CUDA + 使用 C 计算 int 元素出现次数

Question

在主机端，我正在读取一个 128 x 128 整数数组，其随机值介于 0-31 之间。我有一个存储值 0-31 的 Occurrences 数组，然后在设备上我试图执行一个内核，循环遍历 128 x 128 数组中的值，然后计算 0-31 出现的次数。

我遇到了如何在 CUDA 中拆分 blocks/threads 以及如何让内核向主机提供通信并打印出每个 element.This 的出现次数的问题我第一次使用 CUDA，如果有任何建设性的建议，我将不胜感激！到目前为止，这是我的代码：

 #include <stdio.h>
#include <stdlib.h>
#include <cuda.h>


#define MAXR 16
#define MAXC 16
#define N 256
__global__ void count(int *arrayONE_d, int *occurrences_d, int *occurrences_final_d) {

    int count = 0;
    //provide unique thread ID
    int idx = threadIdx.x + blockIdx.x * blockDim.x;

    int k;
    //for(k=0; k < 32;k++) {
    //  occurrences_d[k]=k;
//  }


    if(idx < N) {
        //for(k=0; k < MAXR*MAXC; k++) {
    for(int j=0; j<32; j++) {
            count =0;
        if(arrayONE_d[idx]==occurrences_d[j]){

            count+=1;
            occurrences_final_d[j] =count;
        }
        else {}


    }
    }
    //occurrences_final_d[0] = 77;
    }
}


int main(void) {



    //const int N = MAXR*MAXC;

    int arr1_h[MAXR][MAXC];
    //int *occurrences_h[0][32];
    //creating arrays for the device (GPU)
    //int *arr1_d;
    int occurrences_h[32];
    int *occurrences_d;

    int *occurrences_final_h[32] = {0};
    int *occurrences_final_d;

    int *arrayONE_h[256] = {0};
    int *arrayONE_d;

    int i, j;

    // allocating memory for the arrays on the device
    cudaMalloc( (void**) &arrayONE_d, MAXR*MAXC*sizeof(int)); // change to 16384 when using 128x128
    cudaMalloc( (void**) &occurrences_d,  32* sizeof(int));
    cudaMalloc( (void**) &occurrences_final_d, 32*sizeof(int));

    /*
    for(i=0; i < 32; i++) {

        occurrences_h[i] = i;

    }
/*
 *
 */
    //Reading in matrix from .txt file and storing it in arr1 on the host (CPU)
    FILE *fp;
    fp =fopen("arrays16.txt","r");

     // this loop takes the information from .txt file and puts it into arr1 matrix
    for(i=0;i<MAXR;i++) {


        for(j=0;j<MAXC;j++)
        {
            fscanf(fp,"%d\t", &arr1_h[i][j]);
        }

    }

    for(i=0;i<MAXR;i++) {
        printf("\n");

        for(j=0;j<MAXC;j++) {
            //printf("d\t", arr1_h[i][j]);
        }

        printf("\n\n");
    }


    int x,y;
    int z=0;
// this loop flattens the 2d array and makes it a 1d array of length MAXR*MAXC
    for(x=0;x<MAXR;x++)
    {
        for(y=0;y<MAXC;y++)
        {
            //  printf("**%d   ",arr1_h[x][y]);

            arrayONE_h[z]= &arr1_h[x][y];
            z++;

        }
    }


    for(x=0; x < 256; x++) {
        printf("%d\n", *arrayONE_h[x]);
        //return 0;

    }

    int length = sizeof(arrayONE_h)/sizeof(arrayONE_h[0]);

    printf("\n\n");
    printf("**LENGTH = %d", length);

    // copying the arrays/memory from the host to the device (GPU)
    cudaMemcpy(arrayONE_d, &arrayONE_h, MAXR*MAXC*sizeof(int), cudaMemcpyHostToDevice);
    cudaMemcpy(occurrences_d, &occurrences_h, 32*sizeof(int), cudaMemcpyHostToDevice);
    cudaMemcpy(occurrences_final_d, &occurrences_final_h, 32*sizeof(int), cudaMemcpyHostToDevice);

    // how many blocks we will allocate
    //dim3 DimGrid();
    //how many threads per block we will allocate
    dim3 DimBlock(256);

    //kernel launch against the GPU
    count<<<1, DimBlock>>>(arrayONE_d,occurrences_d,occurrences_final_d);

