开发者

Is there a faster/cleaner way to copy structs to an array in C#?

I have a float4x4 struct which simply containts 16 floats:

struct float4x4
{
        public float M11; public float M12; public float M13; public float M14;
        public float M21; public float M22; public float M23; public float M24;
        public float M31; public float M32; public float M33; public float M34;
        pub开发者_JS百科lic float M41; public float M42; public float M43; public float M44;
}

I want to copy an array of these structs into a big array of floats. This is as far as I know a 1:1 copy of a chunk of memory

What I do know is rather ugly, and not that fast:

        int n = 0;
        for (int i = 0; i < Length; i++)
        {
            array[n++] = value[i].M11;
            array[n++] = value[i].M12;
            array[n++] = value[i].M13;
            array[n++] = value[i].M14;

            array[n++] = value[i].M21;
            array[n++] = value[i].M22;
            array[n++] = value[i].M23;
            array[n++] = value[i].M24;

            array[n++] = value[i].M31;
            array[n++] = value[i].M32;
            array[n++] = value[i].M33;
            array[n++] = value[i].M34;

            array[n++] = value[i].M41;
            array[n++] = value[i].M42;
            array[n++] = value[i].M43;
            array[n++] = value[i].M44;
        }

If I was using a lower level language, I would simply use memcpy, what can I use as an equivilant in C#?


You can't use a memory copy, as you can't blindly assume anything about how the members are stored inside the structure. The JIT compiler could decide to store them with a few bytes of padding between them, if that would make it faster.

Your structure is way too large for the recommended size of a structure anyway, so you should make it a class. Also, structures should not be mutable, which also talks for a class.

If you store the properties in an array internally, you can use that for copying the values:

class float4x4 {

  public float[] Values { get; private set; } 

  public float4x4() {
    Values = new float[16];
  }

  public float M11 { get { return Values[0]; } set { Values[0] = value; } }
  public float M12 { get { return Values[0]; } set { Values[0] = value; } }
  ...
  public float M43 { get { return Values[14]; } set { Values[14] = value; } }
  public float M44 { get { return Values[15]; } set { Values[15] = value; } }

}

Now you can get the Values array from the object and copy to the array using the Array.CopyTo method:

int n = 0;
foreach (float4x4 v in values) {
  v.Values.CopyTo(array, n);
  n += 16;
}


This is perhaps equally ugly, but is very fast.

using System.Runtime.InteropServices;

namespace ConsoleApplication23 {
  public class Program {
    public static void Main() {
      var values=new[] {
        new float4x4(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16),
        new float4x4(-1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -12, -13, -14, -15, -16)
      };
      var result=Transform(values);
    }

    public static unsafe float[] Transform(float4x4[] values) {
      var array=new float[values.Length*16];
      fixed(float* arrayStart=array) {
        var destp=arrayStart;
        fixed(float4x4* valuesStart=values) {
          int count=values.Length;
          for(var valuesp=valuesStart; count>0; ++valuesp, --count) {
            var sourcep=valuesp->data;
            for(var i=0; i<16/4; ++i) {
              *destp++=*sourcep++;
              *destp++=*sourcep++;
              *destp++=*sourcep++;
              *destp++=*sourcep++;
            }
          }
        }
        return array;
      }
    }

    [StructLayout(LayoutKind.Explicit)]
    public unsafe struct float4x4 {
      [FieldOffset(0)] public float M11;
      [FieldOffset(4)] public float M12;
      [FieldOffset(8)] public float M13;
      [FieldOffset(12)] public float M14;
      [FieldOffset(16)] public float M21;
      [FieldOffset(20)] public float M22;
      [FieldOffset(24)] public float M23;
      [FieldOffset(28)] public float M24;
      [FieldOffset(32)] public float M31;
      [FieldOffset(36)] public float M32;
      [FieldOffset(40)] public float M33;
      [FieldOffset(44)] public float M34;
      [FieldOffset(48)] public float M41;
      [FieldOffset(52)] public float M42;
      [FieldOffset(56)] public float M43;
      [FieldOffset(60)] public float M44;

      //notice the use of "fixed" keyword to make the array inline
      //and the use of the FieldOffset attribute to overlay that inline array on top of the other fields
      [FieldOffset(0)] public fixed float data[16];

      public float4x4(float m11, float m12, float m13, float m14,
        float m21, float m22, float m23, float m24,
        float m31, float m32, float m33, float m34,
        float m41, float m42, float m43, float m44) {
        M11=m11; M12=m12; M13=m13; M14=m14;
        M21=m21; M22=m22; M23=m23; M24=m24;
        M31=m31; M32=m32; M33=m33; M34=m34;
        M41=m41; M42=m42; M43=m43; M44=m44;
      }
    }
  }
}


OK this is my test harness. My project properties are Release Build, "optimize code" and also "Allow unsafe code" checked.

Surprisingly (to me anyway) the performance is very different inside and outside the IDE. When run from the IDE there are noticeable differences (and the x64 difference is huge). When run outside the IDE, it's a wash.

So this is kind of weird, and I can't explain the results for IDE+x64. Maybe this is interesting to some people, but because it no longer purports to provide an answer to the poster's original question, maybe this should be moved to some other topic?

