Error handling in CUDA kernels

Specific Domains GPU
,
1

In my CUDA kernel I check for a specific condition and would like to terminate the execution if the condition is fulfilled. What is the proper way to do so? How can I throw an error from inside a CUDA kernel?

1 Like
2

You can just call error("") but the user will only see the specific error when they run with julia -g2

1 Like
3

Thank you. It works.

In case of normal (non debug run) the thrown error is accompanied by continuously repeating message

ERROR: a exception was thrown during kernel execution.
       Run Julia on debug level 2 for device stack traces.

I guess there is a typo in this message: should be “an exception” instead of “a exception”.

4

The proper solution is to restructure your control flow such that you can do an early return. One alternative is to use the exit PTX instruction using LLVM.jl’s @asmcall, but not every GPU supports that instruction, and we’ve encountered miscompilations in the presence of such control flow.

5

I can do the early return. But how I can signal that the return was triggered by an error? In normal functions I can return some variable as an error code, but CUDA kernels do not allow to return anything.

6

You can allocate a global flag and write to it from your kernel. This can be a single-element CuArray you pass as an argument, or something fancier (e.g., look for exception_flag in CUDA.jl, that’s a global flag allocated in CPU memory mapped into GPU address space so that we can more easily read the value without synchronizing the GPU).

1 Like
7

Thank you. I will try.