Apple Inc.

Nintendo Co., Ltd.

Winnie W Yeung

Anthony P DeLaurier

Zelin Zhang

Cheng Li

Leela Kishore Kothamasu

Benjiman L Goodman

Howard H Cheng

Yoong Chert Foo

Kenneth Christian Dyke

Tyson J Bergland

Srinivasa Rangan Sridharan

Karthik Ramani

Farhad Fouladi

David A Gotwalt

Muditha Kanchana

Karl D Mann

Wolfgang H Klingauf

Luc R Semeria

Guanlan Xu

Eddie M Robinson

Terence M Potter

Jonathan M Redshaw

Aroun Demeure

Donald R Desota

David K Li

Hungse Cha

Techniques are disclosed relating to graphics processor data caches. In some embodiments, datapath executes instructions that operate on input operands from architectural registers. Data cache circuitry caches architectural register data for the datapath circuitry. Scoreboard circuitry tracks, for a given architectural register: map information that indicates whether the architectural register is mapped to an entry of the data cache circuitry and a pointer to the entry of the data cache circuitry. Tiered scoreboard circuitry and data storage circuitry may be implemented (e.g., to provide fast scoreboard access for active threads and to give a landing spot for long-latency data retrieval operations). Various disclosed techniques may improve cache performance, reduce power consumption, reduce area, or some combination thereof.

Scoreboard for register data cache

Techniques are disclosed relating to multi-block fetches for cache misses. In some embodiments, cache tag circuitry maintains a tag value that is shared by multiple cache blocks. In response to a miss, the cache may initiate a fetch request to a next level cache or memory. Aggregation circuitry may aggregate multiple fetch requests for cache blocks that share the tag value and fetch circuitry may initiate a single multi-block fetch operation to the next level cache or memory that returns cache blocks for the aggregated multiple fetch requests. In various embodiments, disclosed techniques may improve performance (e.g., by reducing fetch bus transactions), reduce power consumption, or both, relative to traditional techniques.

Multi-block cache fetch techniques

Techniques are disclosed relating to eviction control for cache lines that store register data. In some embodiments, memory hierarchy circuitry is configured to provide memory backing for register operand data in one or more cache circuits. Lock circuitry may control a first set of lock indicators for a set of registers for a first thread, including to assert one or more lock indicators for registers that are indicated, by decode circuitry, as being utilized by decoded instructions of the first thread. The lock circuitry may preserve register operand data in the one or more cache circuits, including to prevent eviction of a given cache line from a cache circuit based on an asserted lock indicator. The lock circuitry may clear the first set of lock indicators in response to a reset event. Disclosed techniques may advantageously retain relevant register information in the cache with limited control circuit area.

Cache control to preserve register data

Techniques are disclosed relating to cache footprint management. In some embodiments, execution circuitry is configured to perform operations for instructions from multiple threads in parallel. Cache circuitry may store information operated on by threads executed by the execution circuitry. Scheduling circuitry may arbitrate among threads to schedule threads for execution by the execution circuitry. Tracking circuitry may determine one or more performance metrics for the cache circuitry. Control circuitry may, based on the one or more performance metrics meeting a threshold, reduce a limit on a number of threads considered for arbitration by the scheduling circuitry, to control a footprint of information stored by the cache circuitry. Disclosed techniques may advantageously reduce or avoid cache thrashing for certain processor workloads.

Cache footprint management

Techniques are disclosed relating to arbitration for computer memory resources. In some embodiments, an apparatus includes queue circuitry that implements multiple queues configured to queue requests to access a memory bus. Control circuitry may, in response to detecting a first threshold condition associated with the queue circuitry, generate a first snapshot that indicates numbers of requests in respective queues of the multiple queues at a first time. The control circuitry may generate a second snapshot that indicates numbers of requests in respective queues of the multiple queues at a second time that is subsequent to the first time. The control circuitry may arbitrate between requests from the multiple queues to select requests to access the memory bus, where the arbitration is based on snapshots to which requests from the multiple queues belong. Disclosed techniques may approximate age-based scheduling while reducing area and power consumption.

Snapshot arbitration techniques for memory requests

Techniques are disclosed relating to compression of data stored at different cache levels. In some embodiments, a memory system implements a storage hierarchy that includes first cache circuitry and second cache circuitry at different levels of the hierarchy. Processor circuitry generates write data to be written to the memory system. In some embodiments, first compression circuitry is configured to compress a first block of write data in response to full accumulation of the first block in the first cache circuitry and second compression circuitry is configured to compress a second block of write data in response to full accumulation of the second block in the second cache circuitry. Write circuitry may write the first and second compressed blocks of data in a single combined write to a higher level in the storage hierarchy.

Compression techniques and hierarchical caching

Techniques are disclosed relating to caching for address translation. In some embodiments, address translation circuitry is configured to process requests to translate addresses in a first address space to addresses in a second address space. The translation circuitry may include cache circuitry configured to store translation information, arbitration circuitry configured to arbitrate among ready requests for access to entries of the cache, and hazard circuitry. The hazard circuitry may assign a first request to an ready status the arbitration circuitry based on detection of an absence of hazards for a first address of the first request and add a second request to a queue of requests for the arbitration circuitry based on detection of a hazard for a second address of the second request. Independent arbitration for requests without hazards may improve performance in various aspects, relative to traditional techniques.

Cache arbitration for address translation requests

Techniques are disclosed relating to compression of data stored at different cache levels. In some embodiments, programmable shader circuitry is configured to execute program instructions of compute kernels that write pixel data. In some embodiments, a first cache is configured to store pixel write data from the programmable shader circuitry and first compression circuitry is configured to compress a first block of pixel write data in response to full accumulation of the first block in the first cache circuitry. In some embodiments, second cache circuitry is configured to store pixel write data from the programmable shader circuitry at a higher level in a storage hierarchy than the first cache circuitry and second compression circuitry is configured to compress a second block of pixel write data in response to full accumulation of the second block in the second cache circuitry. In some embodiments, write circuitry is configured to write the first and second compressed blocks of pixel data in a combined write to a higher level in the storage hierarchy.

Compression techniques for pixel write data

Systems, apparatuses, and methods for efficiently allocating data in a cache are described. In various embodiments, a processor decodes an indication in a software application identifying a temporal data set. The data set is flagged with a data set identifier (DSID) indicating temporal data to drop after consumption. When the data set is allocated in a cache, the data set is stored with a non-replaceable attribute to prevent a cache replacement policy from evicting the data set before it is dropped. A drop command with an indication of the DSID of the data set is later issued after the data set is read (consumed). A copy of the data set is not written back to the lower-level memory although the data set is removed from the cache. An interrupt is generated to notify firmware or other software of the completion of the drop command.

Cache drop feature to increase memory bandwidth and save power

A graphics system including a custom graphics and audio processor produces exciting 2D and 3D graphics and surround sound. The system includes a graphics and audio processor including a 3D graphics pipeline and an audio digital signal processor. A memory controller performs a wide range of memory control related functions including arbitrating between various competing resources seeking access to main memory, handling memory latency and bandwidth requirements of the resources requesting memory access, buffering writes to reduce bus turn around, refreshing main memory, and protecting main memory using programmable registers. The memory controller minimizes memory read/write switching using a 'global' write queue which queues write requests from various diverse competing resources. In this fashion, multiple competing resources for memory writes are combined into one resource from which write requests are obtained. Memory coherency issues are addressed both within a single resource that has both read and write capabilities and among different resources by efficiently flushing write buffers associated with a resource.

Growing community of inventors

San Jose, CA, United States of America

Winnie W Yeung