我有三个阵列需要在一个三维数组中组合。以下代码显示Performance Explorer中的性能降低。有更快的解决方案吗?
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];
}
}
更新
我实际上是在尝试执行以下操作:
//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};
答案 0 :(得分:71)
使用Buffer.BlockCopy。它的全部目的是快速执行(见Buffer):
与System.Array类中的类似方法相比,此类为操作基元类型提供了更好的性能。
不可否认,我没有做任何基准测试,但那是文档。它也适用于多维数组;只需确保您始终指定要复制的字节的数量,而不是指定的元素数量,以及您正在处理基本数组。
此外,我还没有对此进行过测试,但是如果将委托绑定到System.Buffer.memcpyimpl
并直接调用它,那么可能能够从系统中挤出更多性能。签名是:
internal static unsafe void memcpyimpl(byte* src, byte* dest, int len)
它确实需要指针,但我相信它已针对可能的最高速度进行了优化,因此我认为没有任何方法可以获得更快的速度,即使您手头有组装。
<强>更新强>:
由于要求(并满足我的好奇心),我测试了这个:
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();
}
}
}
结果:
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
现在更改顺序:
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
现在再次更改订单:
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
换句话说:它们非常有竞争力;作为一般规则,memcpyimpl
是最快的,但它不一定值得担心。
答案 1 :(得分:11)
答案 2 :(得分:4)
对于基本类型数组(如double
),您可以快速复制,即使对于带指针的多维数组也是如此。
在下面的代码中,我使用值1到100初始化一个二维数组A[10,10]
。然后我将这些值复制到一维数组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}
}
}
}
速度有多快我的测试显示它比原生C#copy和Buffer.BlockCopy()
快许多倍。你试试你的情况并告诉我们。
修改1 我用四种方法比较了复制。 1)两个嵌套循环,2)一个串行循环,3)指针,4)BlockCopy。我测量了各种尺寸阵列的每个刻度的副本数量。
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
很明显,谁是胜利者。指针复制的顺序比任何其他方法都要好。
编辑2 显然我不公平地利用编译器/ JIT优化,因为当我移动委托后面的循环来均衡比赛场地时,数字发生了巨大的变化。
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
缓冲副本位于顶部(感谢@Mehrdad),指针副本第二。现在的问题是为什么指针复制不如缓冲区方法快?
答案 3 :(得分:0)
如果在.NET Core上运行,则可以考虑使用source.AsSpan().CopyTo(destination)
(不过请注意在Mono上)。
Method | Job | Runtime | Mean | Error | StdDev | Ratio | RatioSD |
---------------- |----- |-------- |----------:|----------:|----------:|------:|--------:|
ArrayCopy | Clr | Clr | 60.08 ns | 0.8231 ns | 0.7699 ns | 1.00 | 0.00 |
SpanCopy | Clr | Clr | 99.31 ns | 0.4895 ns | 0.4339 ns | 1.65 | 0.02 |
BufferBlockCopy | Clr | Clr | 61.34 ns | 0.5963 ns | 0.5578 ns | 1.02 | 0.01 |
| | | | | | | |
ArrayCopy | Core | Core | 63.33 ns | 0.6843 ns | 0.6066 ns | 1.00 | 0.00 |
SpanCopy | Core | Core | 47.41 ns | 0.5399 ns | 0.5050 ns | 0.75 | 0.01 |
BufferBlockCopy | Core | Core | 59.89 ns | 0.4713 ns | 0.3936 ns | 0.94 | 0.01 |
| | | | | | | |
ArrayCopy | Mono | Mono | 149.82 ns | 1.6466 ns | 1.4596 ns | 1.00 | 0.00 |
SpanCopy | Mono | Mono | 347.87 ns | 2.0589 ns | 1.9259 ns | 2.32 | 0.02 |
BufferBlockCopy | Mono | Mono | 61.52 ns | 1.1691 ns | 1.0364 ns | 0.41 | 0.01 |
答案 4 :(得分:0)
如果采用以下形式的锯齿状数组,则可以避免复制:
double[][] leftNode = new double[3][];
leftNode[0] = sortedIndex;
leftNode[1] = sortedInstances;
leftNode[2] = sortedLabels;