Nvidia Corporation

Daniel Robert Johnson

Stephen William Keckler

Mark William Stephenson

Sana Damani

Jack H Choquette

James Michael O'Connor

Michael Patrick McKeown

William James Dally

John Erik Lindholm

Olivier Giroux

Niladrish Chatterjee

Iuri Frosio

Yun-Ta Tsai

David Tarjan

Mark Alan Gebhart

Ram Rangan

Rishkul Kulkarni

Chih-Chi Cheng

Jason Lavar Clemons

Oliver Giroux

Minsoo Rhu

Warp sharding techniques to switch execution between divergent shards on instructions that trigger a long stall, thereby interleaving execution between diverged threads within a warp instead of across warps. The technique may be applied to mitigate pipeline stalls in applications with low warp occupancy and high divergence. Warp data cache locality may also be improved by concentrating memory accesses within a warp rather than spreading them across warps.

Techniques for divergent thread group execution scheduling

Convergence of threads executing common code sections is facilitated using instructions inserted at strategic locations in computer code sections. The inserted instructions enable the threads in a warp or other group to cooperate with a thread scheduler to promote thread convergence.

Convergence among concurrently executing threads

A method and system for a DRAM having a first bank that includes a first sub-array (SA) and a second SA. The first SA includes a first storage unit coupled to a first row-buffer in a first sub-channel (FSC) and a second storage unit in a second sub-channel (SSC). The second SA includes a third storage unit and a fourth storage unit coupled to a second row-buffer. The first SA is associated with a first row address (RA) and the FSC is associated with a first column address (CA) stored in the FSC. The second SA is associated with a second RA and the SSC is associated with a second CA stored in the SSC. The first and second CAs are used to select portions of data from the first and second row-buffers, respectively, for output to a data bus.

Systems and methods for dynamic random access memory (DRAM) sub-channels

A patch memory system for accessing patches from a memory is disclosed. A patch is an abstraction that refers to a contiguous, array of data that is a subset of an N-dimensional array of data. The patch memory system includes a tile cache, and is configured to fetch data associated with a patch by determining one or more tiles associated with an N-dimensional array of data corresponding to the patch, and loading data for the one or more tiles from the memory into the tile cache. The N-dimensional array of data may be a two-dimensional (2D) digital image comprising a plurality of pixels. A patch of the 2D digital image may refer to a 2D subset of the image.

Patch memory system

A system, method, and computer program product are provided for providing prioritized access for multithreaded processing. The method includes the steps of allocating threads to process a workload and assigning a set of priority tokens to at least a portion of the threads. Access to a resource, by each one of the threads, is based on the priority token assigned to the thread and the threads are executed by a multithreaded processor to process the workload.

Cache utilization and eviction based on allocated priority tokens

One embodiment of the present invention sets forth a technique for scheduling thread execution in a multi-threaded processing environment. A two-level scheduler maintains a small set of active threads called strands to hide function unit pipeline latency and local memory access latency. The strands are a sub-set of a larger set of pending threads that is also maintained by the two-leveler scheduler. Pending threads are promoted to strands and strands are demoted to pending threads based on latency characteristics. The two-level scheduler selects strands for execution based on strand state. The longer latency of the pending threads is hidden by selecting strands for execution. When the latency for a pending thread is expired, the pending thread may be promoted to a strand and begin (or resume) execution. When a strand encounters a latency event, the strand may be demoted to a pending thread while the latency is incurred.

Two-level scheduler for multi-threaded processing

Growing community of inventors

Austin, TX, United States of America

Daniel Robert Johnson