Any faster way of copying arrays in C#?

问题:

I have three arrays that need to be combined in one three-dimension array. The following code shows slow performance in Performance Explorer. Is there a faster solution?

for (int i = 0; i < sortedIndex.Length; i++) {
    if (i < num_in_left)
    {    
        // add instance to the left child
        leftnode[i, 0] = sortedIndex[i];
        leftnode[i, 1] = sortedInstances[i];
        leftnode[i, 2] = sortedLabels[i];
    }
    else
    { 
        // add instance to the right child
        rightnode[i-num_in_left, 0] = sortedIndex[i];
        rightnode[i-num_in_left, 1] = sortedInstances[i];
        rightnode[i-num_in_left, 2] = sortedLabels[i];
    }                    
}

Update:

I'm actually trying to do the following:

//given three 1d arrays
double[] sortedIndex, sortedInstances, sortedLabels;
// copy them over to a 3d array (forget about the rightnode for now)
double[] leftnode = new double[sortedIndex.Length, 3];
// some magic happens here so that
leftnode = {sortedIndex, sortedInstances, sortedLabels};

回答1:

Use Buffer.BlockCopy. Its entire purpose is to perform fast (see Buffer):

This class provides better performance for manipulating primitive types than similar methods in the System.Array class.

Admittedly, I haven't done any benchmarks, but that's the documentation. It also works on multidimensional arrays; just make sure that you're always specifying how many bytes to copy, not how many elements, and also that you're working on a primitive array.

Also, I have not tested this, but you might be able to squeeze a bit more performance out of the system if you bind a delegate to System.Buffer.memcpyimpl and call that directly. The signature is:

internal static unsafe void memcpyimpl(byte* src, byte* dest, int len)

It does require pointers, but I believe it's optimized for the highest speed possible, and so I don't think there's any way to get faster than that, even if you had assembly at hand.

---

Update:

Due to requests (and to satisfy my curiosity), I tested this:

using System;
using System.Diagnostics;
using System.Reflection;

unsafe delegate void MemCpyImpl(byte* src, byte* dest, int len);

static class Temp
{
    //There really should be a generic CreateDelegate<T>() method... -___-
    static MemCpyImpl memcpyimpl = (MemCpyImpl)Delegate.CreateDelegate(
        typeof(MemCpyImpl), typeof(Buffer).GetMethod("memcpyimpl",
            BindingFlags.Static | BindingFlags.NonPublic));
    const int COUNT = 32, SIZE = 32 << 20;

    //Use different buffers to help avoid CPU cache effects
    static byte[]
        aSource = new byte[SIZE], aTarget = new byte[SIZE],
        bSource = new byte[SIZE], bTarget = new byte[SIZE],
        cSource = new byte[SIZE], cTarget = new byte[SIZE];


    static unsafe void TestUnsafe()
    {
        Stopwatch sw = Stopwatch.StartNew();
        fixed (byte* pSrc = aSource)
        fixed (byte* pDest = aTarget)
            for (int i = 0; i < COUNT; i++)
                memcpyimpl(pSrc, pDest, SIZE);
        sw.Stop();
        Console.WriteLine("Buffer.memcpyimpl: {0:N0} ticks", sw.ElapsedTicks);
    }

    static void TestBlockCopy()
    {
        Stopwatch sw = Stopwatch.StartNew();
        sw.Start();
        for (int i = 0; i < COUNT; i++)
            Buffer.BlockCopy(bSource, 0, bTarget, 0, SIZE);
        sw.Stop();
        Console.WriteLine("Buffer.BlockCopy: {0:N0} ticks",
            sw.ElapsedTicks);
    }

    static void TestArrayCopy()
    {
        Stopwatch sw = Stopwatch.StartNew();
        sw.Start();
        for (int i = 0; i < COUNT; i++)
            Array.Copy(cSource, 0, cTarget, 0, SIZE);
        sw.Stop();
        Console.WriteLine("Array.Copy: {0:N0} ticks", sw.ElapsedTicks);
    }

    static void Main(string[] args)
    {
        for (int i = 0; i < 10; i++)
        {
            TestArrayCopy();
            TestBlockCopy();
            TestUnsafe();
            Console.WriteLine();
        }
    }
}

The results:

Buffer.BlockCopy: 469,151 ticks
Array.Copy: 469,972 ticks
Buffer.memcpyimpl: 496,541 ticks

Buffer.BlockCopy: 421,011 ticks
Array.Copy: 430,694 ticks
Buffer.memcpyimpl: 410,933 ticks

Buffer.BlockCopy: 425,112 ticks
Array.Copy: 420,839 ticks
Buffer.memcpyimpl: 411,520 ticks

Buffer.BlockCopy: 424,329 ticks
Array.Copy: 420,288 ticks
Buffer.memcpyimpl: 405,598 ticks

Buffer.BlockCopy: 422,410 ticks
Array.Copy: 427,826 ticks
Buffer.memcpyimpl: 414,394 ticks

Now change the order:

Array.Copy: 419,750 ticks
Buffer.memcpyimpl: 408,919 ticks
Buffer.BlockCopy: 419,774 ticks

Array.Copy: 430,529 ticks
Buffer.memcpyimpl: 412,148 ticks
Buffer.BlockCopy: 424,900 ticks

