cudaMemGetInfo在GTX 690的两个设备上返回相同的可用内存量

问题描述

我遇到了与Geforce GTX 690的问题，同时试图追踪内存使用。一个简单的测试程序：

I have run into problems with Geforce GTX 690 while trying to track down the memory usage. A simple test program:

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：

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，在分配之后，设备1上的可用内存量与设备0上的完全相同，这不是这种情况，因此问题必须在cudaMemGetInfo和/或cudaSetDevice。

The problem is that after the allocation the amount of free memory on device 1 is exactly the same as on device 0, which shouldn't be the case, hence the problem has to be in cudaMemGetInfo and/or cudaSetDevice. Anyone run on the same problem, or is there something else fundamentally wrong in the test that someone can point out?

在Windows 7，Visual studio 2010，Cuda上运行代码的任何人都会遇到同样问题， SDK 5.0，使用代码生成进行编译：compute_30，sm_30

Running the code on Windows 7, Visual studio 2010, Cuda SDK 5.0, compiling with code generation: compute_30,sm_30

EDIT 22.4.2013

EDIT 22.4.2013

问题，似乎cudaSetDevice工作正常，可以从cudaGetDevice调用的结果验证。我在内存分配测试后添加了设备0的重置，似乎cudaMemGetInfo返回的可用内存的大小对于两个设备都是相同的。我已经在我自己的代码中检查了所有的cuda_error_t的返回值，所有的函数调用返回cudaSuccess。

I continued experimenting with this issue and it seems that cudaSetDevice works fine as can be verified from the result of cudaGetDevice calls. I added a reset of device 0 after the memory allocation test and it seems that the size of free memory returned by cudaMemGetInfo is again same for both devices. I have checked all return values of cuda_error_t in my own code and all function calls return cudaSuccess. Have anyone run into similar problems with GTX 690 with the setup descibed above?

最新的测试代码：

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

推荐答案

这是通过更改WDDM驱动程序设置解决的，如下所示：

This was solved by changed WDDM driver settings as follows:

从作为社区Wiki条目的注释添加，以从CUDA标记的未应答队列中获取问题]

[This answer added from comments as a community wiki entry to get the question off the unanswered queue for the CUDA tag]

这篇关于cudaMemGetInfo在GTX 690的两个设备上返回相同的可用内存量的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！