CUDA 第一次运行或者修改代码后点ctrl+F5时结果为-842150451, 第二次运行结果正常正常

[Lyn1993]

我的环境是VS2010, win32 控制台程序, cuda4.1
每次第一次运行程序(ctrl+F5), 或者修改代码后第一次运行程序, 结果都会出错, 但是第二次就正常了, 不知道为什么. 附上我的源码:

//kernel function
//get the gradiant, for signal channel graph, kernel is [-1, 0, 1] and inverse of [-1, 0, 1]
__global__ static void gradiant(unsigned char* img, int* GradiantX, int* GradiantY, float* GradiantMagnitude, float* GradiantOrientation, int imgWidth, int imgHeight){
    int mulx, muly;
    int posx, posy;
    int X, Y, Z;
    mulx=ceil(imgWidth/(float)THREAD_NUM);
    muly=ceil(imgHeight/(float)BLOCK_NUM);

    for(int i=0; i<muly; i++){
        for(int j=0; j<mulx; j++){
            posy= i*BLOCK_NUM+ blockIdx.x;
            posx= j*THREAD_NUM+ threadIdx.x;
            if(posx<imgWidth&&posy<imgHeight){
                int index= posx+ posy* imgWidth;
                int up= posy-1; if(up<0)up=0;        int down= posy+1; if(down>=imgHeight)down= imgHeight-1;
                int indexUp= up*imgWidth+posx;    int indexDown= down*imgWidth+ posx;
                GradiantY[index]= img[indexDown]- img[indexUp];

                int left= posx-1; if(left<0)left=0;        int right= posx+1; if(right>=imgWidth)right=imgWidth-1;
                int indexLeft= posy*imgWidth+left;        int indexRigt= posy* imgWidth+ right;
                GradiantX[index]= img[indexRigt]- img[indexLeft];

                GradiantMagnitude[index]= sqrt( (float)(GradiantX[index]*GradiantX[index]+GradiantY[index]*GradiantY[index]) );
                //float angle= atan( ( (float)GradiantY[index] ) / (float)(GradiantX[index]+0.1 ) );
                //if(angle<0)angle+= PI;
                //GradiantOrientation[index]= angle;
                GradiantOrientation[index]= (float) (GradiantX[index]+GradiantY[index]);
            }
        }
    }
}

extern "C" void getGradiant(unsigned char* img, int* GradiantX, int* GradiantY, float* GradiantMagnitude, float* GradiantOrientation, int imgWidth, int imgHeight){
    unsigned char* d_img;
    int* d_GradiantX;
    int* d_GradiantY;
    float* d_GradiantMagnitude;
    float* d_GradiantOrientation;

    cudaMalloc( (void**)&d_img, sizeof(unsigned char)* imgWidth* imgHeight);
    cudaMalloc( (void**)&d_GradiantX, sizeof(int)* imgWidth* imgHeight);
    cudaMalloc( (void**)&d_GradiantY, sizeof(int)* imgWidth* imgHeight);
    cudaMalloc( (void**)&d_GradiantMagnitude, sizeof(float)* imgWidth* imgHeight);
    cudaMalloc( (void**)&d_GradiantOrientation, sizeof(float)* imgWidth* imgHeight);
    cudaMemcpy(d_img, img, sizeof(unsigned char)* imgWidth* imgHeight, cudaMemcpyHostToDevice);

    gradiant<<<BLOCK_NUM, THREAD_NUM>>>(d_img, d_GradiantX, d_GradiantY, d_GradiantMagnitude, GradiantOrientation, imgWidth, imgHeight);


    cudaMemcpy(GradiantX, d_GradiantX, sizeof(int)* imgWidth* imgHeight, cudaMemcpyDeviceToHost);
    cudaMemcpy(GradiantY, d_GradiantY, sizeof(int)* imgWidth* imgHeight, cudaMemcpyDeviceToHost);
    cudaMemcpy(GradiantMagnitude, d_GradiantMagnitude, sizeof(float)* imgWidth* imgHeight, cudaMemcpyDeviceToHost);
    cudaMemcpy(GradiantOrientation, d_GradiantOrientation, sizeof(float)*imgWidth*imgHeight, cudaMemcpyDeviceToHost);

}