private static int Diff (int c0,int c1)
{
unsafe {
byte* pc0 = (byte*) &c0;
byte* pc1 = (byte*) &c1;
int d0 = pc0[0] - pc1[0];
int d1 = pc0[1] - pc1[1];
int d2 = pc0[2] - pc1[2];
int d3 = pc0[3] - pc1[3];
d0 *= d0;
d1 *= d1;
d2 *= d2;
d3 *= d3;
return d0 + d1 + d2 + d3;
}
}
最明显的是,这将有益于SIMD,但让我们假设我不想去那里,因为它有点麻烦.
>对于较低级别的东西(调用C库,在GPGPU上执行)相同
>多线程 – 我会用的
编辑：为方便起见,一些反映真实环境和用例的测试代码. (实际上,涉及更多的数据,并且数据不在单个大块中进行比较,而是在多个几kb的数据块中进行比较.)
public static class ByteCompare
{
private static void Main ()
{
const int n = 1024 * 1024 * 20;
const int repeat = 20;
var rnd = new Random (0);
Console.Write ("Generating test data... ");
var t0 = Enumerable.Range (1,n)
.Select (x => rnd.Next (int.MinValue,int.MaxValue))
.ToArray ();
var t1 = Enumerable.Range (1,int.MaxValue))
.ToArray ();
Console.WriteLine ("complete.");
GC.Collect (2,GCCollectionMode.Forced);
Console.WriteLine ("GCs: " + GC.CollectionCount (0));
{
var sw = Stopwatch.StartNew ();
long res = 0;
for (int reps = 0; reps < repeat; reps++) {
for (int i = 0; i < n; i++) {
int c0 = t0[i];
int c1 = t1[i];
res += ByteDiff_REGULAR (c0,c1);
}
}
sw.Stop ();
Console.WriteLine ("res=" + res + ",t=" + sw.Elapsed.TotalSeconds.ToString ("0.00") + "s - ByteDiff_REGULAR");
}
{
var sw = Stopwatch.StartNew ();
long res = 0;
for (int reps = 0; reps < repeat; reps++) {
for (int i = 0; i < n; i++) {
int c0 = t0[i];
int c1 = t1[i];
res += ByteDiff_UNSAFE (c0,t=" + sw.Elapsed.TotalSeconds.ToString ("0.00") + "s - ByteDiff_UNSAFE_PTR");
}
Console.WriteLine ("GCs: " + GC.CollectionCount (0));
Console.WriteLine ("Test complete.");
Console.ReadKey (true);
}
public static int ByteDiff_REGULAR (int c0,int c1)
{
var c00 = (byte) (c0 >> (8 * 0));
var c01 = (byte) (c0 >> (8 * 1));
var c02 = (byte) (c0 >> (8 * 2));
var c03 = (byte) (c0 >> (8 * 3));
var c10 = (byte) (c1 >> (8 * 0));
var c11 = (byte) (c1 >> (8 * 1));
var c12 = (byte) (c1 >> (8 * 2));
var c13 = (byte) (c1 >> (8 * 3));
var d0 = (c00 - c10);
var d1 = (c01 - c11);
var d2 = (c02 - c12);
var d3 = (c03 - c13);
d0 *= d0;
d1 *= d1;
d2 *= d2;
d3 *= d3;
return d0 + d1 + d2 + d3;
}
private static int ByteDiff_UNSAFE (int c0,int c1)
{
unsafe {
byte* pc0 = (byte*) &c0;
byte* pc1 = (byte*) &c1;
int d0 = pc0[0] - pc1[0];
int d1 = pc0[1] - pc1[1];
int d2 = pc0[2] - pc1[2];
int d3 = pc0[3] - pc1[3];
d0 *= d0;
d1 *= d1;
d2 *= d2;
d3 *= d3;
return d0 + d1 + d2 + d3;
}
}
}
这对我来说(在i5上运行为x64 Release)：
Generating test data... complete. GCs: 8 res=18324555528140,t=1.46s - ByteDiff_REGULAR res=18324555528140,t=1.15s - ByteDiff_UNSAFE res=18324555528140,t=1.73s - Diff_Alex1 res=18324555528140,t=1.63s - Diff_Alex2 res=18324555528140,t=3.59s - Diff_Alex3 res=18325828513740,t=3.90s - Diff_Alex4 GCs: 8 Test complete.
解决方法
Most obvIoUsly,this would benefit from SIMD,but let us suppose I don’t want to go there because it is a bit of a hassle.
如果你想要避免它,但它实际上是相当好的支持直接从C#.如果较大的算法适用于SIMD处理,那么我们不希望将其卸载到GPU,我希望这是迄今为止最大的性能优胜者.
http://www.drdobbs.com/architecture-and-design/simd-enabled-vector-types-with-c/240168888
Multithreading
当然,每个cpu核心使用一个线程.您也可以使用像Parallel.For这样的结构,让.NET整理出要使用的线程数.这是非常好的,但由于你知道这肯定是cpu限制你可能(或者可能不会)通过自己管理线程获得更好的结果.
对于加速实际的代码块,使用位掩码和位移可能会更快地获取各个值的工作,而不是使用指针.这有额外的好处,你不需要不安全的代码块,例如
byte b0_leftmost = (c0 & 0xff000000) >> 24;
原文地址：https://www.jb51.cc/csharp/97147.html
版权声明：本文内容由互联网用户自发贡献，该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容， 请发送邮件至 dio@foxmail.com 举报，一经查实，本站将立刻删除。
