cuda中,兩個維度的數組如何使用 也是看了別人的文章,完後自己也實現一下

原文鏈接：https://blog.csdn.net/SUSU0203/article/details/83111221

cudaimprocess.h

#define M 16
#define N 16

__global__ void build_TransMat(float* d_transMat,size_t pitch);

void D2DimTest();

cudaimprocess.cu

__global__ void build_TransMat(float * d_transMat, size_t pitch)
{
	int count = 1;
	int tid_y = blockIdx.y*blockDim.y + threadIdx.y; //這類相當於height 。這麼理解就可以了
	int offset_y = blockDim.y * gridDim.y;

	int tid_x = blockIdx.x*blockDim.x + threadIdx.x; //這裏相當於width
	int offset_x = blockDim.x * gridDim.x;

	for (int j = tid_y; j < M; j += offset_y)
	{
		float * row_d_transMat = (float*)((char*)d_transMat+j*pitch);//這裏要地址對齊，所以不是j*row+d_transMat
		for (int i = tid_x; i < N; i += offset_x)
		{
			if (i == j)
			{
				row_d_transMat[i] = 7;
				printf("%d  %d - \n",i,j);
			}
			else
			{
				row_d_transMat[i] = 0;
			}
		
			count++ ;
		}
	}

	__syncthreads();
}

void D2DimTest()
{
	float * d_transMat;
	float * transMat;
	size_t pitch;
	transMat = new float[M*N];

	cudaMallocPitch(&d_transMat,&pitch,sizeof(float)*N,M);

	cudaMemset2D(d_transMat,pitch,0,sizeof(float)*N,M);

	
	dim3 threadsPerBlock(32,32);
	dim3 blocksPergrid(M+threadsPerBlock.x-1/128, N+threadsPerBlock.y-1/128);
	build_TransMat << <blocksPergrid, threadsPerBlock >> > (d_transMat,pitch);
	cudaMemcpy2D(transMat,sizeof(float)*N,d_transMat,pitch,sizeof(float)*N,M
	,cudaMemcpyKind::cudaMemcpyDeviceToHost);

	std::cout << "二維測試案例:   pitch: " <<pitch<<std::endl;
	for (int j = 0,k=0; j < M; j++)
	{
		for (int i = 0; i < N; i++,k++)
		{
			std::cout << transMat[k] << " ";
		}
		std::cout << std::endl;
	}
	

	cudaFree(d_transMat);
	delete []transMat;
}

這是出來的結果,打印 pitch ,發現是512, 是128倍數, 這個是gpu自己分配的,

那麼這樣一來,以後寫二維數組就簡單多了,直接使用這種方式來處理,不用將二維轉成一維, 雖然犧牲了點效率,但是後續代碼可維護性以及可讀性就強了很多…