. Compute instructions may be performed on the operands, generating results in an output memory. One or more extract instructions may be supported to move data from the output memory to the input memory, permitting additional computation on the data in the output memory without moving the results to main memory.

Computation engine with extract instructions to minimize memory access

Systems, apparatuses, and methods for efficiently moving data for storage and processing a compression unit within a processor includes multiple hardware lanes, selects two or more input words to compress, and for assigns them to two or more of the multiple hardware lanes. As each assigned input word is processed, each word is compared to an entry of a plurality of entries of a table. If it is determined that each of the assigned input words indexes the same entry of the table, the hardware lane with the oldest input word generates a single read request for the table entry and the hardware lane with the youngest input word generates a single write request for updating the table entry upon completing compression. Each hardware lane generates a compressed packet based on its assigned input word.

Systems and methods for performing memory compression

In an embodiment, a processor may implement an access map-pattern match (AMPM)-based prefetch circuit for a multi-level cache system. The access patterns that are matched to the access maps may include prefetches for different cache levels. Centralizing the generation of prefetches into one prefetch circuit may provide better observability and controllability of prefetching at various levels of the cache hierarchy, in an embodiment. Prefetches at different levels may be controlled individually based on the accuracy of those prefetches, in an embodiment. Additionally, in an embodiment, access patterns that are longer that a given threshold may have the granularity of the prefetches change so that more data is prefetched and the prefetches occur farther in advance, in some embodiments.

Unified prefetch circuit for multi-level caches

Systems, apparatuses, and methods for optimizing a load-store dependency predictor (LSDP). When a younger load instruction is issued before an older store instruction and the younger load is dependent on the older store, the LSDP is trained on this ordering violation. A replay/flush indicator is stored in a corresponding entry in the LSDP to indicate whether the ordering violation resulted in a flush or replay. On subsequent executions, a dependency may be enforced for the load-store pair if a confidence counter is above a threshold, with the threshold varying based on the status of the replay/flush indicator. If a given load matches on multiple entries in the LSDP, and if at least one of the entries has a flush indicator, then the given load may be marked as a multimatch case and forced to wait to issue until all older stores have issued.

Load/store dependency predictor optimization for replayed loads

Growing community of inventors

Los Altos, CA, United States of America

Gerard Richard Williams