时间戳

首页 > 解决方案 > Numba 无法并行化循环

python - Numba 无法并行化循环

问题描述

以下代码运行，但 prange 循环无法并行化，尽管它显然是可并行化的：


import numpy as np
from numba import njit, prange


@njit(parallel=True)
def f1(money, u, v, cost_u, cost_v):

    # task: find index of u iu_opt and index of v iv_opt 
    # such that u[iu_opt] + v[iv_opt] is maximal subject
    # to cost_u[iu_opt] + cost_v[iv_opt] < money


    na = money.size

    ncu = cost_u.size
    ncv = cost_v.size

    iu_opt = np.empty((na,),dtype=np.int16)
    iv_opt = np.empty((na,),dtype=np.int16)

    for ia in prange(na):


        money_i = money[ia]
        ivbest = 0 # initially pick iv = 0

        # find max iu corresponding to iv = 0
        for iu in range(ncu-1,-1,-1):
            if cost_u[iu] + cost_v[0] < money_i: break

        iubest = iu
        # compute initial score
        score_best = u[iu] + v[0]  


        # then try to increase iv
        for iv in range(1,ncv):

            # it not enough money for u_0
            if cost_v[iv]  + cost_u[0] > money_i: break 

            while cost_v[iv] + cost_u[iu] > money_i:
                iu -= 1

            assert iu >= 0            

            score_now = u[iu] + v[iv] 


            if score_now > score_best:
                ivbest = iv
                iubest = iu
                score_best = score_now

        iu_opt[ia] = iubest
        iv_opt[ia] = ivbest

    return iu_opt, iv_opt



na = 50
ncu = 204
ncv = 205


money = np.arange(na)/(na)


cost_u = np.arange(ncu)/ncu
u = np.cumsum(np.random.random_sample(ncu))
cost_v = np.arange(ncv)/ncv
v = np.cumsum(np.random.random_sample(ncv))

iu, iv = f1(money, u, v, cost_u, cost_v)
f1.parallel_diagnostics(level=4)

如果这有帮助，那么问题的设置如下：对于每个值 money[ ia] 我们想要找到 i 和 j 最大化 u[i] + v[j]，因为 cost_u[i] + cost_v[j] 不能超过 money[ia]。

标签： pythonnumpyparallel-processingnumba

解决方案

如果有人遇到过类似的问题，我最终通过将大循环的内部拆分为另一个 njit 函数来使其工作。这是代码：

import numpy as np
from numba import njit, prange


@njit(parallel=True)
def f1(money, u, v, cost_u, cost_v):

    # task: find index of u iu_opt and index of v iv_opt 
    # such that u[iu_opt] + v[iv_opt] is maximal subject
    # to cost_u[iu_opt] + cost_v[iv_opt] < money


    na = money.size

    iu_opt = np.empty((na,),dtype=np.int16)
    iv_opt = np.empty((na,),dtype=np.int16)

    for ia in prange(na):

        money_i = money[ia]

        iubest, ivbest = f1_int(money_i,u,v,cost_u,cost_v)

        iu_opt[ia] = iubest
        iv_opt[ia] = ivbest

    return iu_opt, iv_opt


@njit
def f1_int(money_i,u,v,cost_u,cost_v):
    ivbest = 0 # initially pick iv = 0

    ncu = cost_u.size
    ncv = cost_v.size

    # find max iu corresponding to iv = 0
    for iu in range(ncu-1,-1,-1):
        if cost_u[iu] + cost_v[0] < money_i: break

    iubest = iu
    # compute initial score
    score_best = u[iu] + v[0]  


    # then try to increase iv
    for iv in range(1,ncv):

        # it not enough money for u_0
        if cost_v[iv]  + cost_u[0] > money_i: break 

        while cost_v[iv] + cost_u[iu] > money_i:
            iu -= 1

        assert iu >= 0            

        score_now = u[iu] + v[iv] 


        if score_now > score_best:
            ivbest = iv
            iubest = iu
            score_best = score_now
    return iubest, ivbest

na = 50
ncu = 204
ncv = 205


money = np.arange(na)/(na)


cost_u = np.arange(ncu)/ncu
u = np.cumsum(np.random.random_sample(ncu))
cost_v = np.arange(ncv)/ncv
v = np.cumsum(np.random.random_sample(ncv))

iu, iv = f1(money, u, v, cost_u, cost_v)
f1.parallel_diagnostics(level=4)

这并不能真正回答为什么会发生问题的问题，但会以某种方式起作用。

推荐阅读