Cuda 如何将 char** 从内核复制到主机

Question

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <string.h>
#include <malloc.h>
#include <fstream>
#include <algorithm>
#include <time.h>

using namespace std;

__global__ void kern_2D(char **desc, char** merge_char) {

    int idx = threadIdx.x + blockDim.x*blockIdx.x;
    int idy = threadIdx.y + blockDim.y*blockIdx.y;

    if (idx < 10000)
    {
        char* s1 = desc[idx];
        merge_char[idx] = s1;
        //printf("From key = %s\n", merge_char[idx]);
    }

}


int main() {
    cudaError_t err = cudaSuccess;
    size_t max_line_len = 255;
    char line[255];
    size_t line_len;
    size_t max_lines_desc = 10000;
    //---------------------------------------------------------------------------------//

    char **d_desc;
    cudaMalloc(&d_desc, max_lines_desc * sizeof(char *));

    char **m_desc = NULL;
    m_desc = (char**)malloc(max_lines_desc * sizeof(char**));
    char **d_temp_desc = NULL;
    d_temp_desc = (char **)malloc(max_lines_desc * sizeof(char **));

    FILE *f_desc = fopen("desc.txt", "r");
    if (!f_desc)
    {
        fprintf(stderr, "Error opening file!\n");
    }
    int idesc = 0;

    do
    {
        if (!fgets(line, max_line_len, f_desc))
        {
            if (ferror(f_desc) && !feof(f_desc))
            {
                fprintf(stderr, "Error reading from file!\n");
                fclose(f_desc);
            }
            break;
        }

        line_len = strlen(line);
        if ((line_len > 0) && (line[line_len - 1] == '\n'))
        {
            line[line_len - 1] = '\0';
            --line_len;
        }
        m_desc[idesc] = line;
        cudaMalloc(&(d_temp_desc[idesc]), sizeof(line) * sizeof(char));
        cudaMemcpy(d_temp_desc[idesc], m_desc[idesc], sizeof(line) * sizeof(char), cudaMemcpyHostToDevice);
        cudaMemcpy(d_desc + idesc, &(d_temp_desc[idesc]), sizeof(char *), cudaMemcpyHostToDevice);

        ++idesc;
    } while (idesc < max_lines_desc);
    fclose(f_desc);

    //---------------------------------------------------------------------------------//


    char **merge_char;
    cudaMallocManaged(&merge_char, max_lines_desc * sizeof(char *));


    kern_2D << < 1, 1000 >> > (d_desc , merge_char);

    err = cudaDeviceSynchronize();
    if (err != cudaSuccess) {
        fprintf(stderr, "cudaDeviceSynchronize returned error code %s after launching addKernel!\n", cudaGetErrorString(err));
    }


    //---------------------------------------------------------------------------------//

    char** h_dev;

    cudaMalloc((void**)(&h_dev), max_lines_desc * sizeof(char*));
    err = cudaMemcpy(h_dev, merge_char, max_lines_desc * sizeof(char*), cudaMemcpyDeviceToHost);
    if (err == cudaSuccess) printf("2: Okay \n");


    for (int i = 0; i < max_lines_desc; i++)
    {
        printf("%s\n", h_dev[i]);
    }


    return 0;


}
//nvcc - arch = sm_30 - o kernel kernel.cu
// cuda - memcheck . / kernel

I'm sorry for my mistake. I already updated my code. It's completed.

For desc.txt, the file has 10000 lines like below. I checked status after copy from device to host, but I'm wrong. I cannot print char** h_dev.

motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125

Answer 1

I have to say that I don't really understand what your intent is here, because the only thing your kernel is doing is swapping pointers. If that is all you intend to do, you are certainly making things difficult for yourself by using double pointers everywhere. It would be far simpler just to manage the indices.

But to address your question, as far as I can tell your "copy back to host" is indeed incorrect. You are effectively doing a deep-copy of your data from host to device, and so you will need a deep-copy (2-stage copy) in the other direction as well.

To accomplish that, we don't use cudaMalloc on the copy to host. cudaMalloc allocates device memory. If you want to copy something to the host, your copy target is host memory. So we will need a set of cudaMemcpy operations to deep-copy the data back to the host, using host buffers as the targets.

