CUDA推力中的阵列长度

我的CUDA内核使用了按键推进、排序和缩减。 当我使用大于460的数组时，它开始显示不正确的结果

有人能解释这种行为吗？还是与我的机器有关？

尽管大小不同，排序仍能正常工作，但是REDUCE_BY_键的工作情况不好。并返回不正确的结果

有关代码的更多详细信息， 我有4个数组 1） 输入定义为完整序列阵列的键。 2） 在内核中定义的初始值为1的输入值。 3） 输出键用于保存输入键的不同值 4） 输出值是保存与同一输入键对应的输入值之和

有关reduce_by_密钥的更多说明，请访问此页面：

这是我的密码：

#include <cstdlib>
#include <stdlib.h>
#include <stdio.h>
#include <iostream>
#include <vector>
#include <fstream>
#include <string>
#include <cuda.h>
#include <cuda_runtime.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/sort.h>
#include <thrust/reduce.h>
#include <thrust/execution_policy.h>

using namespace std;
#define size 461

__global__ void calculateOccurances(unsigned int *input_keys,
            unsigned int *output_Values) {
    int tid = threadIdx.x;

    const int N = size;
    __shared__ unsigned int input_values[N];

    unsigned int outputKeys[N];

    int i = tid;
    while (i < N) {
            if (tid < N) {
                    input_values[tid] = 1;
            }
            i += blockDim.x;
    }
    __syncthreads();

    thrust::sort(thrust::device, input_keys, input_keys + N);

    thrust::reduce_by_key(thrust::device, input_keys, input_keys + N,
                    input_values, outputKeys, output_Values);

    if (tid == 0) {
            for (int i = 0; i < N; ++i) {
                    printf("%d,", output_Values[i]);
            }
    }

}

int main(int argc, char** argv) {

    unsigned int wholeSequenceArray[size] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
                    11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                    10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8,
                    9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7,
                    8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6,
                    7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5,
                    6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4,
                    5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3,
                    4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2,
                    3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1,
                    2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
                    1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
                    20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
                    19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
                    18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
                    17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
                    16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
                    15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
                    14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
                    13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
                    12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
                    11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                    10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8,
                    9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7,
                    8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,1 };

    cout << "wholeSequenceArray:" << endl;
    for (int i = 0; i < size; i++) {
            cout << wholeSequenceArray[i] << ",";
    }

    cout << "\nStart C++ Array New" << endl;
    cout << "Size of Input:" << size << endl;

    cudaDeviceProp prop;
    cudaGetDeviceProperties(&prop, 0);
    printf("Max threads per block:  %d\n", prop.maxThreadsPerBlock);

    unsigned int counts[size];
    unsigned int *d_whole;
    unsigned int *d_counts;

    cudaMalloc((void**) &d_whole, size * sizeof(unsigned int));
    cudaMalloc((void**) &d_counts, size * sizeof(unsigned int));

    cudaMemcpy(d_whole, wholeSequenceArray, size * sizeof(unsigned int),
                    cudaMemcpyHostToDevice);

    calculateOccurances<<<1, size>>>(d_whole, d_counts);

    cudaMemcpy(counts, d_counts, size * sizeof(unsigned int),
                    cudaMemcpyDeviceToHost);

    cout << endl << "Counts" << endl << endl;
    for (int i = 0; i < size; ++i) {
            cout << counts[i] << ",";
    }
    cout << endl;

    cudaFree(d_whole);
}

#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
使用名称空间std；
#定义尺寸461
__全局无效计算精度（无符号整数*输入键，
无符号整数*输出值）{
int tid=threadIdx.x；
常量int N=大小；
__共享的无符号整数输入值[N]；
无符号整数输出键[N]；
int i=每日三次；
而（icout在内核中调用推力算法时，推力算法从每个CUDA线程整体调度。因此，您的代码对相同的数据执行461次排序操作（从每个CUDA内核线程执行一次）在同一个位置。这意味着在排序操作期间，当每个线程移动数据时，它们将相互踩在一起

如果您只想使用您在问题中概述的方法计算数字的出现次数（有效地进行历史编程），并且您想使用推力，那么您根本不需要编写CUDA内核

如果您真的想在CUDA内核中（正确地）执行此操作，则有必要将推力操作（按_键排序和减少_）限制为仅从单个线程执行（甚至此方法也将限制为单个块）

我真的不认为第二个（CUDA内核）这种方法很有意义，但为了完整性，我修改了您的代码，以包含每个方法的正确示例。请注意，一旦执行了缩减，打印每个数组中的所有461个条目就没有意义了，因此为了清晰起见，我将打印输出限制为每个数组中的前25个条目：

