Question

I am trying to modify the imageDenosing class in CUDA SDK, I need to repeat the filter many time incase to capture the time. But my code doesn’t work properly.

//start

__global__ void F1D(TColor *image,int imageW,int imageH, TColor *buffer)
{  

const int ix = blockDim.x * blockIdx.x + threadIdx.x;
const int iy = blockDim.y * blockIdx.y + threadIdx.y;

    if(iy != 0 && iy < imageH-1  && ix < imageW)
    {

        float4 fresult = get_color(image[imageW * iy + ix]);
        float4 fresult4 = get_color(image[imageW * (iy+1) + ix]);
        float4 fresult5 = get_color(image[imageW * (iy-1) + ix]);

        float4 fresult7; 
            fresult7.x = fresult.x*0.5+fresult4.x*.25+fresult5.x*.25;
            fresult7.y = fresult.y*0.5+fresult4.y*.25+fresult5.y*.25;
            fresult7.z = fresult.z*0.5+fresult4.z*.25+fresult5.z*.25;

        buffer[imageW * iy + ix] =      
            make_color(fresult7.x,fresult7.y,fresult7.z,0);     

    }

    image[imageW * iy + ix] =   buffer[imageW * iy + ix];
    //should be use cudaMemcpy, But it fails
}

//extern

extern "C" void
cuda_F1D(TColor *dst, int imageW, int imageH)
{
dim3 threads(BLOCKDIM_X, BLOCKDIM_Y);
dim3 grid(iDivUp(imageW, BLOCKDIM_X), iDivUp(imageH, BLOCKDIM_Y));
Copy<<<grid, threads>>>(dst, imageW, imageH);

size_t size = imageW*imageH*sizeof(TColor);
TColor *host =(TColor*) malloc(size);
TColor *dst2;
//TColor *dst3;
//TColor *d = new TColor(imageW*imageH*sizeof(TColor));
dim3 threads2(imageW,1);
dim3 grid2(iDivUp(imageW, imageW), iDivUp(imageH, 1));

*for(int i = 0;i<1;i++)
{   
    cudaMalloc( (void **)&dst2, size);
            cudaMemcpy(dst2, dst, imageW*imageH*sizeof(TColor),cudaMemcpyHostToDevice);
//cudaMalloc( (void **)&dst3, imageW*imageH*sizeof(TColor));
//cudaMemcpy(dst3, dst, imageW*imageH*sizeof(TColor),cudaMemcpyHostToDevice);
    F1D<<<grid2, threads2>>>(dst, imageW, imageH,dst2);
//cudaMemcpy(dst, dst3, imageW*imageH*sizeof(TColor),cudaMemcpyDeviceToHost);
    cudaFree(dst2);
}*

}

This code works, but cant synchronise the array of image. and lead to many synchronise problem

Answer 1

Your kernel is running asynchronously – you need to wait for it to complete, e.g.

cudaMalloc((void **)&dst2, size);
cudaMemcpy(dst2, dst, imageW * imageH * sizeof(TColor), cudaMemcpyHostToDevice);
F1D<<<grid2, threads2>>>(dst, imageW, imageH, dst2);
cudaThreadSynchronize(); // *** wait for kernel to complete ***
cudaFree(dst2);