Abstract: Disclosed in some examples, are methods, systems, devices, and machine-readable mediums which provide for more efficient CGRA execution by assigning different initiation intervals to different PEs executing a same code base. The initiation intervals may be a multiple of each other and the PE with the lowest initiation interval may be used to execute instructions of the code that is to be executed at a greater frequency than other instructions than other instructions that may be assigned to PEs with higher initiation intervals.

Abstract: Devices and techniques for thread scheduling control and memory splitting in a processor are described herein. An apparatus includes a hardware interface configured to receive a first request to execute a first thread, the first request including an indication of a workload; and processing circuitry configured to: determine the workload to produce a metric based at least in part on the indication; compare the metric with a threshold to determine that the metric is beyond the threshold; divide, based at least in part on the comparison, the workload into a set of sub-workloads consisting of predefined number of equal parts from the workload; create a second request to execute a second thread, the second request including a first member of the set of sub-workloads; and process a second member of the set of sub-workloads in the first thread.

Abstract: System and techniques for variable execution time atomic operations are described herein. When an atomic operation for a memory device is received, the run length of the operation is measured. If the run length is beyond a threshold, a cache line for the operation is locked while the operation runs. A result of the operation is queued until it can be written to the cache line. At that point, the cache line is unlocked.

Abstract: Representative apparatus, method, and system embodiments are disclosed for a self-scheduling processor which also provides additional functionality. Representative embodiments include a self-scheduling processor, comprising: a processor core adapted to execute a received instruction; and a core control circuit adapted to automatically schedule an instruction for execution by the processor core in response to a received work descriptor data packet. In another embodiment, the core control circuit is also adapted to schedule a fiber create instruction for execution by the processor core, to reserve a predetermined amount of memory space in a thread control memory to store return arguments, and to generate one or more work descriptor data packets to another processor or hybrid threading fabric circuit for execution of a corresponding plurality of execution threads. Event processing, data path management, system calls, memory requests, and other new instructions are also disclosed.

Abstract: A command manager can be configured to enforce respective command execution policies for each of multiple queues according to respective command fence instructions. In an example, the command manager can be configured to receive a first packet comprising a first fence instruction and a first command for a first queue and, responsive to the first fence instruction indicating fence participation, increment a fence counter and provide the first command to a command execution unit, such as a memory controller of a memory device. The command manager can receive a first response message from the command execution unit based on the first command and can decrement the fence counter. In an example, the command manager comprises a portion of an accelerator device that uses an unordered interconnect, such as a Compute Express Link (CXL) interconnect, to communicate with a host device.

Abstract: System and techniques for recall pending cache line eviction are described herein. A queue that includes a deferred memory request is kept for a cache line. Metadata for the queue is stored in a cache line tag. When a recall is needed, the metadata is written from the tag to a first recall storage, referenced by a memory request ID. After the recall request is transmitted, the memory request ID is written to a second recall storage referenced by the message ID of the recall request. Upon receipt of a response to the recall request, the queue for the cache line can be restored by using the message ID in the response to lookup the memory request ID from the second recall storage, then using the memory request ID to lookup the metadata from the first recall storage, and then writing the metadata into the tag for the cache line.

Abstract: A three-dimensional stacked integrated circuit (3D SIC) having a non-volatile memory die, a volatile memory die, and a logic die. The non-volatile memory die, the volatile memory die, and the logic die are stacked. The 3D SIC is partitioned into a plurality of columns that are perpendicular to each of the stacked dies. Each column of the plurality of columns is configurable to be bypassed via configurable routes. When the configurable routes are used, functionality of a failing part of the column is re-routed to a corresponding effective part of a neighboring column.

Abstract: Disclosed in some examples are improvements to memory controllers on distributed memory systems that include a fine-grained data mover component that offloads management of memory commands accessing multiple smaller values to the memory controller. The fine-grained data mover (FGDM) may provide low host processing overhead that enables performance improvements for small task offloads. Work requests (“data mover calls”) may be sent by hosts to the FGDM without OS system calls. The FGDM is a virtually addressed data movement engine architected to transfer data at high transfer rates even during situations where the host has many small data movement requests and where the source and/or destination addresses are not memory controller friendly.

Abstract: Disclosed in some examples, are methods, systems, devices, and machine readable mediums that store instructions for programmable atomic transactions in a memory of the programmable atomic unit prior to execution of the programmable atomic transaction. The memory in some examples may be an instruction RAM. The memory in some examples may be partitioned into partitions of a fixed size that stores a same number of instructions. Each programmable atomic transaction may use one or more contiguously located instruction partitions. By loading the instructions ahead of time, the instructions are ready for execution when the transaction is requested.

Abstract: A three-dimensional stacked integrated circuit (3D SIC) that can have at least a first 3D XPoint (3DXP) die and, in some examples, can have at least a second 3DXP die too. In such examples, the first 3DXP die and the second 3DXP die can be stacked. The 3D SIC can be partitioned into a plurality of columns that are perpendicular to each of the stacked dies. In such examples, when a first column of the plurality of columns is determined as failing, data stored in the first column can be replicated to a second column of the plurality of columns. Also, for example, when a part of a first column of the plurality of columns is determined as failing, data stored in the part of the first column can be replicated to a corresponding part of a second column of the plurality of columns.

