Parallel processing 推力总和的值::减少值不正确
我一直在尝试实现一些需要在推力::设备ptr上调用reduce的代码,结果与处理大值时的CPU实现不一致。我必须处理大的价值观。那么有没有办法: 我的代码:Parallel processing 推力总和的值::减少值不正确,parallel-processing,cuda,nvidia,hpc,thrust,Parallel Processing,Cuda,Nvidia,Hpc,Thrust,我一直在尝试实现一些需要在推力::设备ptr上调用reduce的代码,结果与处理大值时的CPU实现不一致。我必须处理大的价值观。那么有没有办法: 我的代码: #include <cuda_runtime_api.h> #include <stdio.h> #include <thrust/host_vector.h> #include <thrust/device_vector.h> #include <
#include <cuda_runtime_api.h>
#include <stdio.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <iostream>
#define NZ 412//
#define NX 402//
using namespace std;
using real =double;
void allocate_array_2d(real**& preal, const int dim1, const int dim2) {
// Contiguous allocation of 2D arrays
preal = new real * [dim1];
preal[0] = new real[dim1 * dim2];
for (int i = 1; i < dim1; i++) preal[i] = preal[i - 1] + dim2;
for (int i = 0; i < dim1; i++) {
for (int j = 0; j < dim2; j++) {
preal[i][j] = 0;
}
}
}
#define cudaCheckError(code) \
{ \
if ((code) != cudaSuccess) { \
fprintf(stderr, "Cuda failure %s:%d: '%s' \n", __FILE__, __LINE__, \
cudaGetErrorString(code)); \
} \
}
int main()
{
real** a;
std::cout.precision(30);
allocate_array_2d(a, NZ, NX);//input array
for (int i = 0; i < NZ; i++) {
for (int j = 0; j < NX; j++) {
a[i][j] = 2.14748e+09;
}
}
real* da;
cudaCheckError(cudaMalloc(&da, NZ * NX * sizeof(real)));
cudaCheckError(cudaMemcpy(da,a[0], NZ * NX * sizeof(real),cudaMemcpyHostToDevice));
///************************
//CUDA KERNELS ARE HERE
// REMOVED FOR CLEAR QUESTION
///*************************
real sum1=0;
thrust::device_ptr<real> dev_ptr = thrust::device_pointer_cast(da);
sum1 = thrust::reduce(dev_ptr, dev_ptr+NZ*NX, 0, thrust::plus<real>());
cout<<" \nsum gpu "<< sum1<<"\n";
real sum2=0;
////////CPU PART DOING SAME THING//////
for (int i = 0; i < NZ; i++) {
for (int j = 0; j < NX; j++) {
sum2 += a[i][j];
}
}
cout<<"\nsum cpu "<< sum2<<"\n";
if((sum2-sum1)<0.001)
std::cout << "\nSUCESS "<< "\n";
else
std::cout << "\nFailure & by "<<sum2-sum1<< "\n";
}
#包括
#包括
#包括
#包括
#包括
#定义新西兰412//
#定义NX 402//
使用名称空间std;
使用实=双;
void allocate_array_2d(实数**和前置数,常量整数dim1,常量整数dim2){
//二维阵列的连续分配
preal=新的实数*[dim1];
前[0]=新实[dim1*dim2];
对于(inti=1;iinit
值中:
sum1 = thrust::reduce(dev_ptr, dev_ptr+NZ*NX, 0, thrust::plus<real>());
^
sum1=推力::减小(dev_ptr,dev_ptr+NZ*NX,0,推力::plus());
^
该常数的类型为整数类型。如果将其更改为双精度常数:
sum1 = thrust::reduce(dev_ptr, dev_ptr+NZ*NX, 0.0, thrust::plus<real>());
sum1=推力::减小(dev_ptr,dev_ptr+NZ*NX,0.0,推力::plus());
根据我的测试,您可以在CPU和GPU之间获得匹配结果。(您也可以将常量转换为real
类型:(real)0
,然后使用它,还有其他方法可以解决这个问题,例如放弃使用init值和二进制op。)