Parallel vs sequential for loop different results

Hello I’m using Python 3.10.2 and numba 0.55.1. I am trying to do the following operation:

@njit(parallel=True)
def sum_parallel(data, idxs):
    data_len = len(data)
    grid = np.zeros((128, 128), dtype=np.float64)
    for k in prange(data_len):
        i = idxs[1][k]
        j = idxs[0][k]
        grid[i][j] += data[k]
    return grid

def sum_sequential(data, idxs):
    data_len = len(data)
    grid = np.zeros((128, 128), dtype=np.float64)
    for k in prange(data_len):
        i = idxs[1][k]
        j = idxs[0][k]
        grid[i][j] += data[k]
    return grid

data = np.random.random((8000))
idxs = np.random.randint(size=(2, 8000), low=0, high=128)

grid_parallel = sum_parallel(data,idxs)
grid_sequential = sum_sequential(data, idxs)
std_difference = np.std(grid_parallel-grid_sequential)
print(std_difference)
0.014976131305802937

I guess there is a race condition here, although I am not sure if this is a bug or if it is something that I have not taken into consideration when writing my parallel function, is there a way to provide a critical section?

[guilhermeleobas]

Hi @miguelcarcamov, thanks for the report. I can reproduce the error when parallel is enabled. Maybe @stuartarchibald knows a workaround to this issue.

Related discussion: https://numba.discourse.group/t/how-do-i-use-thread-locks-in-nopython-mode/1258 https://numba.discourse.group/t/how-to-use-locks-in-nopython-mode/221

[stuartarchibald]

Closing this question as it seems to be resolved. Numba now has a discourse forum https://numba.discourse.group/ which is great for questions like this, please do consider posting there in future 😃 Thanks!