Cuda 内核返回向量

Cuda kernel returning vectors

本文关键字：向量 返回 内核 Cuda      更新时间：2023-10-16

我有一个单词列表，我的目标是在一个非常长的短语中匹配每个单词。我在匹配每个单词时没有问题，我唯一的问题是返回一个包含每个匹配信息的结构向量。

在代码中：

typedef struct {
    int A, B, C; } Match;
__global__ void Find(veryLongPhrase * _phrase, Words * _word_list, vector<Match> * _matches)
{
    int a, b, c;
    [...] //Parallel search for each word in the phrase
    if(match) //When an occurrence is found
    {
        _matches.push_back(new Match{ A = a, B = b, C = c }); //Here comes the unknown, what should I do here???
    }
}
main()
{
    [...]
    veryLongPhrase * myPhrase = "The quick brown fox jumps over the lazy dog etc etc etc..."
    Words * wordList = {"the", "lazy"};
    vector<Match> * matches; //Obviously I can't pass a vector to a kernel
    Find<<< X, Y >>>(myPhrase, wordList, matches);
    [...]
}

我已经尝试过推力库，但没有任何成功，你能给我任何解决方案吗？

谢谢。

这样的东西应该可以工作（在浏览器中编码，而不是测试）：

// N is the maximum number of structs to insert
#define N 10000
typedef struct {
    int A, B, C; } Match;
__device__ Match dev_data[N];
__device__ int dev_count = 0;
__device__ int my_push_back(Match * mt) {
  int insert_pt = atomicAdd(&dev_count, 1);
  if (insert_pt < N){
    dev_data[insert_pt] = *mt;
    return insert_pt;}
  else return -1;}
__global__ void Find(veryLongPhrase * _phrase, Words * _word_list, vector<Match> * _matches)
{
    int a, b, c;
    [...] //Parallel search for each word in the phrase
    if(match) //When an occurrence is found
    {
        my_push_back(new Match{ A = a, B = b, C = c });    }
}

main()
{
    [...]
    veryLongPhrase * myPhrase = "The quick brown fox jumps over the lazy dog etc etc etc..."
    Words * wordList = {"the", "lazy"};
    Find<<< X, Y >>>(myPhrase, wordList);
    int dsize;
    cudaMemcpyFromSymbol(&dsize, dev_count, sizeof(int));
    vector<Match> results(dsize);
    cudaMemcpyFromSymbol(&(results[0]), dev_data, dsize*sizeof(Match));
    [...]
}

这将需要原子操作的计算能力为 1.1 或更高。

nvcc -arch=sm_11 ...

这是一个工作示例：

$ cat t347.cu
#include <iostream>
#include <vector>
// N is the maximum number of structs to insert
#define N 10000
typedef struct {
    int A, B, C; } Match;
__device__ Match dev_data[N];
__device__ int dev_count = 0;
__device__ int my_push_back(Match & mt) {
  int insert_pt = atomicAdd(&dev_count, 1);
  if (insert_pt < N){
    dev_data[insert_pt] = mt;
    return insert_pt;}
  else return -1;}
__global__ void Find()
{
    if(threadIdx.x < 10) //Simulate a found occurrence
    {
        Match a = { .A = 1, .B = 2, .C = 3 };
        my_push_back(a);    }
}

main()
{
    Find<<< 2, 256 >>>();
    int dsize;
    cudaMemcpyFromSymbol(&dsize, dev_count, sizeof(int));
    if (dsize >= N) {printf("overflow errorn"); return 1;}
    std::vector<Match> results(dsize);
    cudaMemcpyFromSymbol(&(results[0]), dev_data, dsize*sizeof(Match));
    std::cout << "number of matches = " << dsize << std::endl;
    std::cout << "A  =  " << results[dsize-1].A << std:: endl;
    std::cout << "B  =  " << results[dsize-1].B << std:: endl;
    std::cout << "C  =  " << results[dsize-1].C << std:: endl;
}
$ nvcc -arch=sm_11 -o t347 t347.cu
$ ./t347
number of matches = 20
A  =  1
B  =  2
C  =  3
$

请注意，在这种情况下，我的Match结果结构创建是不同的，并且我是通过引用传递的，但概念是相同的。

相关文章：