cuda纹理内存的使用

CUDA纹理内存的访问速度比全局内存要快，因此处理图像数据时，使用纹理内存是一个提升性能的好方法。

贴一段自己写的简单的实现两幅图像加权和的代码，使用纹理内存实现。

输入：两幅图 lena, moon

  

输出：两幅图像加权和

 #include <opencv2\opencv.hpp>
 #include <iostream>
 #include <string>
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <device_launch_parameters.h>

 using namespace std;
 using namespace cv;

 //声明CUDA纹理
 texture <uchar4, cudaTextureType2D, cudaReadModeNormalizedFloat> refTex1;
 texture <uchar4, cudaTextureType2D, cudaReadModeNormalizedFloat> refTex2;
 //声明CUDA数组
 cudaArray* cuArray1;
 cudaArray* cuArray2;
 //通道数
 cudaChannelFormatDesc cuDesc = cudaCreateChannelDesc<uchar4>();

 __global__ void weightAddKerkel(uchar *pDstImgData, int imgHeight, int imgWidth,int channels)
 {
     const int tidx=blockDim.x*blockIdx.x+threadIdx.x;
     const int tidy=blockDim.y*blockIdx.y+threadIdx.y;

     if (tidx<imgWidth && tidy<imgHeight)
     {
         float4 lenaBGR,moonBGR;
         //使用tex2D函数采样纹理
         lenaBGR=tex2D(refTex1, tidx, tidy);
         moonBGR=tex2D(refTex2, tidx, tidy);

         int idx=(tidy*imgWidth+tidx)*channels;
         float alpha=0.5;
         pDstImgData[idx+]=(alpha*lenaBGR.x+(-alpha)*moonBGR.x)*;
         pDstImgData[idx+]=(alpha*lenaBGR.y+(-alpha)*moonBGR.y)*;
         pDstImgData[idx+]=(alpha*lenaBGR.z+(-alpha)*moonBGR.z)*;
         pDstImgData[idx+]=;
     }
 }

 void main()
 {
     Mat Lena=imread("data/lena.jpg");
     Mat moon=imread("data/moon.jpg");
     cvtColor(Lena, Lena, CV_BGR2BGRA);
     cvtColor(moon, moon, CV_BGR2BGRA);
     int imgWidth=Lena.cols;
     int imgHeight=Lena.rows;
     int channels=Lena.channels();

     //设置纹理属性
     cudaError_t t;
     refTex1.addressMode[] = cudaAddressModeClamp;
     refTex1.addressMode[] = cudaAddressModeClamp;
     refTex1.normalized = false;
     refTex1.filterMode = cudaFilterModeLinear;
     //绑定cuArray到纹理
     cudaMallocArray(&cuArray1, &cuDesc, imgWidth, imgHeight);
     t = cudaBindTextureToArray(refTex1, cuArray1);

     refTex2.addressMode[] = cudaAddressModeClamp;
     refTex2.addressMode[] = cudaAddressModeClamp;
     refTex2.normalized = false;
     refTex2.filterMode = cudaFilterModeLinear;
      cudaMallocArray(&cuArray2, &cuDesc, imgWidth, imgHeight);
     t = cudaBindTextureToArray(refTex2, cuArray2);

     //拷贝数据到cudaArray
     t=cudaMemcpyToArray(cuArray1, ,, Lena.data, imgWidth*imgHeight*sizeof(uchar)*channels, cudaMemcpyHostToDevice);
     t=cudaMemcpyToArray(cuArray2, ,, moon.data, imgWidth*imgHeight*sizeof(uchar)*channels, cudaMemcpyHostToDevice);

     //输出图像
     Mat dstImg=Mat::zeros(imgHeight, imgWidth, CV_8UC4);
     uchar *pDstImgData=NULL;
     t=cudaMalloc(&pDstImgData, imgHeight*imgWidth*sizeof(uchar)*channels);

     //核函数，实现两幅图像加权和
     dim3 block(,);
     dim3 grid( (imgWidth+block.x-)/block.x, (imgHeight+block.y-)/block.y );
     weightAddKerkel<<<grid, block, >>>(pDstImgData, imgHeight, imgWidth, channels);
     cudaThreadSynchronize();

     //从GPU拷贝输出数据到CPU
     t=cudaMemcpy(dstImg.data, pDstImgData, imgWidth*imgHeight*sizeof(uchar)*channels, cudaMemcpyDeviceToHost);

     //显示
     namedWindow("show");
     imshow("show", dstImg);
     waitKey();
 }