$ cat t91.cu
#include <cstdlib>
#include <stdlib.h>
#include <stdio.h>
#include <iostream>
#include <vector>
#include <fstream>
#include <string>
#include <cuda.h>
#include <cuda_runtime.h>
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#include <thrust/sort.h>
#include <thrust/reduce.h>
#include <thrust/execution_policy.h>
#include <thrust/iterator/constant_iterator.h>

using namespace std;
#define size 461

__global__ void calculateOccurances(unsigned int *input_keys,
            unsigned int *output_Values) {
    int tid = threadIdx.x;

    const int N = size;
    __shared__ unsigned int input_values[N];

    unsigned int outputKeys[N];

    int i = tid;
    while (i < N) {
            if (tid < N) {
                    input_values[tid] = 1;
            }
            i += blockDim.x;
    }
    __syncthreads();
    if (tid == 0){
      thrust::sort(thrust::device, input_keys, input_keys + N);

      thrust::reduce_by_key(thrust::device, input_keys, input_keys + N,
                    input_values, outputKeys, output_Values);
      }

    if (tid == 0) {
    printf("from kernel:\n");
            for (int i = 0; i < 25; ++i) {
                    printf("%d,", output_Values[i]);
            }
    }

}

int main(int argc, char** argv) {

    unsigned int wholeSequenceArray[size] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
                    11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                    10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8,
                    9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7,
                    8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6,
                    7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5,
                    6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4,
                    5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3,
                    4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2,
                    3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1,
                    2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
                    1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
                    20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
                    19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
                    18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
                    17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
                    16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
                    15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
                    14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
                    13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
                    12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
                    11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8, 9,
                    10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7, 8,
                    9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 1, 2, 3, 4, 5, 6, 7,
                    8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,1 };

    cout << "wholeSequenceArray:" << endl;
    for (int i = 0; i < size; i++) {
            cout << wholeSequenceArray[i] << ",";
    }

    cout << "\nStart C++ Array New" << endl;
    cout << "Size of Input:" << size << endl;

    cudaDeviceProp prop;
    cudaGetDeviceProperties(&prop, 0);
    printf("Max threads per block:  %d\n", prop.maxThreadsPerBlock);

//just using thrust

    thrust::device_vector<int> d_seq(wholeSequenceArray, wholeSequenceArray+size);
    thrust::device_vector<int> d_val_out(size);
    thrust::device_vector<int> d_key_out(size);

    thrust::sort(d_seq.begin(), d_seq.end());
    int rsize = thrust::get<0>(thrust::reduce_by_key(d_seq.begin(), d_seq.end(), thrust::constant_iterator<int>(1), d_key_out.begin(), d_val_out.begin())) - d_key_out.begin();
    std::cout << "rsize:" << rsize <<  std::endl;
    std::cout << "Thrust keys:" << std::endl;
    thrust::copy_n(d_key_out.begin(), rsize, std::ostream_iterator<int>(std::cout, ","));
    std::cout << std::endl << "Thrust vals:" << std::endl;
    thrust::copy_n(d_val_out.begin(), rsize, std::ostream_iterator<int>(std::cout, ","));
    std::cout << std::endl;


// in a cuda kernel


    unsigned int counts[size];
    unsigned int *d_whole;
    unsigned int *d_counts;

    cudaMalloc((void**) &d_whole, size * sizeof(unsigned int));
    cudaMalloc((void**) &d_counts, size * sizeof(unsigned int));

    cudaMemcpy(d_whole, wholeSequenceArray, size * sizeof(unsigned int),
                    cudaMemcpyHostToDevice);

    calculateOccurances<<<1, size>>>(d_whole, d_counts);

    cudaMemcpy(counts, d_counts, size * sizeof(unsigned int),
                    cudaMemcpyDeviceToHost);

    std::cout << "from Host:" << std::endl;
    cout << endl << "Counts" << endl << endl;
    for (int i = 0; i < 25; ++i) {
            cout << counts[i] << ",";
    }
    cout << endl;

    cudaFree(d_whole);
}
$ nvcc -arch=sm_61 -o t91 t91.cu
$ ./t91
wholeSequenceArray:
1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,1,
Start C++ Array New
Size of Input:461
Max threads per block:  1024
rsize:20
Thrust keys:
1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,
Thrust vals:
24,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,
from kernel:
24,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,526324,526325,526325,526327,526329,from Host:

Counts

24,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,23,526324,526325,526325,526327,526329,
$

$cat t91.cu
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
使用名称空间std；
#定义尺寸461
__全局无效计算精度（无符号整数*输入键，
无符号整数*输出值）{
int tid=threadIdx.x；
常量int N=大小；
__共享的无符号整数输入值[N]；
无符号整数输出键[N]；
int i=每日三次；
而（i