Abstract: A method for operating a memory having a plurality of banks accessible in parallel, each bank including a plurality of grains accessible in parallel is provided. The method includes: based on a memory access request that specifies a memory address, identifying a set that stores data for the memory access request, wherein the set is spread across multiple grains of the plurality of grains; and performing operations to satisfy the memory access request, using entries of the set stored across the multiple grains of the plurality of grains.

Abstract: Data reuse cache techniques are described. In one example, a load instruction is generated by an execution unit of a processor unit. In response to the load instruction, data is loaded by a load-store unit for processing by the execution unit and is also stored to a data reuse cache communicatively coupled between the load-store unit and the execution unit. Upon receipt of a subsequent load instruction for the data from the execution unit, the data is loaded from the data reuse cache for processing by the execution unit.

Abstract: A disclosed computing device includes at least one prefetcher and a processing device communicatively coupled to the prefetcher. The processing device is configured to detect a throttling instruction that indicates a start of a throttling region. The computing device is further configured to prevent the prefetcher from being trained on one or more memory instructions included in the throttling region in response to the throttling instruction. Various other apparatuses, systems, and methods are also disclosed.

Abstract: Connection modification based on traffic pattern is described. In accordance with the described techniques, a traffic pattern of memory operations across a set of connections between at least one device and at least one memory is monitored. The traffic pattern is then compared to a threshold traffic pattern condition, such as an amount of data traffic in different directions across the connections. A traffic direction of at least one connection of the set of connections is modified based on the traffic pattern corresponding to the threshold traffic pattern condition.

Abstract: Methods and systems are described. A system includes a redundant shader pipe array that performs rendering calculations on data provided thereto and a shader pipe array that includes a plurality of shader pipes, each of which performs rendering calculations on data provided thereto. The system also includes a circuit that identifies a defective shader pipe of the plurality of shader pipes in the shader pipe array. In response to identifying the defective shader pipe, the circuit generates a signal. The system also includes a redundant shader switch. The redundant shader switch receives the generated signal, and, in response to receiving the generated signal, transfers the data for the defective shader pipe to the redundant shader pipe array.

Abstract: A method and apparatus for training data in a computer system includes reading data stored in a first memory address in a memory and writing it to a buffer. Training data is generated for transmission to the first memory address. The data is transmitted to the first memory address. Information relating to the training data is read from the first memory address and the stored data is read from the buffer and written to the memory area where the training data was transmitted.

Abstract: Systems, apparatuses, and methods for adaptively mapping a machine learning model to a multi-core inference accelerator engine are disclosed. A computing system includes a multi-core inference accelerator engine with multiple inference cores coupled to a memory subsystem. The system also includes a control unit which determines how to adaptively map a machine learning model to the multi-core inference accelerator engine. In one implementation, the control unit selects a mapping scheme which minimizes the memory bandwidth utilization of the multi-core inference accelerator engine. In one implementation, this mapping scheme involves having one inference core of the multi-core inference accelerator engine fetch given data and broadcast the given data to other inference cores of the inference accelerator engine. Each inference core fetches second data unique to the respective inference core.

Abstract: Methods and systems are disclosed for cross-chiplet performance data streaming. Techniques disclosed include accumulating, by a subservient chiplet, event data associated with an event indicative of a performance aspect of the subservient chiplet; sending, by the subservient chiplet, the event data over a chiplet bus to a master chiplet; and adding, by the master chiplet, the received event data to an event record, the event record containing previously received, from the subservient chiplet over the chiplet bus, event data associated with the event.

Abstract: Methods and systems are disclosed for performing dataflow execution by an accelerated processing unit (APU). Techniques disclosed include decoding information from one or more dataflow instructions. The decoded information is associated with dataflow execution of a computational task. Techniques disclosed further include configuring, based on the decoded information, dataflow circuitry, and, then, executing the dataflow execution of the computational task using the dataflow circuitry.

Abstract: Systems and methods related to controlling clock signals for clocking shader engines modules (SEs) and non-shader-engine modules (nSEs) of a graphics processing unit (GPU) are provided. One or more dividers receive a clock signal CLK and output a clock signal CLKA to the SEs and output a clock signal CLKB to the nSEs. The frequencies of CLKA and CLKB are independently selected based on sets of performance counter data monitored at the SEs and nSEs, respectively. The clock signal frequency for either the SEs or the nSEs is reduced when the corresponding sets of performance counter data indicates a comparatively lower processing workload for the SEs or for the nSEs.

Abstract: Disclosed is a system and method for use in a cache for suppressing modification of cache line. The system and method includes a processor and a memory operating cooperatively with a cache controller. The memory includes a coherence directory stored within a cache created to track at least one cache line in the cache via the cache controller. The processor instructs a cache controller to store a first data in a cache line in the cache. The cache controller tags the cache line based on the first data. The processor instructs the cache controller to store a second data in the cache line in the cache causing eviction of the first data from the cache line. The processor compares based on the tagging the first data and the second data and suppresses modification of the cache line based on the comparing of the first data and the second data.

