CUDA view

The CUDA view of the Debug perspective displays information about CUDA kernels, blocks and threads currently running on the GPUs. This view is updated automatically when program stops.

CUDA View Modes

CUDA view has two primary modes that can be toggled with buttons on the view toolbar:

1. Kernels mode - tree with CUDA kernels, blocks and threads. This mode shows threads from every kernel are currently running on the device and their current state.
2. GPUs mode - tree shows state of GPUs, streaming multiprocessors (SM), warps and lanes. This mode makes is designed to provide information what is executing on the hardware.

Note: This view is empty if no CUDA context was created by the application being debugged. Most CUDA runtime calls will result in context creation.

Filtering CUDA View Contents

Contents of this view can be filtered by entering the filter condition to the field above the table. Following filter terms are recognized:

• kernel <num> - will only show elements that belong to a given kernel.
• block <blockIdx> - will only show elements that belong to a given block.
• thread <threadIdx> - will only show threads with given threadIdx. This term is the default so filter string "4,2,1" will show threads with given ID from all blocks.
• device <num> - will only show elements running on specific device.
• sm <num> - will only show elements running on specific SM.
• warp <num> - will only show warps with given number.
• lane <num> - will only show lanes with given number.

Thread and block idx can be specified in following forms:

• 7 matches all threads or blocks that have 7 as x coordinate - e.g. (7,0,0), (7,1,0), (7,2,0), (7,1,1)
• 4,2 matches all threads or blocks that have 4 as x coordinate and 2 as y coordinate - e.g. (4,2,0), (4,2,1), (4,2,2), (4,2,3)
• 1,2,3 matches threads from all blocks (or blocks from all kernels) with this coordinate.

Wildcard character (*) can be used for any dimension, e.g. string "*,2,1" matches (0,2,1), (1,2,1), (2,2,1), etc. Block and thread Idx can be enclosed into parentheses for readability.

Terms can be combined in a filter condition in any combination and all terms are supported in all modes. Some example filter strings are:

• kernel 1 block 2 thread 3 - will show all threads that have 3 for x coordinate for all blocks with 2 as x coordinate from kernel 1. In GPU mode, this will show corresponding devices/sms/warps/lanes.
• block *,2 lane 0 - will show threads that are running on first lanes of warps and belonging to blocks from all kernels with blockIdx matching the mask.
• block (2,1,1) 1 2 3 - shows threads that have 1, 2 or 3 as x coordinate and belong to block 2,1,1 from any kernel.

CUDA view toolbar actions

Icons	Name	Description
	Kernel mode	Switches the view into kernel mode displaying kernels, blocks and device threads.
	GPU mode	Enables GPU mode displaying devices, sms, warps and lanes.
	Pin thread/lane	Adds selected thread or lane to pinned threads list. See Navigating CUDA application state for more details.
	Link selection with the Debug view	When selection linking is enabled any selection changes made to this view are reflected in the Debug view. This in turn may update Variables , Expressions and Registers view. Selection changes in Debug view will also be reflected in this view.
	Collapse All	Collapses all the the currently expanded elements.
	View menu	Opens a view menu with following options: Open New View - opens a new CUDA view copy Coalesce Inactive Elements - will only show one tree node for a continuous range of inactive elements. This setting is enabled by default.

Debug information

Navigating CUDA application state

Debug view
C/C++ Views