有没有办法在numerical python代码中最小化对__new__的调用次数？

Question

我正在努力提高一些用python编写的数值建模代码的速度。它依赖于numpy数组和scipy.linalg的矩阵求解方法。这是一个有限差分PDE求解器。在分析和处理一些较慢的比特之后，built-in method __new__占用了最大的时间块。是否每次声明数字时都会执行__new__方法？有什么方法可以避免它吗？以下是求解器的主循环：

    #iterate over time steps
for j in range(1,N_t):

    #store some numbers
    jm1 = j - 1
    mm2 = m - 2
    mm1 = m - 1
    mp1 = m + 1
    pm1 = p - 1
    pm2 = p - 2

    #load the "b vector"
    vec[0] = v1*C[0,jm1] + v2*C[1,jm1]
    for i in r_vec_a:
        vec[i] = hlam_a*C[i-1,jm1] + v3*C[i,jm1] + hlam_a*C[i+1,jm1]
    vec[mm1] = -f1*C[mm2,jm1] + f5*C[mm1,jm1] - f3*C[m,jm1] - f4*C[mp1,jm1]
    vec[m] = -g1*C[mm2,jm1] - g2*C[mm1,jm1] + g5*C[m,jm1] - g4*C[mp1,jm1]
    for i in r_vec_b:
        vec[i] = hlam_b*C[i-1,jm1] + v4*C[i,jm1] + hlam_b*C[i+1,jm1]
    vec[pm1] = v5*C[pm2,jm1] + v6*C[pm1,jm1]

    #solve the matrix equation for new concentrations
    C[:,j] = scipy.linalg.solve_banded(u1, banded, vec, check_finite = u2,
        overwrite_b = u3)

    #compute boundary values (with 2 extra orders of accuracy because it's easy)
    C_x0[0,j] = w1*C[0,j] - w2*C[1,j] + w3*C[2,j] - w4*C[3,j]
    C_xL[0,j] = e1*C[mm2,j] + e2*C[mm1,j] + e3*C[m,j] + e4*C[mp1,j]
    C_yL[0,j] = K_r*C_xL[0,j]
    C_y0[0,j] = -w4*C[p-4,j] + w3*C[p-3,j] - w2*C[pm2,j] + w1*C[pm1,j]

我的第一个想法是存储所有索引变量和常量，如果它们被重用，但这并没有减少对__new__的调用。