Numba cuda 动态共享 memory：不止一种？

Question

I am aware that I can create a dynamic shared memory array for a numba-compiled CUDA kernel by passing the size in as the forth argument to the kernel call:

    ...
    foo_kernel[grid, block, stream, shared_bytes](...)
    ...

@cuda.jit
def foo_kernel(...) -> None:
    a = cuda.shared.array(0, nb.int32)

From here, I can slice a if I want to treat as several arrays.

However, what if I want to have two arrays of different dtypes? Can I do something like:

    ...
    a = cuda.shared.array(0, nb.int32)
    b = cuda.shared.array(0, nb.float32)
    ...

and then slice b so that I access values appropriately non-overlapping with a ?

Answer 1

Aha -- some googling finds: https://curiouscoding.nl/posts/numba-cuda-speedup/#v15-dynamic-shared-memory

Which confirms that different dtypes are indeed supported, using the trick I guessed (as pictured above).