Abstract: Block data load with transpose techniques are described. In one example, an input is received, at a control unit, specifying an instruction to load a block of data to at least one memory module using a transpose operation. Responsive to the receiving the input by the control unit, the block of data is caused to be loaded to the at least one memory module by transposing the block of data to form a transposed block of data and storing the transposed block of data in the at least one memory.

Abstract: A software-based instruction scoreboard indicates dependencies between closely-issued instructions issued to an arithmetic logic unit (ALU) pipeline. The software-based instruction scoreboard inserts one or more control words into the command stream between the dependent instructions, which is then executed by the ALU pipeline. The control words identify the instruction(s) upon which the dependent instructions depend (parent instructions) so that the GPU hardware can ensure that the ALU pipeline does not stall while the dependent instruction waits for results from the parent instruction.

Abstract: A disclosed technique includes executing, for a first wavefront, a barrier arrival notification instruction, for a first barrier, indicating arrival at a first barrier point; performing, for the first wavefront, work prior to the first barrier point; executing, for the first wavefront, a barrier check instruction; and executing, for the first wavefront, at a control flow path based on a result of the barrier check instruction.

Abstract: A processing system executes wavefronts at multiple arithmetic logic unit (ALU) pipelines of a single instruction multiple data (SIMD) unit in a single execution cycle. The ALU pipelines each include a number of ALUs that execute instructions on wavefront operands that are collected from vector general process register (VGPR) banks at a cache and output results of the instructions executed on the wavefronts at a buffer. By storing wavefronts supplied by the VGPR banks at the cache, a greater number of wavefronts can be made available to the SIMD unit without increasing the VGPR bandwidth, enabling multiple ALU pipelines to execute instructions during a single execution cycle.

Abstract: Methods and systems are disclosed for executing operations on single-instruction-multiple-data (SIMD) units. Techniques disclosed perform a dot product operation on input data during one computer cycle, including convolving the input data, generating intermediate data, and applying one or more transitional operations to the intermediate data to generate output data. Aspects described, wherein the input data is an input to a layer of a convolutional neural network and the generated output data is the output of the layer.

Abstract: Methods and systems are disclosed for executing a collaborative task in a shader system. Techniques disclosed include receiving, by the system, input data and computing instructions associated with the collaborative task, as well as a configuration setting, causing the system to operate in a takeover mode. The system then launches, exclusively in one workgroup processor, a workgroup including wavefronts configured to execute the collaborative task.

Abstract: The address of the draw or dispatch packet responsible for creating an exception is tied to a shader/wavefront back to the draw command from which it originated. In various embodiments, a method of operating a graphics pipeline and exception handling includes receiving, at a command processor of a graphics processing unit (GPU), an exception signal indicating an occurrence of a pipeline exception at a shader stage of a graphics pipeline. The shader stage generates an exception signal in response to a pipeline exception and transmits the exception signal to the command processor. The command processor determines, based on the exception signal, an address of a command packet responsible for the occurrence of the pipeline exception.

Abstract: A processing unit includes a first memory device and a second memory device. The first memory device includes a first plurality of general purpose registers (GPRs) and the second memory device includes a second plurality of GPRs. The second memory device includes fewer GPRs than the first memory device. Program data is stored at the first memory device and the second memory device based on expected frequency of accesses associated with the program data.

Abstract: A software-based instruction scoreboard indicates dependencies between closely-issued instructions issued to an arithmetic logic unit (ALU) pipeline. The software-based instruction scoreboard inserts one or more control words into the command stream between the dependent instructions, which is then executed by the ALU pipeline. The control words identify the instruction(s) upon which the dependent instructions depend (parent instructions) so that the GPU hardware can ensure that the ALU pipeline does not stall while the dependent instruction waits for results from the parent instruction.

Abstract: A processing system executes wavefronts at multiple arithmetic logic unit (ALU) pipelines of a single instruction multiple data (SIMD) unit in a single execution cycle. The ALU pipelines each include a number of ALUs that execute instructions on wavefront operands that are collected from vector general process register (VGPR) banks at a cache and output results of the instructions executed on the wavefronts at a buffer. By storing wavefronts supplied by the VGPR banks at the cache, a greater number of wavefronts can be made available to the SIMD unit without increasing the VGPR bandwidth, enabling multiple ALU pipelines to execute instructions during a single execution cycle.

Abstract: A processing unit includes a plurality of processing elements and one or more caches. A first thread executes a program that includes one or more prefetch instructions to prefetch information into a first cache. Prefetching is selectively enabled when executing the first thread on a first processing element dependent upon whether one or more second threads previously executed the program on the first processing element. The first thread is then dispatched to execute the program on the first processing element. In some cases, a dispatcher receives the first thread four dispatching to the first processing element. The dispatcher modifies the prefetch instruction to disable prefetching into the first cache in response to the one or more second threads having previously executed the program on the first processing element.

Abstract: The address of the draw or dispatch packet responsible for creating an exception is tied to a shader/wavefront back to the draw command from which it originated. In various embodiments, a method of operating a graphics pipeline and exception handling includes receiving, at a command processor of a graphics processing unit (GPU), an exception signal indicating an occurrence of a pipeline exception at a shader stage of a graphics pipeline. The shader stage generates an exception signal in response to a pipeline exception and transmits the exception signal to the command processor. The command processor determines, based on the exception signal, an address of a command packet responsible for the occurrence of the pipeline exception.

Abstract: A system includes a processor configured to operate in at least a first mode and a second mode. In the first mode the first processor operates to execute an instruction for an entire wavefront before executing a next instruction for the entire wavefront. In the second mode the processor operates to execute a set instructions for a portion of a wavefront before executing the set instructions for another portion of the same wavefront. The system further includes a memory coupled to the processor. The memory is configured to store a shader program for execution by the processor, wherein the shader program includes at least one indication associated with one of the first mode or the second mode. The processor is further to implement one of the first mode or the second mode while executing the shader program responsive to the at least one indication present in the first shader program.

Abstract: A processing unit includes a plurality of processing elements and one or more caches. A first thread executes a program that includes one or more prefetch instructions to prefetch information into a first cache. Prefetching is selectively enabled when executing the first thread on a first processing element dependent upon whether one or more second threads previously executed the program on the first processing element. The first thread is then dispatched to execute the program on the first processing element. In some cases, a dispatcher receives the first thread four dispatching to the first processing element. The dispatcher modifies the prefetch instruction to disable prefetching into the first cache in response to the one or more second threads having previously executed the program on the first processing element.

Abstract: A graphics processing unit is disclosed, the graphics processing unit having a processor having one or more SIMD processing units, and a local data share corresponding to one of the one or more SIMD processing units, the local data share comprising one or more low latency accessible memory regions for each group of threads assigned to one or more execution wavefronts, and a global data share comprising one or more low latency memory regions for each group of threads.

Abstract: A graphics processing unit is disclosed, the graphics processing unit having a processor having one or more SIMD processing units, and a local data share corresponding to one of the one or more SIMD processing units, the local data share comprising one or more low latency accessible memory regions for each group of threads assigned to one or more execution wavefronts, and a global data share comprising one or more low latency memory regions for each group of threads.

Abstract: A graphics processing unit is disclosed, the graphics processing unit having a processor having one or more SIMD processing units, and a local data share corresponding to one of the one or more SIMD processing units, the local data share comprising one or more low latency accessible memory regions for each group of threads assigned to one or more execution wavefronts, and a global data share comprising one or more low latency memory regions for each group of threads.

Abstract: A method of allocating a memory to a plurality of concurrent threads is presented. The method includes dynamically determining writer threads each having at least one pending write to the memory; and dynamically allocating respective contiguous blocks in the memory for each of the writer threads. Another method of allocating a memory to a plurality of concurrent threads includes launching the plurality of threads as a plurality of wavefronts, dynamically determining a group of wavefronts each having at least one thread requiring a write to the memory, and dynamically allocating respective contiguous blocks in the memory for each wavefront from the group of wavefronts.

Abstract: Systems and methods to improve performance in a graphics processing unit are described herein. Embodiments achieve power saving in a graphics processing unit by dynamically activating/deactivating individual SIMDs in a shader complex that comprises multiple SIMD units. On-the-fly dynamic disabling and enabling of individual SIMDs provides flexibility in achieving a required performance and power level for a given processing application. Embodiments of the invention also achieve dynamic medium grain clock gating of SIMDs in a shader complex. Embodiments reduce switching power by shutting down clock trees to unused logic by providing a clock on demand mechanism. In this way, embodiments enhance clock gating to save more switching power for the duration of time when SIMDs are idle (or assigned no work). Embodiments can also save leakage power by power gating SIMDs for a duration when SIMDs are idle for an extended period of time.

Abstract: A system and method to optimize processor performance and minimizing average thread latency by selectively loading a cache when a program state, resources required for execution of a program or the program itself change, is described. An embodiment of the invention supports a “cache priming program” that is selectively executed for a first thread/program/sub-routine of each process. Such a program is optimized for situations when instructions and other program data are not yet resident in cache(s), and/or whenever resources required for program execution or the program itself changes. By pre-loading the cache with two resources required for two instructions for only a first thread, average thread latency is reduced because the resources are already present in the cache.