Capturing commands in a multi-engine graphics processing unit

Abstract

The techniques and systems described herein are directed to capturing commands in a multi-engine graphics processing unit (GPU). Captured commands can be played back by a developer to optimize software, hardware, and drivers. To accurately capture commands and memory associated with the commands during execution, dependencies between command buffer segments associated with the various GPU engines may be determined and used to divide a command buffer segment into atomic elements (which may also be referred to as seglets). Command buffer segments are analyzed to identify synchronization commands, which may represent a point in a command buffer segment that relies on an operation to be completed in another command buffer segment. The command buffer segment can be recursively divided into seglets based on the synchronization commands. The resulting seglets represent command segments that, upon execution, operate without synchronization interference from other command buffer segments.