from numba import njit, prange
import numpy as np

@njit(parallel=True)
def nanmean_lowlevel(data, indices, indptr, shape):
    N, M = shape
    sums = np.zeros(N, dtype=np.float64)
    nans = np.zeros(N, dtype=np.int64)
    for i in prange(N):
        start = indptr[i]
        stop = indptr[i+1]
        window = data[start:stop]
        n_nan = np.int64(0)
        i_sum = np.float64(0.)
        for j_val in window:
            if np.isnan(j_val):
                n_nan += 1
            else:
                i_sum += j_val
        sums[i] = i_sum
        nans[i] = n_nan
    sums /= (M - nans)
    return sums