Abstract: Disclosed in some examples, are methods, systems, devices, and machine-readable mediums which provide for more efficient CGRA execution by assigning different initiation intervals to different PEs executing a same code base. The initiation intervals may be a multiple of each other and the PE with the lowest initiation interval may be used to execute instructions of the code that is to be executed at a greater frequency than other instructions than other instructions that may be assigned to PEs with higher initiation intervals.

Abstract: Methods, systems, and devices for shared virtual address spaces are described. In some examples, a globally shared address space may be shared across a plurality of memory devices that are included in one or more domains. A host system may set parameters for determining whether an address (e.g., a virtual address) is included within the globally shared address space, and whether the address is associated with a memory device. When a memory device receives a memory request (e.g., a data packet), a processing unit of the memory device may determine whether an address included in the memory request is associated with the memory device. The processing unit may either initiate an access operation on a physical address of the memory device or transmit the memory request to another memory device.

Abstract: A three-dimensional stacked integrated circuit (3D SIC) that can have at least a first 3D XPoint (3DXP) die and, in some examples, can have at least a second 3DXP die too. In such examples, the first 3DXP die and the second 3DXP die can be stacked. The 3D SIC can be partitioned into a plurality of columns that are perpendicular to each of the stacked dies. In such examples, when a first column of the plurality of columns is determined as failing, data stored in the first column can be replicated to a second column of the plurality of columns. Also, for example, when a part of a first column of the plurality of columns is determined as failing, data stored in the part of the first column can be replicated to a corresponding part of a second column of the plurality of columns.

Abstract: Devices and techniques for chained resource locking are described herein. Threads form a last-in-first-out (LIFO) queue on a resource lock to create a chained lock on the resource. A data store representing the lock for the resource holds the previous thread's identifier, enabling a subsequent thread to wake the previous thread using the identifier when the subsequent thread releases the lock. Generally, the thread releasing the lock need not interact with the data store, reducing contention for the data store among many threads.

Abstract: Disclosed in some examples are methods, systems, devices, and architectures which provide for techniques for memory device and memory fabric redundancy within distributed memory systems. In some examples, two memory devices are paired and each stores a same set of data such that writes to the memory devices are duplicated and reads may be satisfied from either device. In some examples, a memory processing unit (MPU) may be incorporated into the memory architecture to support these paired memory devices. The MPU may be placed between the host and a multi-planed memory fabric which connects to multi-ported CXL memory devices. In some examples, the MPU may also enable the use of alternative fabric links. That is, if a memory fabric link between the MPU and a memory device is unavailable, an alternative link may be utilized to restore connectivity to a memory device.

Abstract: Methods, systems, and devices for virtual addresses for a memory system are described. In some examples, a virtual address space may be shared across a plurality of memory devices that are included in one or more domains. The memory devices may be able to communicate with each other directly. For example, a first memory device may be configured to generate a data packet that includes an identifier and an address that is included in the shared virtual address space. The data packet may be transmitted to a second memory device based on the identifier, and the second memory device may access a physical address based on the address.

Abstract: Linear interpolation is performed within a memory system. The memory system receives a floating-point point index into an integer-indexed memory array. The memory system accesses the two values of the two adjacent integer indices, performs the linear interpolation, and provides the resulting interpolated value. In many system architectures, the critical limitation on system performance is the data transfer rate between memory and processing elements. Accordingly, reducing the amount of data transferred improves overall system performance and reduces power consumption.

Abstract: To implement a hybrid of software and hardware coherency management, a device can receive an unrestricted-access read request for memory that corresponds to a cache line from a first host and record an indication of the unrestricted-access read request with respect to the cache line. Here, the indication can include an identifier for the first host. However, if the device receives a shared-access read request for the cache line from a second host, the device does not record an identifier of the second host. Rather, the device can communicate an invalidation request for the cache line to the first host using the identifier for the first host to provoke the first host to write-back the data and invalidate the cache line.

Abstract: A method performed by a distributed computing system includes receiving a work packet from a separate computing device via a fabric interconnect at a command manager (CM) of a memory controller of a fabric attached memory (FAM) device, wherein the work packet includes a memory access to be performed by a FAM computing resource local to the FAM device; determining a work class of the work packet; placing the work packet in a CM work queue local to the CM for the work class when space is available in the CM work queue for the work class; and when the CM work queue for the work class is full, placing the work packet in a destination work queue according to an address included in the work packet, wherein the destination queue is implemented in a memory array of the FAM device external to the memory controller.

Abstract: A reconfigurable compute fabric can include multiple nodes, and each node can include multiple tiles with respective processing and storage elements. The tiles can be arranged in an array or grid and can be communicatively coupled. In an example, the tiles can be arranged in a one-dimensional array and each tile can be coupled to its respective adjacent neighbor tiles using a direct bus coupling. Each tile can be further coupled to at least one non-adjacent neighbor tile that is one tile, or device space, away using a passthrough bus. The passthrough bus can extend through intervening tiles.