本篇博文转载于https://www.cnblogs.com/1024incn/tag/CUDA/,仅用于学习。
前言
线程的组织形式对程序的性能影响是至关重要的,本篇博文主要以下面一种情况来介绍线程组织形式:
- 2D grid 2D block
线程索引
矩阵在memory中是row-major线性存储的:
在kernel里,线程的唯一索引非常有用,为了确定一个线程的索引,我们以2D为例:
- 线程和block索引
- 矩阵中元素坐标
- 线性global memory 的偏移
首先可以将thread和block索引映射到矩阵坐标:
ix = threadIdx.x + blockIdx.x * blockDim.x
iy = threadIdx.y + blockIdx.y * blockDim.y
之后可以利用上述变量计算线性地址:
idx = iy * nx + ix
上图展示了block和thread索引,矩阵坐标以及线性地址之间的关系,谨记,相邻的thread拥有连续的threadIdx.x,也就是索引为(0,0)(1,0)(2,0)(3,0)...的thread连续,而不是(0,0)(0,1)(0,2)(0,3)...连续,跟我们线代里玩矩阵的时候不一样。
现在可以验证出下面的关系:
thread_id(2,1)block_id(1,0) coordinate(6,1) global index 14 ival 14
下图显示了三者之间的关系:
代码
int main(int argc, char **argv) {printf("%s Starting...\n", argv[0]);// set up deviceint dev = 0;cudaDeviceProp deviceProp;CHECK(cudaGetDeviceProperties(&deviceProp, dev));printf("Using Device %d: %s\n", dev, deviceProp.name);CHECK(cudaSetDevice(dev)); // set up date size of matrixint nx = 1<<14;int ny = 1<<14;int nxy = nx*ny;int nBytes = nxy * sizeof(float);printf("Matrix size: nx %d ny %d\n",nx, ny);// malloc host memoryfloat *h_A, *h_B, *hostRef, *gpuRef;h_A = (float *)malloc(nBytes);h_B = (float *)malloc(nBytes);hostRef = (float *)malloc(nBytes);gpuRef = (float *)malloc(nBytes);// initialize data at host sidedouble iStart = cpuSecond();initialData (h_A, nxy);initialData (h_B, nxy);double iElaps = cpuSecond() - iStart;memset(hostRef, 0, nBytes);memset(gpuRef, 0, nBytes);// add matrix at host side for result checksiStart = cpuSecond();sumMatrixOnHost (h_A, h_B, hostRef, nx,ny);iElaps = cpuSecond() - iStart;// malloc device global memoryfloat *d_MatA, *d_MatB, *d_MatC;cudaMalloc((void **)&d_MatA, nBytes);cudaMalloc((void **)&d_MatB, nBytes);cudaMalloc((void **)&d_MatC, nBytes);// transfer data from host to devicecudaMemcpy(d_MatA, h_A, nBytes, cudaMemcpyHostToDevice);cudaMemcpy(d_MatB, h_B, nBytes, cudaMemcpyHostToDevice);// invoke kernel at host sideint dimx = 32;int dimy = 32;dim3 block(dimx, dimy);dim3 grid((nx+block.x-1)/block.x, (ny+block.y-1)/block.y);iStart = cpuSecond();sumMatrixOnGPU2D <<< grid, block >>>(d_MatA, d_MatB, d_MatC, nx, ny);cudaDeviceSynchronize();iElaps = cpuSecond() - iStart;printf("sumMatrixOnGPU2D <<<(%d,%d), (%d,%d)>>> elapsed %f sec\n", grid.x,grid.y, block.x, block.y, iElaps);// copy kernel result back to host sidecudaMemcpy(gpuRef, d_MatC, nBytes, cudaMemcpyDeviceToHost);// check device resultscheckResult(hostRef, gpuRef, nxy);// free device global memorycudaFree(d_MatA);cudaFree(d_MatB);cudaFree(d_MatC);// free host memoryfree(h_A);free(h_B);free(hostRef);free(gpuRef);// reset devicecudaDeviceReset();return (0);
}
编译运行:
$ nvcc -arch=sm_20 sumMatrixOnGPU-2D-grid-2D-block.cu -o matrix2D $ ./matrix2D
输出:
./a.out Starting... Using Device 0: Tesla M2070 Matrix size: nx 16384 ny 16384 sumMatrixOnGPU2D <<<(512,512), (32,32)>>> elapsed 0.060323 sec Arrays match.
接下来,我们更改block配置为32x16,重新编译,输出为:
sumMatrixOnGPU2D <<<(512,1024), (32,16)>>> elapsed 0.038041 sec
可以看到,性能提升了一倍,直观的来看,我们会认为第二个配置比第一个多了一倍的block所以性能提升一倍,实际上也确实是因为block增加了。但是,如果你继续增加block的数量,则性能又会降低:
sumMatrixOnGPU2D <<< (1024,1024), (16,16) >>> elapsed 0.045535 sec
下图展示了不同配置的性能;
关于性能的分析将在之后的博文中总结,现在只是了解下,本文在于掌握线程组织的方法。