我正在使用Cuda和C++在GPU上执行两个任务(分别分离为2个内核)。作为输入,我采用一个NxM矩阵(以浮点数组的形式存储在主机内存中)。然后我将使用一个内核对此矩阵执行一些操作,使其成为一个NxMxD矩阵。然后我有第二个内核,它对这个3D矩阵执行一些操作(我只读取值,不必向其中写入值)。
在纹理内存中操作似乎对我的任务更快,所以我的问题是是否可以在内核1之后将数据从设备上的全局内存复制并直接传输到内核2的纹理内存中,而无需将其带回主机?
更新:
我添加了一些代码来更好地说明我的问题。
这里是这两个内核。第一个现在只是一个占位符,并将2D矩阵复制到3D矩阵中。
第二步将采用这个3D矩阵,现在被表示为纹理,并对其执行一些操作。暂时为空白。
然后代码的主体如下所示:
在纹理内存中操作似乎对我的任务更快,所以我的问题是是否可以在内核1之后将数据从设备上的全局内存复制并直接传输到内核2的纹理内存中,而无需将其带回主机?
更新:
我添加了一些代码来更好地说明我的问题。
这里是这两个内核。第一个现在只是一个占位符,并将2D矩阵复制到3D矩阵中。
__global__ void computeFeatureVector(float* imData3D_dev, int imX, int imY, int imZ) {
//calculate each thread global index
int xindex=blockIdx.x*blockDim.x+threadIdx.x;
int yindex=blockIdx.y*blockDim.y+threadIdx.y;
#pragma unroll
for (int z=0; z<imZ; z++) {
imData3D_dev[xindex+yindex*imX + z*imX*imY] = tex2D(texImIp,xindex,yindex);
}
}
第二步将采用这个3D矩阵,现在被表示为纹理,并对其执行一些操作。暂时为空白。
__global__ void kernel2(float* resData_dev, int imX) {
//calculate each thread global index
int xindex=blockIdx.x*blockDim.x+threadIdx.x;
int yindex=blockIdx.y*blockDim.y+threadIdx.y;
resData_dev[xindex+yindex*imX] = tex3D(texImIp3D,xindex,yindex, 0);
return;
}
然后代码的主体如下所示:
// declare textures
texture<float,2,cudaReadModeElementType> texImIp;
texture<float,3,cudaReadModeElementType> texImIp3D;
void main_fun() {
// constants
int imX = 1024;
int imY = 768;
int imZ = 16;
// input data
float* imData2D = new float[sizeof(float)*imX*imY];
for(int x=0; x<imX*imY; x++)
imData2D[x] = (float) rand()/RAND_MAX;
//create channel to describe data type
cudaArray* carrayImIp;
cudaChannelFormatDesc channel;
channel=cudaCreateChannelDesc<float>();
//allocate device memory for cuda array
cudaMallocArray(&carrayImIp,&channel,imX,imY);
//copy matrix from host to device memory
cudaMemcpyToArray(carrayImIp,0,0,imData2D,sizeof(float)*imX*imY,cudaMemcpyHostToDevice);
// Set texture properties
texImIp.filterMode=cudaFilterModePoint;
texImIp.addressMode[0]=cudaAddressModeClamp;
texImIp.addressMode[1]=cudaAddressModeClamp;
// bind texture reference with cuda array
cudaBindTextureToArray(texImIp,carrayImIp);
// kernel params
dim3 blocknum;
dim3 blocksize;
blocksize.x=16; blocksize.y=16; blocksize.z=1;
blocknum.x=(int)ceil((float)imX/16);
blocknum.y=(int)ceil((float)imY/16);
// store output here
float* imData3D_dev;
cudaMalloc((void**)&imData3D_dev,sizeof(float)*imX*imY*imZ);
// execute kernel
computeFeatureVector<<<blocknum,blocksize>>>(imData3D_dev, imX, imY, imZ);
//unbind texture reference to free resource
cudaUnbindTexture(texImIp);
// check copied ok
float* imData3D = new float[sizeof(float)*imX*imY*imZ];
cudaMemcpy(imData3D,imData3D_dev,sizeof(float)*imX*imY*imZ,cudaMemcpyDeviceToHost);
cout << " kernel 1" << endl;
for (int x=0; x<10;x++)
cout << imData3D[x] << " ";
cout << endl;
delete [] imData3D;
//
// kernel 2
//
// copy data on device to 3d array
cudaArray* carrayImIp3D;
cudaExtent volumesize;
volumesize = make_cudaExtent(imX, imY, imZ);
cudaMalloc3DArray(&carrayImIp3D,&channel,volumesize);
cudaMemcpyToArray(carrayImIp3D,0,0,imData3D_dev,sizeof(float)*imX*imY*imZ,cudaMemcpyDeviceToDevice);
// texture params and bind
texImIp3D.filterMode=cudaFilterModePoint;
texImIp3D.addressMode[0]=cudaAddressModeClamp;
texImIp3D.addressMode[1]=cudaAddressModeClamp;
texImIp3D.addressMode[2]=cudaAddressModeClamp;
cudaBindTextureToArray(texImIp3D,carrayImIp3D,channel);
// store output here
float* resData_dev;
cudaMalloc((void**)&resData_dev,sizeof(float)*imX*imY);
// kernel 2
kernel2<<<blocknum,blocksize>>>(resData_dev, imX);
cudaUnbindTexture(texImIp3D);
//copy result matrix from device to host memory
float* resData = new float[sizeof(float)*imX*imY];
cudaMemcpy(resData,resData_dev,sizeof(float)*imX*imY,cudaMemcpyDeviceToHost);
// check copied ok
cout << " kernel 2" << endl;
for (int x=0; x<10;x++)
cout << resData[x] << " ";
cout << endl;
delete [] imData2D;
delete [] resData;
cudaFree(imData3D_dev);
cudaFree(resData_dev);
cudaFreeArray(carrayImIp);
cudaFreeArray(carrayImIp3D);
}
我很高兴第一个内核能正常工作,但似乎3D矩阵imData3D_dev没有正确地绑定到纹理texImIp3D上。
答案
我使用了cudaMemcpy3D来解决我的问题。以下是主函数第二部分的修订代码。imData3D_dev包含来自第一个内核的全局内存中的3D矩阵。
cudaArray* carrayImIp3D;
cudaExtent volumesize;
volumesize = make_cudaExtent(imX, imY, imZ);
cudaMalloc3DArray(&carrayImIp3D,&channel,volumesize);
cudaMemcpy3DParms copyparms={0};
copyparms.extent = volumesize;
copyparms.dstArray = carrayImIp3D;
copyparms.kind = cudaMemcpyDeviceToDevice;
copyparms.srcPtr = make_cudaPitchedPtr((void*)imData3D_dev, sizeof(float)*imX,imX,imY);
cudaMemcpy3D(©parms);
// texture params and bind
texImIp3D.filterMode=cudaFilterModePoint;
texImIp3D.addressMode[0]=cudaAddressModeClamp;
texImIp3D.addressMode[1]=cudaAddressModeClamp;
texImIp3D.addressMode[2]=cudaAddressModeClamp;
cudaBindTextureToArray(texImIp3D,carrayImIp3D,channel);
// store output here
float* resData_dev;
cudaMalloc((void**)&resData_dev,sizeof(float)*imX*imY);
kernel2<<<blocknum,blocksize>>>(resData_dev, imX);
// ... clean up