CUDA thrust库sort_by_key函数

问题描述

#include <stdio.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#define SIZE 6

__global__ void kernel1(int* key, int* value)
{
    for(int i = 0; i < SIZE; i++)
    {
        printf("key[%d] = %d value[%d] = %d\n", i, key[i], i, value[i]);
    }
    printf("\n");
}

int main(){
    int key[SIZE] = {24, 43, 27, 42, 25, 40};
    int value[SIZE] = {0, 1, 2, 3, 4, 5};
    int *d_key, *d_value;
    cudaMalloc((void**)&d_key, sizeof(int) * SIZE);
    cudaMalloc((void**)&d_value, sizeof(int) * SIZE);
    cudaMemcpy(d_key, key, sizeof(int) * SIZE, cudaMemcpyHostToDevice);
    cudaMemcpy(d_value, value, sizeof(int) * SIZE, cudaMemcpyHostToDevice);

    kernel1<<<1,1>>>(d_key, d_value);
    cudaDeviceSynchronize();

    //- sort
    thrust::device_ptr<int> t_key(d_key);
    thrust::device_ptr<int> t_value(d_value);
    thrust::sort_by_key(t_value, t_value + SIZE, t_key);
    cudaDeviceSynchronize();

    kernel1<<<1,1>>>(d_key, d_value);
    cudaDeviceSynchronize();
    return 0;
}

sort前后的输出结果分别为：

#- 预期结果                  运行结果
key[0] = 24 value[0] = 0    key[0] = 24 value[0] = 0
key[1] = 43 value[1] = 1    key[1] = 43 value[1] = 1
key[2] = 27 value[2] = 2    key[2] = 27 value[2] = 2
key[3] = 42 value[3] = 3    key[3] = 42 value[3] = 3
key[4] = 25 value[4] = 4    key[4] = 25 value[4] = 4
key[5] = 40 value[5] = 5    key[5] = 40 value[5] = 5

key[0] = 24 value[0] = 0    key[0] = 24 value[0] = 0
key[1] = 25 value[1] = 4    key[1] = 43 value[1] = 1
key[2] = 27 value[2] = 2    key[2] = 27 value[2] = 2
key[3] = 40 value[3] = 5    key[3] = 42 value[3] = 3
key[4] = 42 value[4] = 3    key[4] = 25 value[4] = 4
key[5] = 43 value[5] = 1    key[5] = 40 value[5] = 5