Array.Copy: 424,706 ticks
Buffer.memcpyimpl: 427,861 ticks
Buffer.BlockCopy: 421,929 ticks

Array.Copy: 420,556 ticks
Buffer.memcpyimpl: 421,541 ticks
Buffer.BlockCopy: 436,430 ticks

Array.Copy: 435,297 ticks
Buffer.memcpyimpl: 432,505 ticks
Buffer.BlockCopy: 441,493 ticks

Now change the order again:

Buffer.memcpyimpl: 430,874 ticks
Buffer.BlockCopy: 429,730 ticks
Array.Copy: 432,746 ticks

Buffer.memcpyimpl: 415,943 ticks
Buffer.BlockCopy: 423,809 ticks
Array.Copy: 428,703 ticks

Buffer.memcpyimpl: 421,270 ticks
Buffer.BlockCopy: 428,262 ticks
Array.Copy: 434,940 ticks

Buffer.memcpyimpl: 423,506 ticks
Buffer.BlockCopy: 427,220 ticks
Array.Copy: 431,606 ticks

Buffer.memcpyimpl: 422,900 ticks
Buffer.BlockCopy: 439,280 ticks
Array.Copy: 432,649 ticks

or, in other words: they're very competitive; as a general rule, memcpyimpl is fastest, but it's not necessarily worth worrying about.

回答2:

You can use Array.Copy.

EDIT

Array.Copy does work for multidimensional arrays: see this topic.

回答3:

For primitive type arrays (like double) you can copy fast, even for multidimensional array with pointers.

In the code below I initialize a 2D array A[10,10] with the values 1 through 100. Then I copy these values into a 1D array B[100]

unsafe class Program
{ 
    static void Main(string[] args)
    {
        double[,] A = new double[10, 10];

        for(int i = 0; i < 10; i++)
        {
            for(int j = 0; j < 10; j++)
            {
                A[i, j] = 10 * i + j + 1;
            }
        }
        // A has { { 1 ,2 .. 10}, { 11, 12 .. 20}, .. { .. 99, 100} }
        double[] B = new double[10 * 10];

        if (A.Length == B.Length)
        {
            fixed (double* pA = A, pB = B)
            {
                for(int i = 0; i < B.Length; i++)
                {
                    pB[i] = pA[i];
                }
            }
            // B has {1, 2, 3, 4 .. 100}
        }
    }
}

How fast is it. My testing has shown it to be many times faster then native C# copy and Buffer.BlockCopy(). You try it for your case and let us know.

Edit 1 I compared copying with four methods. 1) Two Nested loops, 2) One Serial loop, 3) Pointers, 4) BlockCopy. I measured the # of copies per tick for various size arrays.

N =   10x  10 (cpy/tck) Nested = 50,  Serial = 33, Pointer =    100, Buffer =    16
N =   20x  20 (cpy/tck) Nested = 133, Serial = 40, Pointer =    400, Buffer =   400
N =   50x  50 (cpy/tck) Nested = 104, Serial = 40, Pointer =   2500, Buffer =  2500
N =  100x 100 (cpy/tck) Nested = 61,  Serial = 41, Pointer =  10000, Buffer =  3333
N =  200x 200 (cpy/tck) Nested = 84,  Serial = 41, Pointer =  40000, Buffer =  2666
N =  500x 500 (cpy/tck) Nested = 69,  Serial = 41, Pointer = 125000, Buffer =  2840
N = 1000x1000 (cpy/tck) Nested = 33,  Serial = 45, Pointer = 142857, Buffer =  1890
N = 2000x2000 (cpy/tck) Nested = 30,  Serial = 43, Pointer = 266666, Buffer =  1826
N = 5000x5000 (cpy/tck) Nested = 21,  Serial = 42, Pointer = 735294, Buffer =  1712

It is clear here who is the winner. Pointer copy is orders of magnitudes better than any other method.

Edit 2 Apparently I was unfairly taking advantage of a compiler/JIT optimization because when I moved the loops behind delegates to equalize the playing field the numbers changed dramatically.

N =   10x  10 (cpy/tck) Nested =  0, Serial =  0, Pointer =      0, Buffer =     0
N =   20x  20 (cpy/tck) Nested = 80, Serial = 14, Pointer =    100, Buffer =   133
N =   50x  50 (cpy/tck) Nested =147, Serial = 15, Pointer =    277, Buffer =  2500
N =  100x 100 (cpy/tck) Nested = 98, Serial = 15, Pointer =    285, Buffer =  3333
N =  200x 200 (cpy/tck) Nested =106, Serial = 15, Pointer =    272, Buffer =  3076
N =  500x 500 (cpy/tck) Nested =106, Serial = 15, Pointer =    276, Buffer =  3125
N = 1000x1000 (cpy/tck) Nested =101, Serial = 11, Pointer =    199, Buffer =  1396
N = 2000x2000 (cpy/tck) Nested =105, Serial =  9, Pointer =    186, Buffer =  1804
N = 5000x5000 (cpy/tck) Nested =102, Serial =  8, Pointer =    170, Buffer =  1673

The buffered copy is top here (thanks to @Mehrdad) with pointer copy second. The question now is why isn't pointer copy as fast as buffer methods?