Check number and types of args passed to RawKernel function?

Hello CuPy developers, I noticed that if you pass the wrong number or types of input arguments to RawKernel that CuPy crashes. What do you think about adding some checks?

source = "extern "C" __global__ void test(float *arg1, float *arg2, const int n) {  } " 
test = cupy.RawKernel(source, "test")
grid=(1,)
block=(1,)

# 1. Too few input arguments
args = (arg1, arg2)
# Segmentation fault (CUDA_ERROR_ILLEGAL_ADDRESS)
# test(grid, block, args)

# 2. Too many input arguments
args = (arg1, arg2, arg2, n)

# 3. Wrong type for input argument #1 (0-based)
args = (arg1, n, arg2)

If you think it is a good idea, I would be happy to try to contribute a PR.

[mnicely]

@leofang No problem. I just need a few days to create an example to submit. I’m swamped with my new position!

[mnicely]

I suspect cuLaunchKernel doesn’t really check (or have no way to check) this.

We might need to file a bug. Having too many parameter should return a illegal address error