Numba cuda.jit assertion fail silently, code still runs, but results are wrong

Numba Community Support
1

Hello, all.

I have the following code where the assertion is wrong but the code will still run as usual. Finally, it gives an incorrect result.

It took me a long time to identify the problem and after removing the assert line the code is working fine.

I’m wondering if failing the assertion the code would still run and give a random result is normal behavior for numba cuda? It seems that assert support doesn’t make any sense.

@cuda.jit()
def add_tile(a, b):
    '''
    a: [tile_size, tile_size, 3]
    b: [tile_size, tile_size, 3]
    '''
    TILE_SIZE = 16
    # This assertion is wrong, I forgot that the first dimension is the block dimension
    # But the code still runs, just the result is wrong
    assert a.shape[0] == TILE_SIZE

    tile_ix, tile_iy = cuda.threadIdx.x, cuda.threadIdx.y
    tile_id = cuda.blockIdx.x

    for i in range(3):
        a[tile_id, tile_ix, tile_iy, i] += b[tile_id, tile_ix, tile_iy, i]

# some random test
blocks = 32
threads = (TILE_SIZE, TILE_SIZE)

a = np.random.randn(blocks, threads[0], threads[1], 3).astype(np.float32)
b = np.random.randn(blocks, threads[0], threads[1], 3).astype(np.float32)
print((a + b)[1, 1, 1, :])

device_a = cuda.to_device(a)
device_b = cuda.to_device(b)

add_tile[blocks, threads](device_a, device_b)
print(device_a.copy_to_host()[1, 1, 1, :])