How to step out of a CUDA kernel using CUDA-GDB? - Stack Overflow

I am using CUDA-GDB to explore a large CUDA project. I can discover the first CUDA kernel using the following command:

set cuda break_on_launch application

I can then step-debug through the kernel, and GDB's continue will conveniently take me to the next kernel launch.

As explained here I can also disable these breaks, returning to the default behaviour:

set cuda break_on_launch none

I would like to exit the kernel, and return to the host code. The GDB finish command doesn't help here, reporting:

(cuda-gdb) finish
"finish" not meaningful in the outermost frame.

Is it possible to exit a CUDA kernel and return to the subsequent line of host code?

I am using CUDA-GDB to explore a large CUDA project. I can discover the first CUDA kernel using the following command:

set cuda break_on_launch application

I can then step-debug through the kernel, and GDB's continue will conveniently take me to the next kernel launch.

As explained here I can also disable these breaks, returning to the default behaviour:

set cuda break_on_launch none

I would like to exit the kernel, and return to the host code. The GDB finish command doesn't help here, reporting:

(cuda-gdb) finish
"finish" not meaningful in the outermost frame.

Is it possible to exit a CUDA kernel and return to the subsequent line of host code?

• cuda
• cuda-gdb

Share Improve this question Follow edited Jan 8 at 17:27 paleonix 3,13855 gold badges1717 silver badges3939 bronze badges asked Jan 8 at 13:11 user2023370user2023370 11.2k66 gold badges5858 silver badges9191 bronze badges 2

• set a breakpoint in host code, after a cudaDeviceSynchronize() call placed immediately after the kernel call. – Robert Crovella Commented Jan 9 at 2:54
• I don't know where the kernel calls are. I'm using break_on_launch as a way to find them, and especially find which ones are called. Once I know where they are, there isn't a problem. – user2023370 Commented Jan 9 at 9:43

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 1

It is possible to switch the focus to any host thread with the thread command the same way you would do in the original GDB, as mentioned in the docs:

[...] the GDB command to display the host threads and switch to host thread 1 are, respectively:

(cuda-gdb) info threads
(cuda-gdb) thread 1

As kernel launches are in general asynchronous, "subsequent line of host code" is not necessarily well defined. But you can set the environment variable CUDA_LAUNCH_BLOCKING to 1 to get synchronous operation between CPU and GPU which should help exploration (and debugging).