Python和LabVIEW計算性能對比

為了對比python和LabVIEW的性能，決定跑一個計算平面上1000個點的最大距離的演算法，測試對象是一個1000行的文本文件：

為了避免文件讀取造成的誤差，時間戳只取計算部分

LabVIEW結果

Importnumpyasnpnimporttimeastmnndocument=open("testspark-points-2.txt")nlines=document.read().split("n")npointsArr=[]nforlineinlines:nargs=line.split(",")niflen(args)==2:nx=float(args[0])ny=float(args[1])npointsArr.append(np.array([x,y]))npointsNum=len(pointsArr)ntempresults=[]nstartTime=tm.time()nfori_ainrange(pointsNum):nfori_binrange(pointsNum):nifi_a<i_b:nvectA=pointsArr[i_a]nvectB=pointsArr[i_b]ntempresults.append(np.sqrt(np.sum(np.square(vectA-vectB))))ntempresultsNP=np.array(tempresults)nmaxResult=np.max(tempresultsNP)nendTime=tm.time()nprint(endTime-startTime)nprint(maxResult)n

2.902729034423828

13.8759846889

Process finished with exit code 0

懷疑是拼接數組的鍋

importnumpyasnpnimporttimeastmnnndocument=open("testspark-points-2.txt")nlines=document.read().split("n")npointsArr=[]nforlineinlines:nargs=line.split(",")niflen(args)==2:nx=float(args[0])ny=float(args[1])npointsArr.append(np.array([x,y]))npointsNum=len(pointsArr)nbuffer=np.full((pointsNum,pointsNum),0.)nstartTime=tm.time()nfori_ainrange(pointsNum):nfori_binrange(pointsNum):nifi_a<i_b:nvectA=pointsArr[i_a]nvectB=pointsArr[i_b]nbuffer[i_a][i_b]=np.sqrt(np.sum(np.square(vectA-vectB)))ntempresultsNP=np.array(buffer)nmaxResult=np.max(tempresultsNP)nendTime=tm.time()nprint(endTime-startTime)nprint(maxResult)n

3.0922529697418213

13.8759846889

Process finished with exit code 0

發現替換數組元素的方法運行速度更慢了，換數學庫直接運算：

importnumpyasnpnimporttimeastmnimportmathnndocument=open("testspark-points-2.txt")nlines=document.read().split("n")npointsArr=[]nforlineinlines:nargs=line.split(",")niflen(args)==2:nx=float(args[0])ny=float(args[1])npointsArr.append([x,y])nnpointsNum=len(pointsArr)ntempresults=[]nstartTime=tm.time()nfori_ainrange(pointsNum):nfori_binrange(pointsNum):nifi_a<i_b:ntempresults.append(math.sqrt(nmath.pow((pointsArr[i_a][0]-pointsArr[i_b][0]),2)+nmath.pow((pointsArr[i_a][1]-pointsArr[i_b][1]),2)))nntempresultsNP=np.array(tempresults)nmaxResult=np.max(tempresultsNP)nendTime=tm.time()nprint(endTime-startTime)nprint(maxResult)n

0.5273771286010742

13.8759846889

Process finished with exit code 0

這下就只有一個數量級的差距了，也許下次應該試試別的演算法。

結果如下：

Python

LabVIEW

0.527

0.014

不知道Python還能怎麼優化一下？

用numba優化Python代碼執行速度

聽說python也可以抱LLVM大腿了，把這個演算法用numba改造了一下

importnumpyasnpnimporttimeastmnimportmathnfromnumbaimportjitnn@jitndefmax_distance(pointsArr):nnpointsNum=len(pointsArr)ntempresults=[]nfori_ainrange(pointsNum):nfori_binrange(pointsNum):nifi_a<i_b:ntempresults.append(math.sqrt(nmath.pow((pointsArr[i_a][0]-pointsArr[i_b][0]),2)+nmath.pow((pointsArr[i_a][1]-pointsArr[i_b][1]),2)))nntempresultsNP=np.array(tempresults)nmaxResult=np.max(tempresultsNP)nreturnmaxResultnnndocument=open("testspark-points-2.txt")nlines=document.read().split("n")npointsArr=[]nforlineinlines:nargs=line.split(",")niflen(args)==2:nx=float(args[0])ny=float(args[1])npointsArr.append([x,y])npointsArr2=np.array(pointsArr)nnnstartTime=tm.time()nnnmaxResult=max_distance(pointsArr2)nnendTime=tm.time()nprint(endTime-startTime)nprint(maxResult)n

Math-numba.py

0.3585703372955322