2

我在尝试追踪内存使用情况时遇到了 Geforce GTX 690 的问题。一个简单的测试程序：

BOOST_AUTO_TEST_CASE(cudaMemoryTest) {

size_t mem_tot_0 = 0;
size_t mem_free_0 = 0;
size_t mem_tot_1 = 0;
size_t mem_free_1 = 0;

unsigned int mem_size = 100*1000000;

float* h_P = new float[mem_size];
for(size_t i = 0; i < mem_size; i++) {
    h_P[i] = 0.f;
}

cudaSetDevice(0);
cudaDeviceReset();
cudaMemGetInfo  (&mem_free_0, & mem_tot_0);
std::cout<<"Free memory before copy dev 0: "<<mem_free_0<<std::endl;

cudaSetDevice(1);
cudaDeviceReset();
cudaMemGetInfo  (&mem_free_1, &mem_tot_1);
std::cout<<"Free memory before copy dev 1: "<<mem_free_1<<std::endl;

cudaSetDevice(0);
float* P;
cudaMalloc((void**)&P, mem_size*sizeof(float));
cudaMemcpy((void*)P, h_P,  mem_size*sizeof(float), cudaMemcpyHostToDevice);

cudaSetDevice(0);
cudaMemGetInfo(&mem_free_0, & mem_tot_0);
std::cout<<"Free memory after copy  dev 0: "<<mem_free_0<<std::endl;

cudaSetDevice(1);
cudaMemGetInfo(&mem_free_1, &mem_tot_1);
std::cout<<"Free memory after copy dev 1: "<<mem_free_1<<std::endl;

BOOST_CHECK(mem_free_0 != mem_free_1);

cudaError_t err;
err = cudaGetLastError();
if(err!=cudaSuccess)
    std::cout<<"an error occurred"<<std::endl;

cudaSetDevice(0);
destroyMem(P);
delete [] h_P;
}

测试打印出来：

1>  Free memory before copy dev 0: 1733173248
1>  Free memory before copy dev 1: 1688424448
1>  Free memory after copy  dev 0: 1289940992
1>  Free memory after copy dev 1: 1289940992     
CudaUtilsTest.cpp(47): error in "cudaMemoryTest": check mem_free_0 != mem_free_1 failed

问题是分配后设备 1 上的空闲内存量与设备 0 上的完全一样，这不应该是这种情况，因此问题必须在 cudaMemGetInfo 和/或 cudaSetDevice 中。任何人都在同样的问题上运行，或者有人可以指出测试中是否存在其他根本性的错误？

在 Windows 7、Visual Studio 2010、Cuda SDK 5.0 上运行代码，使用代码生成进行编译：compute_30,sm_30

编辑 22.4.2013

我继续试验这个问题，似乎 cudaSetDevice 工作正常，可以从 cudaGetDevice 调用的结果中得到验证。我在内存分配测试之后添加了设备 0 的重置，并且似乎 cudaMemGetInfo 返回的可用内存大小对于两个设备来说再次相同。我已经在自己的代码中检查了 cuda_error_t 的所有返回值，并且所有函数调用都返回 cudaSuccess。有没有人在上面描述的设置中遇到过与 GTX 690 类似的问题？

最新测试代码：

BOOST_AUTO_TEST_CASE(cudaMemoryTest) {
size_t mem_tot_0 = 0;
size_t mem_free_0 = 0;
size_t mem_tot_1 = 0;
size_t mem_free_1 = 0;

int device_num = 0;

unsigned int mem_size = 100*1000000;

float* h_P = new float[mem_size];
for(size_t i = 0; i < mem_size; i++) {
    h_P[i] = 0.f;
}

cudaSetDevice(0);
cudaGetDevice(&device_num);
cudaDeviceReset();
cudaMemGetInfo  (&mem_free_0, & mem_tot_0);
std::cout<<"Free memory before copy dev 0: "<<mem_free_0<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

cudaSetDevice(1);
cudaGetDevice(&device_num);
cudaDeviceReset();
cudaMemGetInfo  (&mem_free_1, & mem_tot_1);
std::cout<<"Free memory before copy dev 1: "<<mem_free_1<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

cudaSetDevice(0);
cudaGetDevice(&device_num);
float* P;
cudaMalloc((void**)&P, mem_size*sizeof(float));
cudaMemcpy((void*)P, h_P,  mem_size*sizeof(float), cudaMemcpyHostToDevice);
cudaMemGetInfo(&mem_free_0, & mem_tot_0);
std::cout<<"Free memory after copy  dev 0: "<<mem_free_0<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

cudaSetDevice(1);
cudaGetDevice(&device_num);
cudaMemGetInfo(&mem_free_1, &mem_tot_1);
std::cout<<"Free memory after copy dev 1: "<<mem_free_1<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

BOOST_CHECK(mem_free_0 != mem_free_1);

cudaError_t err;
err = cudaGetLastError();
if(err!=cudaSuccess)
    std::cout<<"an error occurred"<<std::endl;

// Reset only device 0 and check both
cudaSetDevice(0);
cudaGetDevice(&device_num);
cudaDeviceReset();
cudaMemGetInfo  (&mem_free_0, & mem_tot_0);
std::cout<<"Free memory after second reset of device 0, dev 0: "<<mem_free_0<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

cudaSetDevice(1);
cudaGetDevice(&device_num);
cudaMemGetInfo  (&mem_free_1, & mem_tot_1);
std::cout<<"Free memory after second device reset of device 0, dev 1: "<<mem_free_1<<" Device: "<<device_num<<std::endl;
cudaDeviceSynchronize();

delete [] h_P;
    }

测试输出：

1>  Free memory before copy dev 0: 1794379776 Device: 0
1>  Free memory before copy dev 1: 1751728128 Device: 1
1>  Free memory after copy  dev 0: 1351696384 Device: 0
1>  Free memory after copy dev 1: 1351696384 Device: 1
1>  CudaUtilsTest.cpp(353): error in "cudaMemoryTest": check mem_free_0 != mem_free_1 failed
1>  Free memory after second reset of device 0, dev 0: 1751728128 Device: 0
1>  Free memory after second device reset of device 0, dev 1: 1751728128 Device: 1
4

1 回答 1

4

这已通过更改 WDDM 驱动程序设置解决，如下所示：

从 NVIDIA 控制面板的“3D 设置”->“配置多 GPU、环绕声、PhysX”中切换“禁用多 GPU 模式”。

[此答案从评论中添加为社区 wiki 条目，以将问题从 CUDA 标签的未回答队列中删除]

于 2016-01-28T17:38:41.840 回答