CUDA笔记3：代码实践



    本文用CUDA实现两矩阵相乘。下图为C=A*B的直观解释，C的每个元素为A对应行向量与B对应列向量的点积。可见，为了计算元素C(i,j)，需要访问全局内存A.width + B.height次。为了提升计算效率，矩阵乘法还可以用共享内存实现。首先按照Block大小将A,B,C拆分成子矩阵形式，那么C中一个Block内的所有Thread可以同步访问A,B对应的子区域（从全局内存拷贝到共享内存，共享内存设置成Block区域大小），然后进行子区域的乘积计算。实际上利用共享内存可以将全局访问次数降低为A.width/BLOCK_SIZE + B.height/BLOCK_SIZE.

                                      矩阵相乘原理图

                                            利用共享内存可以提高计算效率

以下Code参照SDK VectorAdd代码示例改下，用来计算两个1024*1024方阵的乘机，并且添加了CPU和GPU计时功能，可以方便对比运算时间。

/**
 * 两矩阵（方阵）相乘示例，
 使用共享内存 比 不使用共享内存 效率高出一倍
 */

#include <stdio.h>
#include <time.h>
// For the CUDA runtime routines (prefixed with "cuda_")
#include <cuda_runtime.h>

/**
 * CUDA Kernel Device code
 * 
 * Computes the matrix multiplication of A and B into C. 
 * without using shared memory
 */
__global__ void matrixMul_noSharedM(const float *A, const float *B, float *C, int N)
{
	int col = blockDim.x * blockIdx.x + threadIdx.x;
	int row = blockDim.y * blockIdx.y + threadIdx.y;

	if (col < N && row < N)
	{
		C[row*N + col] = 0;
		for (int i=0; i<N; i++)
		{
			C[row*N + col] += A[row*N + i] * B[i*N + col];
		}
	}
}

//using shared memory
__global__ void matrixMul_SharedM(const float *A, const float *B, float *C, int N)
{
	__shared__ float Mds[32][32];
	__shared__ float Nds[32][32];
	
	int col = blockDim.x * blockIdx.x + threadIdx.x;
	int row = blockDim.y * blockIdx.y + threadIdx.y;

	if (col < N && row < N)
	{
		float pValue = 0.0;
		for (int m=0; m<32; m++)
		{
			Mds[threadIdx.y][threadIdx.x] = A[row*N + m*32 + threadIdx.x];
			Nds[threadIdx.y][threadIdx.x] = B[(threadIdx.y+m*32)*N + col];

			__syncthreads(); //同步同一个block的所有thread
			for(int k = 0; k < 32; k++)
			{
				pValue += Mds[threadIdx.y][k] * Nds[k][threadIdx.x];
			}
			__syncthreads();
		}
		C[row*N + col] = pValue;
	}
}

/**
 * Host main routine
 */
int main(void)
{
    // Error code to check return values for CUDA calls
    cudaError_t err = cudaSuccess;

    // Define the squared matrix, and compute its size
    int numElements = 32*32;
    int size = numElements * numElements * sizeof(float);
	int NN = numElements * numElements;

    

    // Allocate the host input vector A
    float *h_A = new float[size];//(float *)malloc(size);

    // Allocate the host input vector B
    float *h_B = new float[size];//(float *)malloc(size);

    // Allocate the host output vector C
    float *h_C = new float[size];//(float *)malloc(size);

    // Verify that allocations succeeded
    if (h_A == NULL || h_B == NULL || h_C == NULL)
    {
        fprintf(stderr, "Failed to allocate host matrix!\n");
        exit(EXIT_FAILURE);
    }

    // Initialize the host input matrix
    for (int i = 0; i < NN; i++)
    {
        h_A[i] = rand()/(float)RAND_MAX * 10.0;
        h_B[i] = rand()/(float)RAND_MAX * 10.0; 
    }

    // Allocate the device input matrix A
    float *d_A = NULL;
    err = cudaMalloc((void **)&d_A, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device matrix A (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Allocate the device input matrix B
    float *d_B = NULL;
    err = cudaMalloc((void **)&d_B, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device matrix B (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Allocate the device output matrix C
    float *d_C = NULL;
    err = cudaMalloc((void **)&d_C, size);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to allocate device matrix C (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Copy the host input matrix A and B in host memory to the device input matrix in
    // device memory
   // printf("Copy input data from the host memory to the CUDA device\n");
    err = cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy matrix A from host to device (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    err = cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy matrix B from host to device (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Launch the Matrix multiplication CUDA Kernel
    dim3 dimGrid;
	dim3 dimBlock(32, 32);
	dimGrid.x  = numElements / dimBlock.x ;
	dimGrid.y  = numElements / dimBlock.y ;	

	cudaEvent_t start, stop; // measure GPU time
	float time;
	cudaEventCreate(&start);
	cudaEventCreate(&stop);
	cudaEventRecord(start, 0);

    //matrixMul_noSharedM<<<dimGrid, dimBlock>>>(d_A, d_B, d_C, numElements);
	matrixMul_SharedM<<<dimGrid, dimBlock>>>(d_A, d_B, d_C, numElements);

	cudaEventRecord(stop, 0);
	cudaEventSynchronize(stop);
	cudaEventElapsedTime(&time, start, stop); printf("GPU time: %f \n", time);
	cudaEventDestroy(start);
	cudaEventDestroy(stop);




    err = cudaGetLastError();

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to launch vectorAdd kernel (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Copy the device result matrix in device memory to the host result matrix
    // in host memory.
   // printf("Copy output data from the CUDA device to the host memory\n");
    err = cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to copy matrix C from device to host (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }


	// Compute the groundtrueth
	clock_t start2, finish;   // measure CPU time
    double  duration = 0.0;   
	start2 = clock();   

	float* GT_C = new float[NN];
	for (int r=0; r<numElements; r++)
	{
		for (int c=0; c<numElements; c++)
		{
			GT_C[r*numElements + c] = 0;
			for (int i=0; i<numElements; i++)
			{
				GT_C[r*numElements + c] += h_A[r*numElements + i] * h_B[i*numElements + c];
			}
		}
	}
	finish = clock();   
	duration = (double)(finish - start2); //输出单位ms  
	printf("CPU time: %f\n", duration);

	// Verify that the result vector is correct
	for (int i = 0; i < NN; ++i)
	{
		//printf("%d  %f  %f\n", i,  GT_C[i], h_C[i]);
		if (fabs(GT_C[i] - h_C[i]) > 1e-5)
		{
			fprintf(stderr, "Result verification failed at element %d!\n", i);
			exit(EXIT_FAILURE);
		}
	}
	printf("Test PASSED\n");

    // Free device global memory
    err = cudaFree(d_A);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device matrix A (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
    err = cudaFree(d_B);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device matrix B (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }
    err = cudaFree(d_C);

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to free device matrix C (error code %s)!\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    // Free host memory
    delete []h_A;
    delete []h_B;
    delete []h_C;

    // Reset the device and exit
    err = cudaDeviceReset();

    if (err != cudaSuccess)
    {
        fprintf(stderr, "Failed to deinitialize the device! error=%s\n", cudaGetErrorString(err));
        exit(EXIT_FAILURE);
    }

    printf("Done\n");
    return 0;
}

参考：http://blog.csdn.net/csgxy123/article/details/10018531