k Nearest Neighbor Search by CUDA

kNN（k Nearest Neighbor）是常用的群集算法（Cluster Algorithm）用于空间搜索。目前最快的kNN方法莫过于KDTree的版本，不过基本上都是CPU的比如ANN C++ Library。对于GPU来说，实现加速结构比较复杂，因为没有栈所以无法递归，而且执行资源有限，不能像CPU一样舒舒服服的顺序执行。

为了方便起见我直接用komrade粘合C++与CUDA，下面是测试程序的代码。 

#include <komrade/device_vector.h>
#include <komrade/transform.h>
#include <komrade/range.h>
#include <komrade/copy.h>
#include <komrade/fill.h>
#include <komrade/sort.h>
#include <komrade/replace.h>
#include <komrade/functional.h>
#include <iostream>
#include <iterator>
#include <math.h>
#include <memory>
#include <boost/timer.hpp>

using namespace std;

#include <ANN/ANN.h>
#pragma comment(lib,"ANN.lib")

struct dist
{
    dist(float qx, float qy, float qz) : x(qx), y(qy), z(qz)
    {
    }
    float x,y,z;
    
    __host__ __device__ float operator()(const float3& p) const
    {
        float a = p.x - x;
        float b = p.y - y;
        float c = p.z - z;
        return a*a + b*b + c*c;
    }
};

int main(int argc, char* argv[])
{
    if( argc < 2 )
        return -1;
            
    int N = 0;
    sscanf(argv[1],"%d",&N);
    
    boost::timer Stopwatch;
    ANNpointArray ANNPh = annAllocPts(N,3);
    komrade::host_vector<float3> Ph(N);
    komrade::device_vector<int> Idx(N);
    komrade::device_vector<float> Dist(N);
    
    
    for(int i=0; i<N; ++i)
    {
        float x = (float)rand() / (float)RAND_MAX;;
        float y = (float)rand() / (float)RAND_MAX;;
        float z = (float)rand() / (float)RAND_MAX;;
        
        Ph[i].x = x;
        Ph[i].y = y;
        Ph[i].z = z;
        
        ANNPh[i][0] = x;
        ANNPh[i][1] = y;
        ANNPh[i][2] = z;
                
        if( N < 32 )
            printf("<%f %f %f>\n",x,y,z);
        Idx[i] = i;
    }
    cout<<fixed<<"Generate Data Used "<<Stopwatch.elapsed()<<endl;
    Stopwatch.restart();
        
    float QueryPos[3];
    QueryPos[0] = (float)rand() / (float)RAND_MAX * 0.1f + 0.5f;
    QueryPos[1] = (float)rand() / (float)RAND_MAX * 0.1f + 0.5f;
    QueryPos[2] = (float)rand() / (float)RAND_MAX * 0.1f + 0.5f;
    if( N < 32 )
        printf("[%f %f %f]\n",QueryPos[0],QueryPos[1],QueryPos[2]);
        
    komrade::device_vector<float3> Pd = Ph;
    cout<<fixed<<"Copy Data Used "<<Stopwatch.elapsed()<<endl;
    Stopwatch.restart();
    
    komrade::transform(Pd.begin(), Pd.end(), Dist.begin(), dist(QueryPos[0], QueryPos[1], QueryPos[2]));
    komrade::sort_by_key(Dist.begin(), Dist.end(), Idx.begin());
    cout<<fixed<<"Sort Data by CUDA Used "<<Stopwatch.elapsed()<<endl;
    Stopwatch.restart();
    
    ANNkd_tree ANNTree(ANNPh,N,3);
    ANNidx ANNIdx[1];
    ANNdist ANNDst[1];
    ANNTree.annkSearch(QueryPos,1,ANNIdx,ANNDst);
    cout<<fixed<<"Sort Data by ANN Used "<<Stopwatch.elapsed()<<endl;

    
    if( N < 32 )
    {
        copy(Dist.begin(), Dist.end(), ostream_iterator<float>(cout," "));
        cout<<endl;
        copy(Idx.begin(), Idx.end(), ostream_iterator<int>(cout," "));
    }
    
    
    return 0;
}

下面是性能分析对比，可以看到CUDA版本在大量点的情况下还是非常有优势的，不过在纯RenderFarm就用不上了，恐怕只能在装备有G80以上芯片的桌面环境或者工作站上执行。测试环境如下，

• Intel [email protected]
• Apacer 2Gx2
• 9800GT 512M
• CUDA 2.2, komrade 0.9

	10	100	1000	10000	100000	1000000
CUDA	0.003	0.003	0.00300	0.007	0.012	0.038
ANN	0.000	0.000	0.00100	0.008	0.131	2.490

 

 

极其明显的，当排序数目数量级超过10e5的时候，CUDA显示出了巨大的速度优势。数目比较少的时候还是CPU的比较快速。

启发自《Fast k Nearest Neighbor Search using GPU》