    //copy the arrays post-computation from the device back to the host (CPU)
    cudaMemcpy(&occurrences_final_h, occurrences_final_d, 32*sizeof(int), cudaMemcpyDeviceToHost);
    cudaMemcpy(&occurrences_h, occurrences_d, 32*sizeof(int), cudaMemcpyDeviceToHost);

    // some error checking - run this with cuda-memcheck when executing your code
    cudaError_t errSync  = cudaGetLastError();
    cudaError_t errAsync = cudaDeviceSynchronize();
    if (errSync != cudaSuccess)
        printf("Sync kernel error: %s\n", cudaGetErrorString(errSync));
    if (errAsync != cudaSuccess)
        printf("Async kernel error: %s\n", cudaGetErrorString(errAsync));

    //free up the memory of the device arrays
    cudaFree(arrayONE_d);
    cudaFree(occurrences_d);
    cudaFree(occurrences_final_d);

    //print out the number of occurrences of each 0-31 value
    for(i=0;i<32;i++) {
        printf("\n");

        printf("%d\n",occurrences_final_h[i]);

    }

}

Answer 1

正如我在评论中提到的，您对指针的理解存在缺陷。我已经在您的代码中的许多地方进行了更改以解决此问题。我用评论 // mod 标记了其中的大部分，但我可能遗漏了一些。

此外，当多个线程可以更新同一位置时，您的内核根本无法跟踪元素。解决这个问题的一种方法是使用原子（我已经演示过）。还有各种其他方法，例如并行缩减，但其中 none 是对内核的微不足道的更改。此外，您的内核逻辑在几个方面被破坏了。

接下来是我可以对您的代码进行最少数量的修改以获得合理的结果。您可以使用一些编译开关来探索不同的内核行为：

没有开关 - 接近你的内核，但它不能正常工作
-DUSE_ATOMICS 将演示对内核的修改以使其正确计数。
-DUSE_ALT_KERNEL 探索了一种不同的内核逻辑方法：为每个直方图 bin 分配一个线程，并让每个线程遍历整个数组，跟踪属于该 bin 的元素。由于只有一个线程写入每个 bin 结果，因此不需要原子。然而，我们只能拥有与垃圾箱一样多的线程（通过这种微不足道的实现）。没有太多困难，这种方法可能会扩展到每个 bin 一个 warp，using warp shuffle to do a final warp-level reduction 在让一个线程将最终结果写入 bin 之前。这会在一定程度上提高内存访问效率。但是，这也会给内核带来您可能尚未了解的复杂性。

代码如下：

$ cat t316.cu
 #include <stdio.h>
#include <stdlib.h>
#include <cuda.h>


#define MAXR 16
#define MAXC 16
#define BINS 32
#define N (MAXR*MAXC)
__global__ void count(int *arrayONE_d, int *occurrences_d, int *occurrences_final_d) {

    //provide unique thread ID
    int idx = threadIdx.x + blockIdx.x * blockDim.x;
#ifndef USE_ALT_KERNEL
    if(idx < N) {
        //for(k=0; k < MAXR*MAXC; k++) {
    for(int j=0; j<32; j++) {
        if(arrayONE_d[idx]==occurrences_d[j]){
#ifndef USE_ATOMICS
            occurrences_final_d[j]++;
#else
         atomicAdd(occurrences_final_d+j, 1);
#endif

        }
        else {}


    }
    }
#else
   // use one thread per histo bin
   if (idx < BINS){
     int count = 0;
     int myval = occurrences_d[idx];
     for (int i = 0; i < N; i++) if (arrayONE_d[i] == myval) count++;
     occurrences_final_d[idx] = count;
     }

#endif
    }


int main(void) {



    //const int N = MAXR*MAXC;

    int arr1_h[MAXR][MAXC];
    //int *occurrences_h[0][32];
    //creating arrays for the device (GPU)
    //int *arr1_d;
    int occurrences_h[32]; // mod
    int *occurrences_d;

    int occurrences_final_h[32] = {0};  // mod
    int *occurrences_final_d;

    int arrayONE_h[256] = {0};  // mod
    int *arrayONE_d;

    int i, j;

    // allocating memory for the arrays on the device
    cudaMalloc( (void**) &arrayONE_d, MAXR*MAXC*sizeof(int)); // change to 16384 when using 128x128
    cudaMalloc( (void**) &occurrences_d,  32* sizeof(int));
    cudaMalloc( (void**) &occurrences_final_d, 32*sizeof(int));

