bennycen — Tue, 26 Jul 2011 09:01:00 GMT

�q�几天研�I�了一下CUDA�Q�发现其�q�行的思想和普通的CPU多线�E�思想不太一��_��但还是挺不错。主要是��Q务划分成一个个block�Q�然后每个block里面再划分成�l�的�U�程。然后每个线�E�做自己做的
事情。这�U��ƈ行思想很适用于像矩阵�q�算�q�些元素与元素之间的�q�算�q�不耦合得很厉害�Q�但整体数据很大的情况，�q�只是我对CUDA的初步感觉�?br />矩阵�怹�的CPU�E�序如下�Q?/div>

//C = A*B
void MatrixMulCPU(float* _C,const float *_A,const float *_B,int _wa,int _ha,int _wb)
{
    float sum = 0;
    for (int i = 0; i < _ha; ++i)
    {
        for (int j = 0; j < _wb; ++j)
        {
            sum = 0;
            for (int k = 0; k < _wa; ++k)
            {
                sum += (float)_A[i*_wa+k]*(float)_B[k*_wb+ j];
            }
            _C[i*_wb+j] = (float)sum;
        }
    }
}

从上面可以看出，C(i,j) = sum { A(i,k)*B(k,j) } 0<=k < _wa;耦合�E�度很小�Q�所以我们可以通过划分区域的方法，让每个线�E�负责一个区域�?br />怎么划分呢？首先最初的��x��是让每一个线�E�计��一个C(i,j)�Q�那么估��一下，应该需要height_c*width_c�Q�也��是ha*wb个线�E�。进一步，我们��矩阉|��一个大�Ҏ��Grid划分�Q�如果一�?br />�Ҏ��Grid大小�?6*16�Q�那么矩�?0*48的可以表�C�Zؓ5(*16) * 3(*16)�Q�即16*16个大格子(block)�Q�每一个格子内�Q�自然就�?height_c/16) *(width_c/16)个线�E�了�?br />好了�Q�划分完后，内核代码如下�Q?/div>计算版本0�Q?br />

__global__ void matrix_kernel_0(float* _C,const float* _A,const float *_B,int _wa,int _wb)
{
    float sum = 0;
    //扑և�该线�E�所在的行列
    int row = blockIdx.y*blockDim.y + threadIdx.y;
    int col = blockIdx.x*blockDim.x + threadIdx.x;

    //�U�程Thread(row,col)负责计算C(row,col)
    for (int i = 0; i < _wa; ++i)
    {
        sum += _A[row*_wa + i]*_B[i*_wb + col];
    }
    _C[row*_wb + col] = sum;
}

另外一�U�思�\�Q�我们不让每一个线�E�完整计��一个C(i,j)�Q�通过C(i,j) = sum { A(i,k)*B(k,j) }发现�Q�我们还可以再细度划分：
Csub(i,j) = sum{A(i,ksub+offsetA)*B(ksub+offsetB,j)} 0<=ksub < blockSize
C(i,j) = sum{Csub(i,j)}
��是把矩阵分成n*n个大的子块，然后每一个block负责计算子块i �?子块j的子乘积�Q�计��完毕后加�v来则可。这里主要��用了�׃�n昑֭�作优化�?/div>
计算版本1�Q?br />

__global__ void matrix_kernel_1(float* _C,const float* _A,const float *_B,int _wa,int _wb)
{
    int bx = blockIdx.x;
    int by = blockIdx.y;
    int tx = threadIdx.x;
    int ty = threadIdx.y;

    //该block要处理的A
    int aBegin = _wa*(by*BLOCK_SIZE);//A(0,by)
    int aEnd = aBegin + _wa - 1;
    int aStep = BLOCK_SIZE;//offsetA

    int bBegin = BLOCK_SIZE*bx;//B(bx,0)
    int bStep = BLOCK_SIZE*_wb;//offsetB

    float cSub = 0;
    for (int a = aBegin,b = bBegin; a <= aEnd; a += aStep,b += bStep)
    {
        __shared__ float As[BLOCK_SIZE][BLOCK_SIZE];
        __shared__ float Bs[BLOCK_SIZE][BLOCK_SIZE];
        //每个�U�程负责一个元素拷�?/span>
        As[ty][tx] = _A[a + _wa*ty + tx];
        Bs[ty][tx] = _B[b + _wb*ty + tx];