Abstract: A processing system includes hull shader circuitry that launches thread groups including one or more primitives. The hull shader circuitry also generates tessellation factors that indicate subdivisions of the primitives. The processing system also includes throttling circuitry that estimates a primitive launch time interval for the domain shader based on the tessellation factors and selectively throttles launching of the thread groups from the hull shader circuitry based on the primitive launch time interval of the domain shader and a hull shader latency. In some cases, the throttling circuitry includes a first counter that is incremented in response to launching a thread group from the buffer and a second counter that modifies the first counter based on a measured latency of the domain shader.

Abstract: Techniques for invalidating cache lines are provided. The techniques include issuing, to a first level of a memory hierarchy, a weak exclusive read request for a speculatively executing store instruction; determining whether to invalidate one or more cache lines associated with the store instruction in one or more memories; and issuing the weak invalidation request to additional levels of the memory hierarchy.

Abstract: Systems, apparatuses, and methods for implementing duplicated registers for access by initiators across multiple semiconductor dies are disclosed. A system includes multiple initiators on multiple semiconductor dies of a chiplet processor. One of the semiconductor dies is the master die, and this master die has copies of registers which can be accessed by the multiple initiators on the multiple semiconductor dies. When a given initiator on a given secondary die generates a register access, the register access is routed to the master die and a particular duplicate copy of the register maintained for the given secondary die. From the point of view of software, the multiple semiconductor dies appear as a single die, and the multiple initiators appear as a single initiator. Multiple types of registers can be maintained by the master die, with a flush register being one of the register types.

Abstract: Systems, apparatuses, and methods for performing command buffer gang submission are disclosed. A system includes at least first and second processors and a memory. The first processor (e.g., CPU) generates a command buffer and stores the command buffer in the memory. A mechanism is implemented where a granularity of work provided to the second processor (e.g., GPU) is increased which, in turn, increases the opportunities for parallel work. In gang submission mode, the user-mode driver (UMD) specifies a set of multiple queues and command buffers to execute on those multiple queues, and that work is guaranteed to execute as a single unit from the GPU operating system scheduler point of view. Using gang submission, synchronization between command buffers executing on multiple queues in the same submit is safe. This opens up optimization opportunities for application use (explicit gang submission) and for internal driver use (implicit gang submission).

Abstract: Generating optimization instructions for data processing pipelines is described. A pipeline optimization system computes resource usage information that describes memory and compute usage metrics during execution of each stage of the data processing pipeline. The system additionally generates data storage information that describes how data output by each pipeline stage is utilized by other stages of the pipeline. The pipeline optimization system then generates the optimization instructions to control how memory operations are performed for a specific data processing pipeline during execution. In implementations, the optimization instructions cause a memory system to discard data (e.g., invalidate cache entries) without copying the discarded data to another storage location after the data is no longer needed by the pipeline. The optimization instructions alternatively or additionally control at least one of evicting, writing-back, or prefetching data to minimize latency during pipeline execution.

Abstract: A memory controller coupled to a memory module receives both processing-in-memory (PIM) requests and memory requests from a host (e.g., a host processor). The memory controller issues PIM requests to one group of memory banks and concurrently issues memory requests to one or more other groups of memory banks. Accordingly, memory requests are performed on groups of memory banks that would otherwise be idle while PIM requests are performed on the one group of memory banks. Optionally, the memory controller coupled to the memory module also takes various actions when switching between operating in a PIM mode and a non-processing-in-memory mode to reduce or hide overhead when switching between the two modes.

Abstract: Multi-level cell memory management techniques are described. In one example, the memory controller is configured to control whether a single-level cell operation or a multi-level cell operation to be used using different mapping schemes. The single-level cell operation, for instance, is usable to store a data word using two states whereas the multi-level cell operation is usable to store the data word by also using an intermediate state. In order to store the data word using two states, the memory controller is configurable to separate the data word across two word lines in the physical memory. In an implementation, use of the different operations and corresponding mapping schemes by the memory controller alternates between adjacent word lines in physical memory.

Abstract: A/D bit storage, processing, and mode management techniques through use of a dense A/D bit representation are described. In one example, a memory management unit employs an A/D bit representation generation module to generate the dense A/D bit representation. In an implementation, the A/D bit representation is stored adjacent to existing page table structures of the multilevel page table hierarchy. In another example, memory management unit supports use of modes as part of A/D bit storage.

Abstract: Predicates for processing in memory is described. In accordance with the described techniques, a predicate instruction to compute a conditional value based on data stored in a memory is provided to a processing-in-memory component. A response that includes the conditional value computed by the processing-in-memory component is received, and the conditional value is stored in a predicate register. One or more conditional instructions are provided to the processing-in-memory component based on the conditional value stored in the predicate register.