Nvidia Corporation

Gautam Chakrabarti

Yuan Lin

Jaydeep Marathe

Amit Sabne

Okwan Kwon

Vinod Grover

Ronny Krashinsky

Xiangyun Kong

Yunsup Lee

A system and method for compiling or runtime executing a fork-join data parallel program with function calls. In one embodiment, the system includes: (1) a partitioner operable to partition groups into a master group and at least one worker group and (2) a thread designator associated with the partitioner and operable to designate only one thread from the master group for execution and all threads in the at least one worker group for execution.

System and method for compiling or runtime executing a fork-join data parallel program with function calls on a single-instruction-multiple-thread processor

A system and method of translating functions of a program. In one embodiment, the system includes: (1) a local-scope variable identifier operable to identify local-scope variables employed in the at least some of the functions as being either thread-shared local-scope variables or thread-private local-scope variables and (2) a function translator associated with the local-scope variable identifier and operable to translate the at least some of the functions to cause thread-shared memory to be employed to store the thread-shared local-scope variables and thread-private memory to be employed to store the thread-private local-scope variables.

System and method for translating program functions for correct handling of local-scope variables and computing system incorporating the same

A system and method for allocating shared memory of differing properties to shared data objects and a hybrid stack data structure. In one embodiment, the system includes: (1) a hybrid stack creator configured to create, in the shared memory, a hybrid stack data structure having a lower portion having a more favorable property and a higher portion having a less favorable property and (2) a data object allocator associated with the hybrid stack creator and configured to allocate storage for shared data object in the lower portion if the lower portion has a sufficient remaining capacity to contain the shared data object and alternatively allocate storage for the shared data object in the higher portion if the lower portion has an insufficient remaining capacity to contain the shared data object.

System and method for allocating memory of differing properties to shared data objects

A system and method for executing sequential code in the context of a single-instruction, multiple-thread (SIMT) processor. In one embodiment, the system includes: (1) a pipeline control unit operable to create a group of counterpart threads of the sequential code, one of the counterpart threads being a master thread, remaining ones of the counterpart threads being slave threads and (2) lanes operable to: (2) execute certain instructions of the sequential code only in the master thread, corresponding instructions in the slave threads being predicated upon the certain instructions and (2) broadcast branch conditions in the master thread to the slave threads.

System and method for executing sequential code using a group of threads and single-instruction, multiple-thread processor incorporating the same

Basic blocks within a thread program are characterized for convergence based on variance analysis or corresponding instructions. Each basic block is marked as divergent based on transitive control dependence on a block that is either divergent or comprising a variant branch condition. Convergent basic blocks that are defined by invariant instructions are advantageously identified as candidates for scalarization by a thread program compiler.

Growing community of inventors

Santa Clara, CA, United States of America

Gautam Chakrabarti