The following code represents the simplest modifications I could make to what you have shown to accomplish this, and it seems to work for me for my simple test case:

$ cat desc.txt
1motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
2motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
3motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
4motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
5motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
6motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap1
$ cat t301.cu
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <string.h>
#include <malloc.h>
#include <fstream>
#include <algorithm>
#include <time.h>

using namespace std;

__global__ void kern_2D(char **desc, char** merge_char) {

    int idx = threadIdx.x + blockDim.x*blockIdx.x;
    int idy = threadIdx.y + blockDim.y*blockIdx.y;

    if (idx < 10000)
    {
        char* s1 = desc[idx];
        merge_char[idx] = s1;
        //printf("From key = %s\n", merge_char[idx]);
    }

}


int main() {
    cudaError_t err = cudaSuccess;
    size_t max_line_len = 255;
    char line[255];
    size_t line_len;
    size_t max_lines_desc = 10000;
    //---------------------------------------------------------------------------------//

    char **d_desc;
    cudaMalloc(&d_desc, max_lines_desc * sizeof(char *));

    char **m_desc = NULL;
    m_desc = (char**)malloc(max_lines_desc * sizeof(char**));
    char **d_temp_desc = NULL;
    d_temp_desc = (char **)malloc(max_lines_desc * sizeof(char **));

    FILE *f_desc = fopen("desc.txt", "r");
    if (!f_desc)
    {
        fprintf(stderr, "Error opening file!\n");
    }
    int idesc = 0;

    do
    {
        if (!fgets(line, max_line_len, f_desc))
        {
            if (ferror(f_desc) && !feof(f_desc))
            {
                fprintf(stderr, "Error reading from file!\n");
                fclose(f_desc);
            }
            break;
        }

        line_len = strlen(line);
        if ((line_len > 0) && (line[line_len - 1] == '\n'))
        {
            line[line_len - 1] = '\0';
            --line_len;
        }
        m_desc[idesc] = line;
        cudaMalloc(&(d_temp_desc[idesc]), sizeof(line) * sizeof(char));
        cudaMemcpy(d_temp_desc[idesc], m_desc[idesc], sizeof(line) * sizeof(char), cudaMemcpyHostToDevice);
        cudaMemcpy(d_desc + idesc, &(d_temp_desc[idesc]), sizeof(char *), cudaMemcpyHostToDevice);

        ++idesc;
    } while (idesc < max_lines_desc);
    fclose(f_desc);

    //---------------------------------------------------------------------------------//


    char **merge_char;
    cudaMallocManaged(&merge_char, max_lines_desc * sizeof(char *));


    kern_2D << < 1, 1000 >> > (d_desc , merge_char);

    err = cudaDeviceSynchronize();
    if (err != cudaSuccess) {
        fprintf(stderr, "cudaDeviceSynchronize returned error code %s after launching addKernel!\n", cudaGetErrorString(err));
    }


    //---------------------------------------------------------------------------------//

    char** h_dev;

    h_dev = (char **)malloc(max_lines_desc * sizeof(char*));
    err = cudaMemcpy(h_dev, merge_char, max_lines_desc * sizeof(char*), cudaMemcpyDeviceToHost);
    if (err == cudaSuccess) printf("2: Okay \n");


    for (int i = 0; i < 6; i++)
    {
        cudaMemcpy(line, h_dev[i], sizeof(line), cudaMemcpyDeviceToHost);
        printf("%s\n", line);
    }


    return 0;


}
$ nvcc -o t301 t301.cu
t301.cu(15): warning: variable "idy" was declared but never referenced

$ cuda-memcheck ./t301
========= CUDA-MEMCHECK
2: Okay
1motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
2motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
3motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
4motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
5motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap125
6motorcycle ckd new apsonic ckd 2017 ckd 2018 motorcycle apsoni new motorcycle apsonic no 125 motorcycle apsonic ap125 new motorcycle apsonic ap1
========= ERROR SUMMARY: 0 errors
$