Nvidia Corporation

Christopher Frederick Lamb

Lacky V Shah

Philip Alexander Cuadra

Stephen Jones

Jerome Francis Duluk, Jr

Monier Maher

Ignacio Llamas

Sanjay J Patel

Daniel Elliot Wexler

Vivek Kini

Jack H Choquette

Jean Pierre Bordes

Mark Hairgrove

Peter Y Hsu

Rajballav Dash

Ming Liang Milton Lei

Apparatuses, systems, and techniques for performing nested kernel execution within a parallel processing subsystem. In at least one embodiment, a parent thread launches a nested child grid on the parallel processing subsystem, and enables the parent thread to perform a thread synchronization barrier on the child grid for proper execution semantics between the parent thread and the child grid.

Technique for computational nested parallelism

Techniques are disclosed for sharing of data exchange among kernels (each a set of instructions) executing on a system having multiple processing units. In an embodiment, each processing unit includes an on-chip scratchpad memory that can be accessed by the kernels executing on the processing unit. All or a portion of the scratchpad memory can be allocated and configured, for example, such that the scratchpad is accessible to multiple kernels in parallel, to one or more kernels in serial, or a combination of both.

Persistent scratchpad memory for data exchange between programs

One embodiment of the present invention sets forth a technique for instruction level execution preemption. Preempting at the instruction level does not require any draining of the processing pipeline. No new instructions are issued and the context state is unloaded from the processing pipeline. Any in-flight instructions that follow the preemption command in the processing pipeline are captured and stored in a processing task buffer to be reissued when the preempted program is resumed. The processing task buffer is designated as a high priority task to ensure the preempted instructions are reissued before any new instructions for the preempted context when execution of the preempted context is restored.

Software-assisted instruction level execution preemption

One embodiment sets forth a method for assigning priorities to kernels launched by a software application and executed within a stream of work on a parallel processing subsystem that supports dynamic parallelism. First, the software application assigns a maximum nesting depth for dynamic parallelism. The software application then assigns a stream priority to a stream. These assignments cause a driver to map the stream priority to a device priority and, subsequently, associate the device priority with the stream. As part of the mapping, the driver ensures that each device priority is at least the maximum nesting depth higher than the device priorities associated with any lower priority streams. Subsequently, the driver launches any kernel included in the stream with the device priority associated with the stream. Advantageously, by strategically assigning the maximum nesting depth and prioritizing streams, an application developer may increase the overall processing efficiency of the software application.

Techniques for sharing priorities between streams of work and dynamic parallelism

One embodiment of the present invention sets forth a technique for performing nested kernel execution within a parallel processing subsystem. The technique involves enabling a parent thread to launch a nested child grid on the parallel processing subsystem, and enabling the parent thread to perform a thread synchronization barrier on the child grid for proper execution semantics between the parent thread and the child grid. This technique advantageously enables the parallel processing subsystem to perform a richer set of programming constructs, such as conditionally executed and nested operations and externally defined library functions without the additional complexity of CPU involvement.

One embodiment sets forth a method for guiding the order in which a parallel processing subsystem executes memory copies. A driver creates semaphores for all but the lowest priority included in a plurality of priorities and associates one priority with each copy hardware channel included in the parallel processing subsystem. The driver then aliases prioritized streams to the copy hardware channels based on the priorities. Upon receiving a request to execute a memory copy within one of the streams, the driver inserts commands into the aliased copy hardware channel. These commands use the semaphores to direct the parallel processing subsystem to execute the memory copy based on the priority of the copy hardware channel. Advantageously, by assigning priorities to streams and, subsequently, strategically requesting memory copies within the prioritized streams, an application developer may fine-tune their software application to increase the overall processing efficiency of the software application.

Techniques for assigning priorities to memory copies

A system for performing data-parallel operations and task-parallel operations. A first switch fabric node (SFN) includes first and second lane processing engines (LPEs). The first LPE includes a first set of lane processing units (LPUs) configured to perform data-parallel operations, where each LPU performs a set of operations, and each LPU uses a different set of data for the set of operations, and each LPU within the first set of LPUs uses a different set of data for the set of operations. The second LPE includes a second set of LPUs configured to perform task-parallel operations, where each LPU performs a different set of operations. A processing control engine (PCE) is configured to distribute instructions and data to the first LPE and the second LPE. Advantageously, data parallel operations and task parallel operations are able to be performed on the same processor simultaneously.

System of lanes of processing units receiving instructions via shared memory units for data-parallel or task-parallel operations

An integrated circuit comprises an external memory, a plurality of parallel connected Vector Processing Engines (VPEs), and an External Memory Unit (EMU) providing a data transfer path between the VPEs and the external memory. Each VPE contains a plurality of data processing units and a message queuing system adapted to transfer messages between the data processing units and other components of the integrated circuit.

Growing community of inventors

Los Gatos, CA, United States of America

Christopher Frederick Lamb