cuda改写C++代码，进行移植GPU上优化速度反而变慢？

[-阿白-]

将一个C++代码进行移植到GPU上，为了并行计算，实现运行时间减少加速。

但是在运行过程中，发现：移植到GPU反而耗时更长，后面检索了一下资料和进行测试，发现应该是cudaMemcpy这部分耗时了，计算部分在GPU上很快便执行了，求各路大神，提点意见，如何让我的代码移植到GPU上并行计算，且加速！谢谢！

以下是改写前的部分代码：

mat2D<int>* s::GetResidual(mat2D<int> *img, mat2D<int>* kernel)
{
	mat2D<int> *residual = new mat2D<int>(img->rows-kernel->rows+1, img->cols-kernel->cols+1);
	for (int ir=0; ir < (img->rows - kernel->rows + 1); ir++)
	{
		for (int ic=0; ic < (img->cols - kernel->cols + 1); ic++)
		{
			int convVal = 0;
			for (int kr=0; kr < kernel->rows; kr++)
			{
				for (int kc=0; kc < kernel->cols; kc++)
				{
					convVal = convVal + img->Read(ir+kr, ic+kc) * kernel->Read(kr, kc);
				}
			}
			residual->Write(ir, ic, convVal);
		}
	}
	return residual;
}

改写为cuda的代码：

__global__ void GPU_GetResidual(int *img_arr, int img_cols, int *kernel_arr, int kernel_rows, int kernel_cols, int *res_arr, int res_cols)
{
	int ir = blockIdx.x;
	int ic = blockIdx.y;

	int convVal = 0;
	
	for (int kr=0; kr < kernel_rows; kr++)
	{
		for (int kc=0; kc < kernel_cols; kc++)
		{
			convVal = convVal + img_arr[(ir+kr)*img_cols+(ic+kc)] * kernel_arr[kr*kernel_cols+kc];
		}
	}

	res_arr[ir*res_cols+ic] = convVal;
}

mat2D<int>* s::GetResidual(mat2D<int> *img, mat2D<int>* kernel)
{
	mat2D<int> *residual = new mat2D<int>(img->rows-kernel->rows+1, img->cols-kernel->cols+1);

	std::vector<int> img_ve = img->getVector();
	std::vector<int> kernel_ve = kernel->getVector();
	std::vector<int> res_ve = residual->getVector();

	int *img_arr;
	cudaMalloc((void**) &img_arr, img_ve.size()*sizeof(int));
	cudaMemcpy(img_arr, &img_ve[0], img_ve.size()*sizeof(int), cudaMemcpyHostToDevice);

	int *kernel_arr;
	cudaMalloc((void**) &kernel_arr, kernel_ve.size()*sizeof(int));
	cudaMemcpy(kernel_arr, &kernel_ve[0], kernel_ve.size()*sizeof(int), cudaMemcpyHostToDevice);

	int *res_arr;
	cudaMalloc((void**) &res_arr, residual->rows * residual->cols * sizeof(int));
	
	dim3 blocks(img->rows - kernel->rows + 1, img->cols - kernel->cols + 1);

	GPU_GetResidual<<<blocks, 1>>>(img_arr, img->cols, kernel_arr, kernel->rows, kernel->cols, res_arr, residual->cols);

	cudaMemcpy(&res_ve[0], res_arr, residual->rows * residual->cols * sizeof(int), cudaMemcpyDeviceToHost);
	
	residual->setVector(res_ve);
	
	cudaFree(res_arr);
	cudaFree(kernel_arr);
	cudaFree(img_arr);
	
	return residual;
}

[stellarman]

内存的读写img_arr[(ir+kr)*img_cols+(ic+kc)] 跨行超过cache了，这个CUDA肯定很慢