Question

我在代码中使用了很多东西，因为它是一个很棒的包装器并且提供了非常有用的实用程序，我更加确信，因为已经添加了对异步行为的支持。

我的代码在使用cuda推力方面运行良好，直到我最近尝试在我的应用程序中添加多GPU支持。我经历过烦人的事情

CUDA Runtime API错误77：遇到非法内存访问

我之前从未显示过任何边界问题的部分代码。

我在我的代码中添加了冗长的内容，看来我的thrust :: device_vector指针地址在执行过程中发生了变化，没有明显的原因，在手写内核中产生错误77.

我可能误解了UVA概念及其最终的副作用＆＃34;但是，我仍然对理解导致指针更新的过程感兴趣。

我无法正确再现我的问题，我不使用临时主变量来存储cuda内存指针，但在内核包装器调用中需要时只需要push :: raw_pointer_cast。

但是我编写了一个小程序，显示我可能遇到的哪种错误，请注意这不是很强大，你需要在你的系统上至少有2个gpu才能运行它：

/********************************************************************************************
** Compile using nvcc ./test.cu -gencode arch=compute_35,code=sm_35 -std=c++11 -o test.exe **
********************************************************************************************/

//Standard Library
#include <iostream>
#include <vector>

//Cuda
#include "cuda_runtime.h"

//Thrust
#include <thrust/device_vector.h>
#include <thrust/functional.h>
#include <thrust/transform.h>

inline void __checkCudaErrors( cudaError err, const char *file, const int line )
{
    if( err != cudaSuccess )
    {
        printf("%s(%i) : CUDA Runtime API error %i : %s \n",file ,line, (int)err, cudaGetErrorString(err) );
    }
};

#define checkCudaErrors(err)    __checkCudaErrors (err, __FILE__, __LINE__)

__global__ void write_memory( float* buf, float value )
{
    printf("GPU TALK: Raw pointer is %p \n",buf);
    buf[0] = value;
}

int main( int argc, char* argv[] )
{
    //declare a vector of vector
    std::vector<thrust::device_vector<float> > v;
    float test;
    float* tmp;

    //Initialize first vector on GPU 0
    cudaSetDevice( 0 );
    v.emplace_back( 65536, 1.0f );
    tmp = thrust::raw_pointer_cast( v.at(0).data() );
    std::cout << " Host TALK: Raw pointer of vector 0 at step 0 " << (void*)tmp << std::endl;

    //Try to use it raw pointer
    write_memory<<<1,1,0,0>>>( tmp, 2.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After first kernel launch, value is " << test << std::endl;

    //Initialize second vector on GPU 1, but we do not use it
    cudaSetDevice( 1 );
    v.emplace_back( 65536, 1.0f );
    std::cout << " Host TALK: Raw pointer of vector 0 at step 1 is now " << (void*)thrust::raw_pointer_cast( v.at(0).data() ) << " != " << (void*)tmp << std::endl;
    std::cout << " Host TALK: Raw pointer of vector 1 at step 1 is " << (void*)thrust::raw_pointer_cast( v.at(1).data() ) << std::endl; 

    //Try to use the first vector : No segmentation fault ?
    test = v.at(0)[0];
    std::cout << " Host TALK: Before second kernel launch, value is " << test << std::endl;
    write_memory<<<1,1,0,0>>>( thrust::raw_pointer_cast( v.at(0).data() ), 3.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After second kernel launch, value is " << test << std::endl;

    //Raw pointer stored elsewhere: generates a segmentation fault
    write_memory<<<1,1,0,0>>>( tmp, 4.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After third kernel launch, value is " << test << std::endl;

    return 0;
}

以下是我在我的机器上生成的输出示例：

Host TALK：步骤0中向量0的原始指针0xb043c0000
GPU TALK：原始指针为0xb043c0000
Host TALK：首次启动内核后，值为2
主机TALK：步骤1中向量0的原始指针现在是0xb08000000！= 0xb043c0000
主机TALK：步骤1中矢量1的原始指针是0xb07fc0000
Host TALK：在第二次内核启动之前，值为2
GPU TALK：原始指针为0xb08000000
Host TALK：第二次内核启动后，值为3
GPU TALK：原始指针为0xb043c0000
./test.cu(68）：CUDA运行时API错误77：遇到非法内存访问终止在抛出＆＃39; thrust :: system :: system_error＆＃39; what（）：遇到非法内存访问

提前感谢您的帮助，我也可以在推荐的github上提出这个问题。

编辑：感谢m.s和Hiura，这里有一个按预期工作的代码：

/********************************************************************************************
** Compile using nvcc ./test.cu -gencode arch=compute_35,code=sm_35 -std=c++11 -o test.exe **
********************************************************************************************/

//Standard Library
#include <iostream>
#include <vector>

//Cuda
#include "cuda_runtime.h"