        __syncthreads();

        //每个�U�程负责计算一个子块i �?nbsp;子块j的子乘积
        for (int k = 0; k < BLOCK_SIZE; ++k)
        {
            cSub += As[ty][k]*Bs[k][tx];
        }

        __syncthreads();
    }

    //全局地址�Q�向全局寄存器写回去
    //一个线�E�负责一个元素，一个block负责一个子�?/span>
    int cIndex = (by*BLOCK_SIZE + ty)*_wb + (bx*BLOCK_SIZE + tx);
    _C[cIndex] = cSub;
}

最后写一个面向Host的接口函敎ͼ�

void matrixMulGPU(float* _C,const float *_A,const float *_B,int _wa,int _ha,int _wb)
{
    float* d_a = myNewOnGPU<float>(_wa*_ha);
    float* d_b = myNewOnGPU<float>(_wb*_wa);
    float* d_c = myNewOnGPU<float>(_wb*_ha);
    copyFromCPUToGPU(_A,d_a,_wa*_ha);
    copyFromCPUToGPU(_B,d_b,_wb*_wa);
    dim3 threads(BLOCK_SIZE,BLOCK_SIZE);
    dim3 blocks(WC/BLOCK_SIZE,HC/BLOCK_SIZE);
    matrix_kernel_0<<<blocks,threads>>>(d_c,d_a,d_b,_wa,_wb);
    cudaThreadSynchronize();
    copyFromGPUToCPU(d_c,_C,_wb*_ha);

    myDeleteOnGPU(d_a);
    myDeleteOnGPU(d_b);
    myDeleteOnGPU(d_c);
}

调用的主函数如下�Q?br />

#include <stdio.h>
#include <cuda_runtime.h>
#include <cutil.h>
#include <cutil_inline.h>
#include <stdlib.h>
#include <time.h>
#include <math.h>
#include <string.h>
#include <Windows.h>
#include "CUDACommon.h"
#include "MatrixMulCPU.h"
#include "MatrixMulGPU.h"

void randomInit(float* _data,int _size)
{
    for (int i = 0; i < _size; ++i)
    {
        _data[i] = rand()/(float)RAND_MAX;
    }
}

bool checkError(const float* _A,const float* _B,int _size)
{
    for (int i = 0 ; i < _size; ++i)
    {
        if (fabs(_A[i] - _B[i]) > 1.0e-3)
        {
            printf("%f \t %f\n",_A[i],_B[i]);
            return false;
        }
    }
    return true;
}

int main(int argc, char* argv[])
{
    srand(13);
    if(!InitCUDA()) {
        return 0;
    }

    float* A = myNewOnCPU<float>(WA*HA);
    float* B = myNewOnCPU<float>(WB*HB);
    randomInit(A,WA*HA);
    randomInit(B,WB*HB);
    float* C = myNewOnCPU<float>(WC*HC);
    memset(C,0,sizeof(float)*WC*HC);

    float* C2 = myNewOnCPU<float>(WC*HC);
    memset(C2,0,sizeof(float)*WC*HC);

    unsigned int tick1 = GetTickCount();
    MatrixMulCPU(C2,A,B,WA,HA,WB);
    printf("CPU use Time : %dms\n",GetTickCount() - tick1);
    unsigned int timer = 0;
    cutilCheckError(cutCreateTimer(&timer));
    cutilCheckError(cutStartTimer(timer));
    {
        matrixMulGPU(C,A,B,WA,HA,WB);
    }
    cutilCheckError(cutStopTimer(timer));
    printf("GPU use time: %f (ms) \n", cutGetTimerValue(timer));
    cutilCheckError(cutDeleteTimer(timer));

    if (checkError(C,C2,WC*HC))
    {
        printf("Accept\n");
    }
    else
    {
        printf("Worng Answer\n");
    }

    myDeleteOnCPU(A);
    myDeleteOnCPU(B);
    myDeleteOnCPU(C);
    myDeleteOnCPU(C2);

    return 0;
}

�q�算�l�果如下�Q?br />版本0�Q?br />

版本1�Q?br />

可以看出�Q�GPU�q�行性能比CPU好很多，而且版本1优于版本0

整个工程下蝲�Q?a href="/Files/bennycen/CUDAMatrixMul.rar">/Files/bennycen/CUDAMatrixMul.rar

bennycen 2011-07-26 17:01 发表评论

国产精品激情电影,久久免费黄色,国产精品免费福利