Inside IDE, platform set to x86

pass 1: old 00:00:09.7505625 new 00:00:08.6897013 percent 0.1088

Inside IDE, platform set to x64

pass 1: old 00:00:14.7584514 new 00:00:08.8835715 percent 0.398068858362741

Running from command line, platform set to x86

pass 1: old 00:00:07.6576469 new 00:00:07.2818252 percent 0.0490779615341104

Running from command line, platform set to x64

pass 1: old 00:00:07.2501032 new 00:00:07.3077479 percent -0.00795087992678504

And this is the code:

using System;
using System.Runtime.InteropServices;

namespace ConsoleApplication23 {
  public class Program {
    public static void Main() {
      const int repeatCount=20;
      const int arraySize=5000000;

      var values=MakeValues(arraySize);

      for(var pass=0; pass<2; ++pass) {
        Console.WriteLine("Starting old");
        var startOld=DateTime.Now;
        for(var i=0; i<repeatCount; ++i) {
          var result=TransformOld(values);
        }
        var elapsedOld=DateTime.Now-startOld;

        Console.WriteLine("Starting new");
        var startNew=DateTime.Now;
        for(var i=0; i<repeatCount; ++i) {
          var result=TransformNew(values);
        }
        var elapsedNew=DateTime.Now-startNew;

        var difference=elapsedOld-elapsedNew;
        var percentage=(double)difference.TotalMilliseconds/elapsedOld.TotalMilliseconds;

        Console.WriteLine("pass {0}: old {1} new {2} percent {3}", pass, elapsedOld, elapsedNew, percentage);
      }
      Console.Write("Press enter: ");
      Console.ReadLine();
    }

    private static float4x4[] MakeValues(int count) {
      var result=new float4x4[count];
      for(var i=0; i<count; ++i) {
        result[i]=new float4x4(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
      }
      return result;
    }

    public static float[] TransformOld(float4x4[] value) {
      var array=new float[value.Length*16];
      int n = 0;
      for(int i = 0; i < value.Length; i++) {
        array[n++] = value[i].M11;
        array[n++] = value[i].M12;
        array[n++] = value[i].M13;
        array[n++] = value[i].M14;

        array[n++] = value[i].M21;
        array[n++] = value[i].M22;
        array[n++] = value[i].M23;
        array[n++] = value[i].M24;

        array[n++] = value[i].M31;
        array[n++] = value[i].M32;
        array[n++] = value[i].M33;
        array[n++] = value[i].M34;

        array[n++] = value[i].M41;
        array[n++] = value[i].M42;
        array[n++] = value[i].M43;
        array[n++] = value[i].M44;
      }
      return array;
    }

    public static unsafe float[] TransformNew(float4x4[] values) {
      var array=new float[values.Length*16];
      fixed(float* arrayStart=array) {
        var destp=arrayStart;
        fixed(float4x4* valuesStart=values) {
          int count=values.Length;
          for(var valuesp=valuesStart; count>0; ++valuesp, --count) {
            var sourcep=valuesp->data;
            for(var i=0; i<16/4; ++i) {
              *destp++=*sourcep++;
              *destp++=*sourcep++;
              *destp++=*sourcep++;
              *destp++=*sourcep++;
            }
          }
        }
        return array;
      }
    }

    [StructLayout(LayoutKind.Explicit)]
    public unsafe struct float4x4 {
      [FieldOffset(0)] public float M11;
      [FieldOffset(4)] public float M12;
      [FieldOffset(8)] public float M13;
      [FieldOffset(12)] public float M14;
      [FieldOffset(16)] public float M21;
      [FieldOffset(20)] public float M22;
      [FieldOffset(24)] public float M23;
      [FieldOffset(28)] public float M24;
      [FieldOffset(32)] public float M31;
      [FieldOffset(36)] public float M32;
      [FieldOffset(40)] public float M33;
      [FieldOffset(44)] public float M34;
      [FieldOffset(48)] public float M41;
      [FieldOffset(52)] public float M42;
      [FieldOffset(56)] public float M43;
      [FieldOffset(60)] public float M44;

      //notice the use of "fixed" keyword to make the array inline
      //and the use of the FieldOffset attribute to overlay that inline array on top of the other fields
      [FieldOffset(0)] public fixed float data[16];

      public float4x4(float m11, float m12, float m13, float m14,
        float m21, float m22, float m23, float m24,
        float m31, float m32, float m33, float m34,
        float m41, float m42, float m43, float m44) {
        M11=m11; M12=m12; M13=m13; M14=m14;
        M21=m21; M22=m22; M23=m23; M24=m24;
        M31=m31; M32=m32; M33=m33; M34=m34;
        M41=m41; M42=m42; M43=m43; M44=m44;
      }
    }
  }
}


Maybe you could alias the array of structs with an array of floats and do absolutely no copying. Check this SO answer for a starting point


It's not necessarily a 1 to 1 copy. The CLR is free to layout the fields in a struct in whichever way it likes. It might reorder them, realign them.

If you add a [StructLayout(LayoutKind.Sequential)] a direct copy might be possible, but I'd still go with something similar to your original code.

0

上一篇:

下一篇:

精彩评论

暂无评论...
验证码 换一张
取 消

最新问答

问答排行榜