//Thrust
#include <thrust/device_vector.h>
#include <thrust/functional.h>
#include <thrust/transform.h>

inline void __checkCudaErrors( cudaError err, const char *file, const int line )
{
    if( err != cudaSuccess )
    {
        printf("%s(%i) : CUDA Runtime API error %i : %s \n",file ,line, (int)err, cudaGetErrorString(err) );
    }
};

#define checkCudaErrors(err)    __checkCudaErrors (err, __FILE__, __LINE__)

__global__ void write_memory( float* buf, float value )
{
    printf("GPU TALK: Raw pointer is %p \n",buf);
    buf[0] = value;
}

int main( int argc, char* argv[] )
{
    //declare a vector of vector
    std::vector<thrust::device_vector<float> > v;
    v.reserve(2);
    float test;
    float* tmp;

    //Initialize first vector on GPU 0
    cudaSetDevice( 0 );
    v.emplace_back( 65536, 1.0f );
    tmp = thrust::raw_pointer_cast( v.at(0).data() );
    std::cout << " Host TALK: Raw pointer of vector 0 at step 0 " << (void*)tmp << std::endl;

    //Try to use it raw pointer
    write_memory<<<1,1,0,0>>>( tmp, 2.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After first kernel launch, value is " << test << std::endl;

    //Initialize second vector on GPU 1, but we do not use it
    cudaSetDevice( 1 );
    v.emplace_back( 65536, 1.0f );
    std::cout << " Host TALK: Raw pointer of vector 0 at step 1 is now " << (void*)thrust::raw_pointer_cast( v.at(0).data() ) << " != " << (void*)tmp << std::endl;
    std::cout << " Host TALK: Raw pointer of vector 1 at step 1 is " << (void*)thrust::raw_pointer_cast( v.at(1).data() ) << std::endl; 

    //Try to use the first vector : No segmentation fault ?
    cudaSetDevice( 0 );
    test = v.at(0)[0];
    std::cout << " Host TALK: Before second kernel launch, value is " << test << std::endl;
    write_memory<<<1,1,0,0>>>( thrust::raw_pointer_cast( v.at(0).data() ), 3.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After second kernel launch, value is " << test << std::endl;

    //Raw pointer stored elsewhere: generates a segmentation fault
    write_memory<<<1,1,0,0>>>( tmp, 4.0f );
    checkCudaErrors( cudaStreamSynchronize( NULL ) );
    test = v.at(0)[0];
    std::cout << " Host TALK: After third kernel launch, value is " << test << std::endl;

    return 0;
}

这是我的代码中的最后一个地方，为了简单起见，我没有使用指针对象的向量而不是对象的向量，但我发现我应该避免这些恼人的移动/复制问题... < / p>

现在输出是：

Host TALK：步骤0中向量0的原始指针0xb043c0000
GPU TALK：原始指针为0xb043c0000
Host TALK：首次启动内核后，值为2
主机TALK：步骤1中向量0的原始指针现在为0xb043c0000！= xb043c0000
主机TALK：步骤1中矢量1的原始指针是0xb07fc0000
Host TALK：在第二次内核启动之前，值为2
GPU TALK：原始指针为0xb043c0000
Host TALK：第二次内核启动后，值为3
GPU TALK：原始指针为0xb043c0000
Host TALK：第三次内核启动后，值为4

Answer 1

所以我快速安装了CUDA以测试我的假设：添加reserve语句会保留地址。

//declare a vector of vector
std::vector<thrust::device_vector<float> > v;
v.reserve(2); // <<-- HERE
float test;
float* tmp;

输出，首先没有补丁。

 $ nvcc thrust.cu  -std=c++11 -o test
 $ ./test 
  Host TALK: Raw pointer of vector 0 at step 0 0x700ca0000
 GPU TALK: Raw pointer is 0x700ca0000 
  Host TALK: After first kernel launch, value is 2
  Host TALK: Raw pointer of vector 0 at step 1 is now 0x700d20000 != 0x700ca0000
  Host TALK: Raw pointer of vector 1 at step 1 is 0x700ce0000
  Host TALK: Before second kernel launch, value is 2
 GPU TALK: Raw pointer is 0x700d20000 
  Host TALK: After second kernel launch, value is 3
 GPU TALK: Raw pointer is 0x700ca0000 
  Host TALK: After third kernel launch, value is 3

补丁：

 $ nvcc thrust.cu  -std=c++11 -o test
 $ ./test 
  Host TALK: Raw pointer of vector 0 at step 0 0x700ca0000
 GPU TALK: Raw pointer is 0x700ca0000 
  Host TALK: After first kernel launch, value is 2
  Host TALK: Raw pointer of vector 0 at step 1 is now 0x700ca0000 != 0x700ca0000
  Host TALK: Raw pointer of vector 1 at step 1 is 0x700ce0000
  Host TALK: Before second kernel launch, value is 2
 GPU TALK: Raw pointer is 0x700ca0000 
  Host TALK: After second kernel launch, value is 3
 GPU TALK: Raw pointer is 0x700ca0000 
  Host TALK: After third kernel launch, value is 4

thrust :: raw_pointer_cast和多GPU，奇怪的行为

1 个答案: