Ubuntu Manpages

cuda-gdb

The NVIDIA CUDA Debugger cuda-gdbserver - The NVIDIA CUDA-GDB server for remote debugging

cuda-gdb [gdb-options] [--cuda-use-lockfile VALUE] [prog [core|procID]]

cuda-gdb [gdb-options] [--cuda-use-lockfile VALUE] --args prog [arguments]

cuda-gdb is the NVIDIA CUDA debugger for Unix platforms. It is gdb with extra commands specific to CUDA debugging.

All the existing gdb commands are supported. On top of those commands, the following CUDA-specific commands are also supported.

Commands to query information about CUDA activities:

information about all the devices
information about all the SMs in the current device
information about all the warps in the current SM
information about all the lanes in the current warp
information about all the active kernels
information about all the contexts
information about all the active blocks in the current kernel
information about all the active threads in the current kernel
information about the parent kernels of the kernel in focus
information about the kernels launched by the kernels in focus

Commands to query or switch the focus of the debugger:

Print or select the current CUDA block
Print or select the current CUDA device
Print or select the current CUDA grid
Print or select the current CUDA kernel
Print or select the current CUDA lane
Print or select the current CUDA SM
Print or select the current CUDA thread
Print or select the current CUDA warp

Commands to configure the CUDA-specific commands:

Set the api_failures to ignore/stop/hide on CUDA driver API call errors
Automatically set a breakpoint at the entrance of kernels
Turn on/off coalescing of the CUDA commands output
Turn on/off context events (push/pop/create/destroy) output messages
Turn on/off deferral of kernel launch messages
Choose whether to disassemble from the device memory (slow) or the ELF image (fast)
Turn on/off GPU busy check the next time the inferior application is run
Set hiding of the internal CUDA frames when printing the call stack
Turn on/off kernel events (launch/termination) output messages
Turn on/off CUDA kernel launch blocking (effective starting from the next run)
Turn on/off CUDA Memory Checker next time the inferior application is run
Thread to notify about CUDA events when no other known candidate
Turn on/off CUDA software preemption debugging the next time the inferior application is run
Set the automatic thread selection policy to use when the current thread cannot be selected

CUDA-GDB supports the same options as GDB, with the addition of the following options:

If VALUE is 0, CUDA-GDB will not create a lock file and will not enforce the exclusive usage of GPUs.

If non-null, the full pathname for an alternate temporary directory. Note that the debugger must have write and exec permission to that directory.

~/.cuda-gdbinit
Per user configuration file. The file format is identical to ~/.gdbinit. See gdb(5) for further details.
/tmp/cuda-dbg/
Temporary directory where the debugger stores intermediate files. The directory must have write and and exec permissions. Use $TMPDIR to override the location.

For more information about cuda-gdb, please see the online documentation at http://docs.nvidia.com/cuda/cuda-gdb/index.html.

For more information about CUDA-related features, type help cuda or help info cuda from within CUDA-GDB, or consult the CUDA-GDB online manual.

cuda-binaries(1), cuda-libraries(7), gdb(1)

Email bugs to cuda-debugger-bugs@nvidia.com

©2013 NVIDIA Corporation. All rights reserved.