数值积分 - 如何并行化?

fac*_*hpc 5 c algorithm parallel-processing opencl

我开始使用OpenCL,我可以看到添加向量示例并理解它.但我在考虑梯形方法.这是[a,b]中x ^ 2的积分计算的代码(C).

double f(double x)
{
    return x*x;
}

double Simple_Trap(double a, double b)
{
    double fA, fB;
    fA = f(a);
    fB = f(b);
    return ((fA + fB) * (b-a)) / 2;
}

double Comp_Trap( double a, double b)
{
    double Suma = 0;
    double i = 0;
    i = a + INC;
    Suma += Simple_Trap(a,i);
    while(i < b)
    {
        i+=INC;
        Suma += Simple_Trap(i,i + INC);
    }
    return Suma;
}
Run Code Online (Sandbox Code Playgroud)

问题是如何使用梯形方法获得用于积分计算的内核?


所以,我正在思考这个想法:partials [i] = integration(a,a + offset),然后创建一个内核来计算部分的总和,如Patrick87所述.

但是,这是最好的方法吗?

mfa*_*mfa 2

这就是我的想法。我没有对该内核进行端到端测试。当我有更多时间时,我会进行更新。

comp_trap 是基于上面代码的基本分而治之方法。comp_trap_multi 通过让每个工作项划分其子部分来提高准确性

您只需在主机中分配一组双精度数,以便每个工作组都有一个双精度数来返回其结果。这应该有助于减少您想要避免的向量分配。

如果有任何问题,请告诉我。

更新:

1) 将所有 double 引用更改为 float,因为 double 在 opencl 中是可选的

2) 将工作组大小硬编码为 64。该值在我的系统上是最佳值,应通过实验确定。我更喜欢对该值进行硬编码,而不是传递本地浮点数组来使用,因为无论如何,主机程序最终将只使用目标系统上的最佳值。

3)修正了一个错误的计算(a1是错误的,现在应该更好了)

/*
numerical-integration.cl
*/

float f(float x)
{
    return x*x;
}

float simple_trap(float a, float b)
{
    float fA, fB;
    fA = f(a);
    fB = f(b);
    return ((fA + fB) * (b-a)) / 2;
}

__kernel void comp_trap(
    float a,
    float b,
    __global float* sums)
{
/*
- assumes 1D global and local work dimensions
- each work unit will calculate 1/get_global_size of the total sum
- the 0th work unit of each group then accumulates the sum for the
group and stores it in __global * sums
- memory allocation: sizeof(sums) = get_num_groups(0) * sizeof(float)
- assumes local scratchpad size is at lease 8 bytes per work unit in the group
ie sizeof(wiSums) = get_local_size(0) * sizeof(float)
*/
    __local float wiSums[64];
    int l_id = get_local_id(0);

    //cumpute range for this work item is: a1, b1 
    float a1 = a+((b-a)/get_global_size(0))*get_global_id(0);
    float b1 = a1+(b-a)/get_global_size(0);

    wiSums[l_id] = simple_trap(a1,b1);

    barrier(CLK_LOCAL_MEM_FENCE);

    int i;
    if(l_id == 0){
        for(i=1;i<get_local_size(0);i++){
            wiSums[0] += wiSums[i];
        }
        sums[get_group_id(0)] = wiSums[0];
    }
}

__kernel void comp_trap_multi(
    float a,
    float b,
    __global float* sums,
    int divisions)
{
/*
- same as above, but each work unit further divides its range into
'divisions' equal parts, yielding a more accurate result
- work units still store only one sum in the local array, which is
used later for the final group accumulation
*/
    __local float wiSums[64];
    int l_id = get_local_id(0);

    float a1 = a+((b-a)/get_global_size(0))*get_global_id(0);
    float b1 = a1+(b-a)/get_global_size(0);
    float range;
    if(divisions > 0){
        range = (b1-a1)/divisions;
    }else{
        range = (b1-a1);
    }

    int i;
    wiSums[l_id] = 0;
    for(i=0;i<divisions;i++){
        wiSums[l_id] += simple_trap(a1+range*i,a1+range*(i+1));
    }

    barrier(CLK_LOCAL_MEM_FENCE);

    if(l_id == 0){
        for(i=1;i<get_local_size(0);i++){
            wiSums[0] += wiSums[i];
        }
        sums[get_group_id(0)] = wiSums[0];
    }
}
Run Code Online (Sandbox Code Playgroud)