Nvidia Corporation

John Clifton Woolley, Jr

John Tran

Jaydeep Marathe

David Fontaine

Kyrylo Perelygin

Ashutosh Jain

Lukasz Krystian Ligowski

Sebastian Piotr Jodlowski

Maciej Marcin Piechotka

Magnus Strengert

Jitendra Pratap Singh Chauhan

Przemyslaw Tredak

Xiaonan Tian

In one embodiment of the present invention, a convolution engine configures a parallel processing pipeline to perform multi-convolution operations. More specifically, the convolution engine configures the parallel processing pipeline to independently generate and process individual image tiles. In operation, for each image tile, the pipeline calculates source locations included in an input image batch based on one or more start addresses and one or more offsets. Subsequently, the pipeline copies data from the source locations to the image tile. The pipeline then performs matrix multiplication operations between the image tile and a filter tile to generate a contribution of the image tile to an output matrix. To optimize the amount of memory used, the pipeline creates each image tile in shared memory as needed. Further, to optimize the throughput of the matrix multiplication operations, the values of the offsets are precomputed by a convolution preprocessor.

Indirectly accessing sample data to perform multi-convolution operations in a parallel processing system

Apparatuses, systems, and techniques to indicate contextual information to be used by available logical processors. In at least one embodiment, one or more circuits are to perform an application programming interface (API) to indicate a first set of contextual information to be used by a first subset of available processors.

Techniques for using contextual information

System and method for pseudo-random number generation based on a recursion with significantly increased multithreaded parallelism. A single pseudo-random generator program is assigned with multiple threads to process in parallel. N state elements indexed incrementally are arranged into a matrix comprising x rows, where a respective adjacent pair of state elements in a same column are related by g=(M+j)mod N, wherein j and g represent indexes of the pair of state elements. x can be determined through an modular manipulative inverse of M and N. The matrix can be divided into sections with each section having a number of columns, and each thread is assigned with a section. In this manner, the majority of the requisite interactions among the state elements occur without expensive inter-thread communications, and further each thread may only need to communicate with a single other thread for a small number of times.

Parallelization of random number generators

Growing community of inventors

San Jose, CA, United States of America

John Clifton Woolley, Jr