在双显卡系统中,cuda运行时NVIDAI显卡必须是当前使用的显卡,否则无法获取GPU设备,cudaGetDeviceCount函数会报错,错误码35。 使用nvidia-prime切换到N卡时,如果只是按照提示logout,再重新login是不行的,必须重启系统,否则会报错,错误码30。 caffe的问题也是同样的道理,因为caffe也要调用cuda的cudaGetDeviceCount函数获取GPU设备。
环境:ubuntu16+nvidia-378 driver+cuda8.0+cudnn5.1+GTX965M显卡 成功安装cuda 8.0,也正常编译了Caffe以后,执行mnist训练程序来测试一下,然鹅报错了:
$ ./build/tools/caffe train –solver=examples/mnist/lenet_solver.prototxt I0312 22:15:25.125078 2171 caffe.cpp:217] Using GPUs 0 I0312 22:15:25.126852 2171 caffe.cpp:222] GPU 0: 0 cg� F0312 22:15:25.127008 2171 common.cpp:151] Check failed: error == cudaSuccess (35 vs. 0) CUDA driver version is insufficient for CUDA runtime version * Check failure stack trace: * @ 0x7f1bb831c5cd google::LogMessage::Fail() @ 0x7f1bb831e433 google::LogMessage::SendToLog() @ 0x7f1bb831c15b google::LogMessage::Flush() @ 0x7f1bb831ee1e google::LogMessageFatal::~LogMessageFatal() @ 0x7f1bb8882192 caffe::Caffe::SetDevice() @ 0x40c950 train() @ 0x4088e8 main @ 0x7f1bb6a51830 __libc_start_main @ 0x4091b9 _start @ (nil) (unknown) 已放弃 (核心已转储)
显然是cuda的问题,于是执行cuda samples程序中的deviceQuery,果然也是报错,,错误码35:
$ ./NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery ./NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery Starting… CUDA Device Query (Runtime API) version (CUDART static linking) cudaGetDeviceCount returned 35 -> CUDA driver version is insufficient for CUDA runtime version Result = FAIL
笔记本电脑是双显卡(i7 cpu有集成显卡),猜测应该是NVIDIA显卡没启用,执行nvidia-setting,在PRIME profile中果然显示当前使用的是Intel 集成显卡,于是切换到nvidia显卡。
PRIME切换到nvidia显卡时提示要logout才能生效,于是登出再重新登录,再执行上面的mnist训练,还是报错,执行deviceQuery也报错,不过这次错误不一样了,错误代码30。
$ ./build/tools/caffe train –solver=examples/mnist/lenet_solver.prototxt I0312 22:17:46.619762 3469 caffe.cpp:217] Using GPUs 0 I0312 22:17:46.639750 3469 caffe.cpp:222] GPU 0: 1(� F0312 22:17:46.639799 3469 common.cpp:151] Check failed: error == cudaSuccess (30 vs. 0) unknown error * Check failure stack trace: * @ 0x7fe1702315cd google::LogMessage::Fail() @ 0x7fe170233433 google::LogMessage::SendToLog() @ 0x7fe17023115b google::LogMessage::Flush() @ 0x7fe170233e1e google::LogMessageFatal::~LogMessageFatal() @ 0x7fe170797192 caffe::Caffe::SetDevice() @ 0x40c950 train() @ 0x4088e8 main @ 0x7fe16e966830 __libc_start_main @ 0x4091b9 _start @ (nil) (unknown) 已放弃 (核心已转储) $ ./NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery ./NVIDIA_CUDA-8.0_Samples/bin/x86_64/linux/release/deviceQuery Starting… CUDA Device Query (Runtime API) version (CUDART static linking) cudaGetDeviceCount returned 30 -> unknown error Result = FAIL
要想到切换显卡时没有重启系统,是不是这个原因生成的呢?于是sudo reboot
重启电脑,再次进入,执行deviceQuery就正常了
再执行mnist训练也正常了。 尼玛这PRIME的提示妥妥的是坑爹嘛,你直接提示切换显卡要reboot不就成了嘛 ,logout显然不管用嘛。