.NET array is slower than list in IronPython?

I did the following matrix multiplication benchmark in IronPython based on code here:

from System import Random
from System.Diagnostics import Stopwatch

def zero(m,n):
    # Create zero matrix
    new_matrix = [[0 for row in range(n)] for col in range(m)]
    return new_matrix

def rand(m,n):
    # Create random matrix
    rnd = Random(1)
    new_matrix = [[rnd.NextDouble() for row in range(n)] for col in range(m)]
    return new_matrix

def show(matrix):
    # Print out matrix
    for col in matrix:
        print col 

def mult(matrix1,matrix2):
    # Matrix multiplication
    if len(matrix1[0]) != len(matrix2):
        # Check matrix dimensions
        print 'Matrices must be m*n and n*p to multiply!'
    else:
        # Multiply if correct dimensions
        watch = Stopwatch()
        print 'mult1 start....'
        watch.Start()
        new_matrix = zero(len(matrix1),len(matrix2[0]))
        for i in range(len(matrix1)):
            for j in range(len(matrix2[0])):
                for k in range(len(matrix2)):
                    new_matrix[i][j] += matrix1[i][k]*matrix2[k][j]
        watch.Stop()
        print 'mult1 end.'
        print watch.ElapsedMilliseconds
        return new_matrix

from System import Array

def ListToArray(matrix):
    n = len(matrix)
    m = len(matrix[0])
    a = Array.CreateInstance(float, n, m)
    for i in range(n):
        for j in range(m):
            a[i,j] = matrix[i][j]
    return a


def mult2(matrix1, matrix2):

    N = len(matrix1)
    K = len(matrix2)
    M = len(matrix2[0])

    m1 = ListToArray(matrix1)
    m2 = ListToArray(matrix2)
    res = ListToArray(rand(len(matrix1), len(matrix2[0])))

    watch = Stopwatch()
    print 'mult2 start...'
    watch.Start()
    for i in range(N):
        for j in range(M):
            for k in range(K):
                res[i,j] += m1[i,k]*m开发者_Go百科2[k,j]
    watch.Stop()
    print 'mult2 ends.'
    print watch.ElapsedMilliseconds
    return res


if __name__ == '__main__':
    #a = rand(280,10304)
    #b = rand(10304,280)

    a = rand(280,10)
    b = rand(10,280)

    c = mult2(a, b)
    d = mult(a, b)

I want to try two big matrices (280 by 10304 and 10304 by 208), but both versions cannot produce a result within a short time.

Then I tried a much smaller one (as show in the code), I the result is following:

mult2 : 7902 ms
mult1 : 420 ms

indicating that using .NET array in IronPython is much slower than the python List.

Also notice that C# uses about ~12 seconds for the two big matrices. IronPython already spends a lot of them on the 10K times smaller case. I am not sure whether the IronPython setting in my computer is wrong or not, if not, IronPython is really slow for numerical code.

---

For your particular question, I think the problem is boxing - in IronPython, list items (and all other variables) are stored boxed, so only boxed values are operated on. CLR Array elements are not boxed, however, and thus IronPython will have to box them when they are extracted from the array, and then unbox them on the way back in. C# can operate on the unboxed values, and has a bunch of other optimizations to make arrays fast that IronPython doesn't have.

If you want fast numerical math, NumPy for IronPython might be a better bet.

---

In our projects we are trying to avoid .Net classes usage until this become really necessary. I guess that there is no any particular need to use Array for matrix multiplication since python gives you several ways to handle matrices like list, collections or numpy.

---

You're likely to get slow speeds one way or another when using a high level language to process matrices. If you want to do it in .net, try http://numerics.mathdotnet.com/ - they've done what's possible to improve matrix operations already.

It's not really an answer to why you see the difference, but might help if you're looking for some fast solution.

---

My guess is what's killing the performance vs. normal Python lists is the use of a multi-dimensional array. I could easily see this being a slow path and that if you switched to arrays of arrays or calculating the index yourself you could see a big speed up.

Others have pointed out that this isn't the best code for a dynamic language like Python. But one of the great benefits of IronPython is how easy and seamless it is to pull in C# snippets to speed up cases like this.