今天忙活了3个小时,竟然被一个苦恼的CUDA小例程给困住了,本来是参照Rachal zhang大神的CUDA学习笔记来一个模仿,结果却自己给自己糊里糊涂,最后还是弄明白了一些。
RZ大神对CUDA关于kernel,memory的介绍还是蛮清楚,看完决定写一个二维数组的加法。如果是C++里的加法,那就简单了,用C[i][j] = A[i][j] +B[i][j]就可以。
1 void CppMatAdd(int A[M][N],int B[M][N],int C[M][N]){ 2 for(int i=0;i<M;i++) 3 for(int j=0;j<N;j++) 4 C[i][j] = A[i][j] + B[i][j]; 5 }
1 int main() 2 { 3 int a[M][N] = {1,2,3,4,5,6,7,8,9,10,11,12}; 4 int b[M][N] = {1,2,3,4,5,6,7,8,9,10,11,12}; 5 int c[M][N] ; 6 CppMatAdd(a,b,c); 7 std::cout<<c[0][0]; 8 }
运行上面代码,就可以实现二维矩阵(也就是数组)的加法运算。
但是CUDA计算是在GPU上实现的,要划分出专门的内存区域给GPU做运算,结果就是,我们必须划分出主机内存、设备内存分别供CPU、GPU访问。
对于一维的情况,我们设置好主机变量,设备变量即可。具体可以参找RZ的博客。
但是二维的情况麻烦就来了,最一开始我也是设置出主机变量,设备变量,一一对应的分配内存,拷贝数据,GPU运算,最后考出结果。但是发现怎么调试结果都不对,最主要的原因是c++的二维数组实际上是一维数组的指针,所以,无法按照一位数组的模式去拷贝数据,结果相映的写法就麻烦许多,其实说到底还是还原成一维数组的方法去做的加法运算,代码如下,具体就不想赘述了,代码能力有限,慢慢来吧,今天算是把指针弄的更清楚了。
/*-------------------------------------------- * Date:2015-3-18 * Author:李根 * FileName:.cpp * Description:CUDA二维数组加法 ------------------------------------------------*/ #include "cuda_runtime.h" #include "device_launch_parameters.h" #include <iostream> #include <stdio.h> static const int M = 4; static const int N = 3; //矩阵加法的kernel __global__ void addMat(int **A,int **B,int **C) { int i = blockIdx.x * blockDim.x + threadIdx.x; int j = blockIdx.y * blockDim.y + threadIdx.y; if(i < M && j < N) C[i][j] = A[i][j] + B[i][j]; } int main() {int **A = (int **)malloc(M*sizeof(int *)); //host memory int **B = (int **)malloc(M*sizeof(int *)); //host memory int **C = (int **)malloc(M*sizeof(int *)); //host memory int *dataA =(int *)malloc(M*N*sizeof(int )); //host memory data int *dataB = (int *)malloc(M*N*sizeof(int )); //host memory data int *dataC =(int *)malloc(M*N*sizeof(int )); //host memory data int **dev_A ; //device memory int **dev_B ; //device memory int **dev_C ; //device memory int *dev_dataA ; //device memory data int *dev_dataB ; //device memory data int *dev_dataC ; //device memory data cudaMalloc((void**)(&dev_A), M*sizeof(int*)); cudaMalloc((void**)(&dev_dataA), M*N*sizeof(int)); cudaMalloc((void**)(&dev_B), M*sizeof(int*)); cudaMalloc((void**)(&dev_dataB), M*N*sizeof(int)); cudaMalloc((void**)(&dev_C), M*sizeof(int*)); cudaMalloc((void**)(&dev_dataC), M*N*sizeof(int)); for(int i=0;i<M*N;i++) { dataA[i] = i; dataB[i] = i+1; dataC[i] =0; } cudaMemcpy((void*)(dev_dataA), (void*)(dataA), M*N*sizeof(int*), cudaMemcpyHostToDevice); cudaMemcpy((void*)(dev_dataB), (void*)(dataB), M*N*sizeof(int*), cudaMemcpyHostToDevice); for(int i=0;i<M;i++) { A[i] = dev_dataA + N*i; B[i] = dev_dataB + N*i; C[i] = dev_dataC + N*i; } cudaMemcpy((void*)(dev_A), (void*)(A), M*sizeof(int*), cudaMemcpyHostToDevice); cudaMemcpy((void*)(dev_B), (void*)(B), M*sizeof(int*), cudaMemcpyHostToDevice); cudaMemcpy((void*)(dev_C), (void*)(C), M*sizeof(int*), cudaMemcpyHostToDevice); dim3 threadPerBlock(16,16); dim3 numBlocks((N+threadPerBlock.x-1)/(threadPerBlock.x), (M+threadPerBlock.y-1)/(threadPerBlock.y)); addMat<<<numBlocks,threadPerBlock>>>(dev_A,dev_B,dev_C); cudaMemcpy((void*)(dataC), (void*)(dev_dataC), M*N*sizeof(int), cudaMemcpyDeviceToHost); for(int i=0;i<M*N;i++) std::cout<<dataC[i]<<" "; cudaFree((void*)dev_dataC); cudaFree((void*)dev_C); free(C); free(dataC); cudaFree((void*)dev_dataB); cudaFree((void*)dev_B); free(B); free(dataB); cudaFree((void*)dev_dataA); cudaFree((void*)dev_A); free(A); free(dataA); getchar(); }
博客恢复更新,慢慢的积累吧
时间: 2024-10-27 16:57:29