CUDA中的并行化,为每列分配线程

时间:2012-04-26 19:43:21

标签: arrays cuda

假设我有一个从MxN 2D矩阵转换的一维数组,我想并行化每一列并进行一些操作。如何为每列分配线程?

例如,如果我有一个3x3矩阵:

1  2  3

4  5  6

7  8  9

我想根据列#添加列中的每个数字(因此第1列将添加1,第2列将添加2 ....),然后它变为:

1+1   2+1   3+1

4+2   5+2   6+2

7+3   8+3   9+3

我如何在CUDA中执行此操作?我知道如何为数组中的所有元素分配线程,但我不知道如何为每个列分配线程。所以,我想要的是发送每一列(1,2,3)(4,5,6)(7,8,9)并进行操作。

2 个答案:

答案 0 :(得分:3)

在您的示例中,您将根据行添加数字。你知道矩阵的行/列长度(你知道它是MxN)。你能做的就是:

__global__ void MyAddingKernel(int* matrix, int M, int N)
{

    int gid = threadIdx.x + blockDim.x*blockIdx.x;
    //Let's add the row number to each element
    matrix[ gid ] += gid % M;
    //Let's add the column number to each element
    matrix[ gid ] += gid % N;

}

如果您想添加其他数字,可以执行以下操作:

matrix[ gid ] += my_col_number_function(gid%N);

答案 1 :(得分:1)

使用更好的网格布局来避免那些模运算。

对最新Cuda上64位范围的行使用唯一块索引。

让线程在所有元素的循环中迭代并添加唯一的线程索引!

如果计算数据在块(行)中是唯一的,则平铺输入数据是一种通用方法,尤其是对于更复杂的计算。

/*
 * @param tileCount 
 */
__global__ void addRowNumberToCells(int* inOutMat_g, 
    const unsigned long long int inColumnCount_s, 
    const int inTileCount_s)
{

    //get unique block index
    const unsigned long long int blockId = blockIdx.x //1D
        + blockIdx.y * gridDim.x //2D
        + gridDim.x * gridDim.y * blockIdx.z; //3D

    /* 
     * check column ranges in case kernel is called 
     * with more blocks then columns 
     * (since its block wide following syncthreads are safe)
     */
    if(blockId >= inColumnCount_s)
        return;

    //get unique thread index
    const unsigned long long int threadId = blockId * blockDim.x + threadIdx.x; 

    /*
     * calculate unique and 1 blockId
     * maybe shared memory is overhead 
     * but it shows concept if calculation is more complex
     */
    __shared__ unsigned long long int blockIdAnd1_s;
    if(threadIdx.x == 0)
        blockIdAnd1_s = blockId + 1;
    __sycnthreads();


    unsigned long long int idx;

    //loop over tiles
    for(int i = 0; i < inTileCount_s)
    {
        //calculate new offset for sequence thread writes
        idx = i * blockDim.x + threadIdx.x;
        //check new index range in case column count is no multiple of blockDim.x
        if(idx >= inColumnCount_s)
            break;
        inOutMat_g[idx] = blockIdAnd1_s;
    }

}

示例Cuda 2.0:

垫[131000] [1000]

必要的blockCount = 131000/65535 = 2 for blockDim.y四舍五入!

inTileCount_s = 1000/192 = 6四舍五入!

(每块192个主题= Cuda 2.0占用100个)

&lt;&lt;(65535,2,1),(192,1,1)&gt;&gt; addRowNumberToCells(mat,1000,6)