    /*
    for(i=0; i < 32; i++) {

        occurrences_h[i] = i;

    }
 */
    //Reading in matrix from .txt file and storing it in arr1 on the host (CPU)

//    FILE *fp;
//    fp =fopen("arrays16.txt","r");

     // this loop takes the information from .txt file and puts it into arr1 matrix
    for(i=0;i<MAXR;i++) {


        for(j=0;j<MAXC;j++)
        {
//            fscanf(fp,"%d\t", &arr1_h[i][j]);
              arr1_h[i][j] = j;  // mod
        }

    }

    for(i=0;i<MAXR;i++) {

        for(j=0;j<MAXC;j++) {
            //printf("d\t", arr1_h[i][j]);
        }

    }


    int x,y;
    int z=0;
// this loop flattens the 2d array and makes it a 1d array of length MAXR*MAXC
    for(x=0;x<MAXR;x++)
    {
        for(y=0;y<MAXC;y++)
        {
            //  printf("**%d   ",arr1_h[x][y]);

            arrayONE_h[z]= arr1_h[x][y];  // mod
            z++;

        }
    }


    for(x=0; x < 256; x++) {
//        printf("%d\n", arrayONE_h[x]);  // mod
        //return 0;

    }

    int length = sizeof(arrayONE_h)/sizeof(arrayONE_h[0]);

    printf("**LENGTH = %d\n", length);

    // copying the arrays/memory from the host to the device (GPU)
    cudaMemcpy(arrayONE_d, arrayONE_h, MAXR*MAXC*sizeof(int), cudaMemcpyHostToDevice);  //mod
    cudaMemcpy(occurrences_d, occurrences_h, 32*sizeof(int), cudaMemcpyHostToDevice);   // mod
    cudaMemcpy(occurrences_final_d, occurrences_final_h, 32*sizeof(int), cudaMemcpyHostToDevice); // mod

    // how many blocks we will allocate
    //dim3 DimGrid();
    //how many threads per block we will allocate
#ifndef USE_ALT_KERNEL
    dim3 DimBlock(N);
#else
    dim3 DimBlock(BINS);
#endif
    //kernel launch against the GPU
    count<<<1, DimBlock>>>(arrayONE_d,occurrences_d,occurrences_final_d);

    //copy the arrays post-computation from the device back to the host (CPU)
    cudaMemcpy(occurrences_final_h, occurrences_final_d, 32*sizeof(int), cudaMemcpyDeviceToHost); // mod
    cudaMemcpy(occurrences_h, occurrences_d, 32*sizeof(int), cudaMemcpyDeviceToHost);  // mod

    // some error checking - run this with cuda-memcheck when executing your code
    cudaError_t errSync  = cudaGetLastError();
    cudaError_t errAsync = cudaDeviceSynchronize();
    if (errSync != cudaSuccess)
        printf("Sync kernel error: %s\n", cudaGetErrorString(errSync));
    if (errAsync != cudaSuccess)
        printf("Async kernel error: %s\n", cudaGetErrorString(errAsync));

    //free up the memory of the device arrays
    cudaFree(arrayONE_d);
    cudaFree(occurrences_d);
    cudaFree(occurrences_final_d);

    //print out the number of occurrences of each 0-31 value
    for(i=0;i<32;i++) {
        printf("%d ",occurrences_final_h[i]);

    }
    printf("\n");
}
$ nvcc -o t316 t316.cu
$ cuda-memcheck ./t316
========= CUDA-MEMCHECK
**LENGTH = 256
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
========= ERROR SUMMARY: 0 errors
$ nvcc -o t316 t316.cu -DUSE_ATOMICS
$ ./t316
**LENGTH = 256
16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16
$ nvcc -o t316 t316.cu -DUSE_ALT_KERNEL
$ cuda-memcheck ./t316
========= CUDA-MEMCHECK
**LENGTH = 256
16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16 16
========= ERROR SUMMARY: 0 errors
$

在上面的输出中，我们看到基本内核产生了不正确的结果。原子内核和备用内核产生正确的结果

（您的代码已修改为使用合成数据，因此不需要打开文件。）

CUDA + 使用 C 计算 int 元素出现次数

CUDA + count int element occurrences using C

c

cuda

gpu

distributed-computing