Abstract: A processor includes a plurality of execution units that perform respective portions of a parallel execution. As part of the parallel execution, each execution unit requests respective execution data via a respective memory request. A request aggregation circuit combines received memory requests from the execution units. Combining the requests includes identifying the memory requests as corresponding to the same execution data, sending a single representative memory request for the execution data, receiving a single instance of the execution data, and providing the respective execution data to each requesting execution unit.

Abstract: A DMA controller iteratively loads regions of tensor data from global memory to a shared memory of a processor to generate an output from matrix multiplication in a format in which rows of data are contiguous in memory. In a first iteration, the DMA controller loads a first region of data that includes a plurality of rows, each row separated by a tile stride from the preceding row, from the tile to a first contiguous region of the shared memory. In a second iteration, the DMA controller loads a second region of data that includes a plurality of rows, each row separated by a tile stride from the preceding row, from the tile to a second contiguous region of the shared memory. The second region of data is offset from the first region of data in global memory by a configurable offset.

Abstract: A direct memory access (DMA) controller issuing memory copy operations on behalf of a shader at a parallel processor stops issuing copy operations upon a context switch at the shader for a wave. The DMA controller or a trap handler associated with the shader saves the incomplete copy operations to a region of global memory, from which the incomplete operations are restored upon a context resume for the wave.

Abstract: A processor employs one or more branch predictors to issue branch predictions for each thread executing at an instruction pipeline. Based on the branch predictions, the processor determines a branch prediction confidence for each of the executing threads, whereby a lower confidence level indicates a smaller likelihood that the corresponding thread will actually take the predicted branch. Because speculative execution of an untaken branch wastes resources of the instruction pipeline, the processor prioritizes threads associated with a higher confidence level for selection at the stages of the instruction pipeline.

Abstract: In an illustrative example, a system includes a resource and a first processor. The first processor is configured to access the resource based on a first physical address space and to generate a request for access to the resource. The request has a first format. The system further includes a second processor configured to access the resource based on a second physical address space. The system also includes a device coupled to the resource and to the first processor. The device is configured to receive the request, to generate a message having a second format based on the request, to send the message to the resource, and to provide a reply to the request to the first processor.

Abstract: An apparatus includes a set of one or more processing cores, a thread dispatcher, and an event register of a first compute unit. The set of one or more processing cores is configured to execute a set of threads. The thread dispatcher is coupled to the set of one or more processing cores and is configured to select threads of the set of threads for execution by the set of one or more processing cores. The thread dispatcher is further configured to refrain from selecting a first thread of the set of threads for execution in response to a first value of one or more bits of the event register and to select the first thread for execution in response to a second value of the one or more bits.

Abstract: A processing device includes one or more queues to convey data between a producing processor unit in a first timing domain and a consuming processor unit in a second timing domain that is asynchronous with the first timing domain. A system management unit configures a first operating frequency of the producing processor unit and a second operating frequency of the consuming processor unit based on a power constraint for the processing device and a target size of the one or more queues.

Abstract: A processor employs a prefetch prediction module that predicts, for each prefetch request, whether the prefetch request is likely to be satisfied from (“hit”) the cache. The arbitration priority of prefetch requests that are predicted to hit the cache is reduced relative to demand requests or other prefetch requests that are predicted to miss in the cache. Accordingly, an arbiter for the cache is less likely to select prefetch requests that hit the cache, thereby improving processor throughput.

Abstract: A processor transfers prefetch requests from their targeted cache to another cache in a memory hierarchy based on a fullness of a miss address buffer (MAB) or based on confidence levels of the prefetch requests. Each cache in the memory hierarchy is assigned a number of slots at the MAB. In response to determining the fullness of the slots assigned to a cache is above a threshold when a prefetch request to the cache is received, the processor transfers the prefetch request to the next lower level cache in the memory hierarchy. In response, the data targeted by the access request is prefetched to the next lower level cache in the memory hierarchy, and is therefore available for subsequent provision to the cache. In addition, the processor can transfer a prefetch request to lower level caches based on a confidence level of a prefetch request.

Abstract: A processor employs a prefetch prediction module that predicts, for each prefetch request, whether the prefetch request is likely to be satisfied from (“hit”) the cache. The arbitration priority of prefetch requests that are predicted to hit the cache is reduced relative to demand requests or other prefetch requests that are predicted to miss in the cache. Accordingly, an arbiter for the cache is less likely to select prefetch requests that hit the cache, thereby improving processor throughput.

Abstract: A processor includes storage elements to store a first and second value, as well as a plurality of hash units coupled to the storage elements. Each hash unit performs a hash operation using the first value and the second value to generate a corresponding hash result value. The processor further includes selection logic to select a hash result value from the hash result values generated by the plurality of hash units responsive to a selection input generated from another hash operation performed using the first value and the second value. A method includes predicting whether a branch instruction is taken based on a prediction value stored at an entry of a branch prediction table indexed by an index value selected from a plurality of values concurrently generated from an address value of the branch instruction and a branch history value representing a history of branch directions at the processor.

Abstract: A processor employs one or more branch predictors to issue branch predictions for each thread executing at an instruction pipeline. Based on the branch predictions, the processor determines a branch prediction confidence for each of the executing threads, whereby a lower confidence level indicates a smaller likelihood that the corresponding thread will actually take the predicted branch. Because speculative execution of an untaken branch wastes resources of the instruction pipeline, the processor prioritizes threads associated with a higher confidence level for selection at the stages of the instruction pipeline.

Abstract: A processor transfers prefetch requests from their targeted cache to another cache in a memory hierarchy based on a fullness of a miss address buffer (MAB) or based on confidence levels of the prefetch requests. Each cache in the memory hierarchy is assigned a number of slots at the MAB. In response to determining the fullness of the slots assigned to a cache is above a threshold when a prefetch request to the cache is received, the processor transfers the prefetch request to the next lower level cache in the memory hierarchy. In response, the data targeted by the access request is prefetched to the next lower level cache in the memory hierarchy, and is therefore available for subsequent provision to the cache. In addition, the processor can transfer a prefetch request to lower level caches based on a confidence level of a prefetch request.