Apple Inc.

Other

International Business Machines Corporation

Robert D Kenney

Terence M Potter

Liang-Kai Wang

Benjiman L Goodman

Christopher A Burns

Andrew M Havlir

Sivayya V Ayinala

Gregory D Roberts

James Han Wang

Aaftab A Munshi

Michael R Seningen

Richard Warren Schreyer

Paul K Miller

Justin A Hensley

Jason Nathaniel Dale

Raymond Cheung Yeung

Ronald Gene Walther

Dzung Q Vu

Sreevathsa Ramachandra

Donald W Glowka

Michael A Geary

Jeffrey B Reed

Liang Xia

Vinod Reddy Nalamalapu

James De Leon

Hani Hasan Mustafa Saleh

Dan Jeffrey Dickinson

Christina Lynne Newman-LaBounty

Techniques are disclosed relating to operand routing among SIMD pipelines. In some embodiments, an apparatus includes a set of multiple hardware pipelines configured to execute a single-instruction multiple-data (SIMD) instruction for multiple threads in parallel, wherein the instruction specifies first and second architectural registers. In some embodiments, the pipelines include execution circuitry configured to perform operations using one or more pipeline stages of the pipeline. In some embodiments, the pipelines include routing circuitry configured to select, based on the instruction, a first input operand for the execution circuitry from among: a value from the first architectural register from thread-specific storage for another pipeline and a value from the second architectural register from thread-specific storage for a thread assigned to another pipeline. In some embodiments, the routing circuitry may support a shift and fill instruction that facilitates storage of an arbitrary portion of a graphics frame in one or more registers.

SIMD operand permutation with selection from among multiple registers

Techniques are disclosed relating to channel stalls or deactivations based on the latency of prior operations. In some embodiments, a processor includes a plurality of channel pipelines for a plurality of channels and a plurality of execution pipelines shared by the channel pipelines and configured to perform different types of operations provided by the channel pipelines. First scheduler circuitry may assign threads to channels and second scheduler circuitry may assign an operation from a given channel to a given execution pipeline based on decode of an operation for that channel. Dependency circuitry may, for a first operation that depends on a prior operation that uses one of the execution pipelines, determine, based on status information for the prior operation from the one of the execution pipelines, whether to stall the first operation or to deactivate a thread that includes the first operation from its assigned channel.

Fence enforcement techniques based on stall characteristics

Techniques are disclosed relating to sharing datapath circuitry among multiple SIMD groups. In some embodiments, pipeline circuitry is configured to perform operations specified by instructions of first and second assigned SIMD groups. The pipeline circuitry may include first and second front-end circuitry configured to decode instructions of the respective SIMD groups. The pipeline circuitry may include shared execution circuitry configured to perform operations specified by the first and second assigned SIMD groups and arbitration circuitry configured to select an instruction from among at least the first and second front-end circuitry for assignment to the shared execution circuitry in a current cycle. The arbitration circuitry may select an instruction based on one or more of: stall counts, whether available instructions are being speculatively executed, whether ones of available instructions target a particular portion of the shared execution circuitry, numbers of execution cycles, and SIMD group ages.

Multi-channel data path circuitry

Techniques are disclosed relating to routing circuitry configured to perform permute operations for operands of threads in a single-instruction multiple-data group. In some embodiments, an apparatus includes hierarchical operand routing circuitry configured to route operands between a set of single-instruction multiple-data (SIMD) pipelines based on a permute instruction. In some embodiments, the routing circuitry includes a first level and a second level. The first level may include a set of multiple crossbar circuits each configured to receive operands from a respective subset of the pipelines and output one or more of the received operands on multiple output lines based on the permute instruction, where the crossbar circuits support full permutation within a respective subset. A second level may be configured to select an operand from a previous level for each of the pipelines, and may select from among only a portion of output operands from the previous level to provide an operand for a respective pipeline.

Routing circuitry for permutation of single-instruction multiple-data operands

Techniques are disclosed relating to sharing operands among SIMD threads for a larger arithmetic operation. In some embodiments, a set of multiple hardware pipelines is configured to execute single-instruction multiple-data (SIMD) instructions for multiple threads in parallel, where ones of the hardware pipelines include execution circuitry configured to perform floating-point operations using one or more pipeline stages of the pipeline and first routing circuitry configured to select, from among thread-specific operands stored for the hardware pipeline and from one or more other pipelines in the set, a first input operand for an operation by the execution circuitry. In some embodiments, a device is configured to perform a mathematical operation on source input data structures stored across thread-specific storage for the set of hardware pipelines, by executing multiple SIMD floating-point operations using the execution circuitry and the first routing circuitry. This may improve performance and reduce power consumption for matrix multiply and reduction operations, for example.

