c++ - CUDA:错误:创建 thrust::device_ptr 时出现 "transfer of control bypasses initialization of"

Question

我的 Cuda - C 应用程序中有以下代码行:

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

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

#include <time.h>

#include <device_functions.h>

int main()
{
    const int size = 32;
    unsigned int * dev_ips_range_end;
    unsigned int * ips_range_end = new unsigned int[size];
    for (int i = 0; i < size; i++)
        ips_range_end[i] = i;



cudaError_t cudaStatus;
    cudaStatus = cudaSetDevice(0);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaSetDevice failed!  Do you have a CUDA-capable GPU installed?");
        goto Error;
    }
    cudaStatus = cudaMalloc((void**)&dev_ips_range_end, size * sizeof(int));
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "Problem !");
        goto Error;
    }

    // Copy input vectors from host memory to GPU buffers.
    cudaStatus = cudaMemcpy(dev_ips_range_end, ips_range_end, size * sizeof(unsigned char), cudaMemcpyHostToDevice);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "Problem !");
        goto Error;
    }

    thrust::device_ptr<unsigned int> dev_ips_range_end_ptr(dev_ips_range_end);
    thrust::inclusive_scan(dev_ips_range_end_ptr, dev_ips_range_end_ptr + size, dev_ips_range_end_ptr);
    return 0;

Error:
    cudaFree(dev_ips_range_end);
}

这是我使用的命令和输出:

[测试]$ nvcc -I/usr/local/cuda/include -L/usr/local/cuda/lib kernel.cu -o test.run kernel.cu(27): 错误:控制转移绕过以下初始化: 变量“dev_ips_range_end_ptr” (42): 这里

kernel.cu(32): 错误:控制转移绕过以下初始化: 变量“dev_ips_range_end_ptr” (42): 这里

kernel.cu(39): 错误:控制转移绕过以下初始化: 变量“dev_ips_range_end_ptr” (42): 这里

编译“/tmp/tmpxft_000022ad_00000000-9_kernel.cpp1.ii”时检测到 3 个错误。

相同的代码在 Windows 上的 visual studio 中运行没有任何问题。 如何解决这个问题？

Answer 1

有些人可能会告诉您，在 C/C++ 中使用 goto 并不是一个好主意。但是为了避免参数，并允许您保持相同的代码结构，您可以在程序的顶部声明您的推力设备指针(在任何 goto 语句之前)，然后在您执行时设置指针值准备好使用它，像这样:

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

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

#include <time.h>

#include <device_functions.h>

int main()
{
    const int size = 32;
    unsigned int * dev_ips_range_end;
    unsigned int * ips_range_end = new unsigned int[size];
    for (int i = 0; i < size; i++)
        ips_range_end[i] = i;



    thrust::device_ptr<unsigned int> dev_ips_range_end_ptr;
cudaError_t cudaStatus;
    cudaStatus = cudaSetDevice(0);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "cudaSetDevice failed!  Do you have a CUDA-capable GPU installed?");
        goto Error;
    }
    cudaStatus = cudaMalloc((void**)&dev_ips_range_end, size * sizeof(int));
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "Problem !");
        goto Error;
    }

    // Copy input vectors from host memory to GPU buffers.
    cudaStatus = cudaMemcpy(dev_ips_range_end, ips_range_end, size * sizeof(unsigned char), cudaMemcpyHostToDevice);
    if (cudaStatus != cudaSuccess) {
        fprintf(stderr, "Problem !");
        goto Error;
    }

    dev_ips_range_end_ptr = thrust::device_pointer_cast(dev_ips_range_end);
    thrust::inclusive_scan(dev_ips_range_end_ptr, dev_ips_range_end_ptr + size, dev_ips_range_end_ptr);
    return 0;

Error:
    cudaFree(dev_ips_range_end);
}