Datapath circuitry for math operations using SIMD pipelines

Techniques are disclosed relating to arbitration among register file accesses. In some embodiments, an apparatus includes a register file configured to store operands for multiple client circuits and arbitration circuitry configured to select from among multiple received requests to access the register file. In some embodiments, the apparatus includes first interface circuitry configured to provide access requests from a first client circuit to the arbitration circuitry and supplemental interface circuitry configured to receive unsuccessful requests from the first client circuit and provide the received unsuccessful requests to the arbitration circuitry. The supplemental interface circuitry may provide additional catch-up bandwidth to clients that lose arbitration, which may result in fairness during bandwidth shortages.

Register file arbitration

Techniques are disclosed relating to scheduling tasks for graphics processing. In one embodiment, a graphics unit is configured to render a frame of graphics data using a plurality of pass groups and the frame of graphics data includes a plurality of frame portions. In this embodiment, the graphics unit includes scheduling circuitry configured to receive a plurality of tasks, maintain pass group information for each of the plurality of tasks, and maintain relative age information for the plurality of frame portions. In this embodiment, the scheduling circuitry is configured to select a task for execution based on the pass group information and the age information. In some embodiments, the scheduling circuitry is configured to select tasks from an oldest frame portion and current pass group before selecting other tasks. This scheduling approach may result in efficient execution of various different types of graphics workloads.

GPU task scheduling

A data queuing and format apparatus is disclosed. A first selection circuit may be configured to selectively couple a first subset of data to a first plurality of data lines dependent upon control information, and a second selection circuit may be configured to selectively couple a second subset of data to a second plurality of data lines dependent upon the control information. A storage array may include multiple storage units, and each storage unit may be configured to receive data from one or more data lines of either the first or second plurality of data lines dependent upon the control information.

Data alignment and formatting for graphics processing unit

Techniques are disclosed relating to sharing an arithmetic logic unit (ALU) between multiple threads. In some embodiments, the threads also have dedicated ALUs for other types of operations. In some embodiments, arbitration circuitry is configured to receive operations to be performed by the shared arithmetic logic unit from the set of threads and issue the received operations to the shared arithmetic logic unit. In some embodiments, the arbitration circuitry is configured to switch to a different one of the set of threads for each instruction issued to the shared arithmetic logic unit. In some embodiments, the shared ALU is configured to perform 32-bit operations and the dedicated ALUs are configured to perform the same operations using 16-bit precision. In some embodiments, the shared ALU is shared between two threads and is physically located adjacent to other datapath circuitry for the two threads.

Techniques for ALU sharing between threads

Techniques are disclosed relating to controlling an operand cache in a pipelined fashion. An operand cache may cache operands fetched from the register file or generated by previous instructions to improve performance and/or reduce power consumption. In some embodiments, instructions are pipelined and separate tag information is maintained to indicate allocation of an operand cache entry and ownership of the operand cache entry. In some embodiments, this may allow an operand to remain in the operand cache (and potentially be retrieved or modified) during an interval between allocation of the entry for another operand and ownership of the entry by the other operand. This may improve operand cache efficiency by allowing the entry to be used while to retrieving the other operand from the register file, for example.

Pipelined allocation for operand cache

Techniques are disclosed relating to selecting store instructions for dispatch to a shared pipeline. In some embodiments, the shared pipeline processes instructions for different target clients with different data rate capabilities. Therefore, in some embodiments, the pipeline is configured to generate state information that is based on a determined amount of work in the pipeline that targets at least one slower target. In some embodiments, the state information indicates whether the amount of work is above a threshold for the particular target. In some embodiments, scheduling circuitry is configured to select instructions for dispatch to the pipeline based on the state information. For example, the scheduling circuitry may refrain from selecting instructions with a slower target when the slower target is above its threshold amount of work in the pipeline. In some embodiments, the shared pipeline is a store pipeline configured to execute store instructions that target memories with different data rate capabilities.

Hints for shared store pipeline and multi-rate targets

Techniques are disclosed relating to a hardware-supported flexible data structure for graphics processing. In some embodiments, dimensions of the data structure are configurable in an X direction, a Y direction, a number of samples per pixel, and an amount of data per sample. In some embodiments, these attributes are configurable using hardware registers. In some embodiments, the data structure is persistent across a tile being processed such that local memory context is accessible to both rendering threads of a render pass and mid-render compute threads.

Local image blocks for graphics processing

A method and apparatus for saving power in integrated circuits is disclosed. An IC includes functional circuit blocks which are not placed into a sleep mode when idle. A power management circuit may monitor the activity levels of the functional circuit blocks not placed into a sleep mode. When the power management circuit detects that an activity level of one of the non-sleep functional circuit blocks is less than a predefined threshold, it reduce the frequency of a clock signal provided thereto by scheduling only one pulse of a clock signal for every N pulses of the full frequency clock signal. The remaining N&#x2212;1 pulses of the clock signal may be inhibited. If a high priority transaction inbound for the functional circuit block is detected, an inserted pulse of the clock signal may be provided to the functional unit irrespective of when a most recent regular pulse was provided.

Growing community of inventors

Austin, TX, United States of America

Robert D Kenney