CUDNN_STATUS_NOT_INITIALIZED when using Caffe with cuDNN

2,752 views
Skip to first unread message

Tiferet Gazit

unread,
Nov 19, 2015, 10:57:58 AM11/19/15
to Caffe Users

I am trying to use cuDNN to speed up my Caffe networks.


When I "make runtest" with cuDNN (and also when I run my own Caffe models) I get the following error for many of the cuDNN layers (for example,cudnn_softmax_layer.cpp):

 

Check failed: status == CUDNN_STATUS_SUCCESS (1 vs. 0)  CUDNN_STATUS_NOT_INITIALIZED

*** Check failure stack trace: ***

    @     0x2b0f3ac86daa  (unknown)

    @     0x2b0f3ac86ce4  (unknown)

    @     0x2b0f3ac866e6  (unknown)

    @     0x2b0f3ac89687  (unknown)

    @     0x2b0f3c62a01e  caffe::CuDNNSoftmaxLayer<>::LayerSetUp()

    @     0x2b0f3c5f9b36  caffe::SoftmaxWithLossLayer<>::LayerSetUp()

    @           0x498be0  caffe::Layer<>::SetUp()

    @           0x498f6f  caffe::GradientChecker<>::CheckGradientExhaustive()

    @           0x597f7f  caffe::SoftmaxWithLossLayerTest_TestGradientUnnormalized_Test<>::TestBody()

    @           0x7cff8d  testing::internal::HandleExceptionsInMethodIfSupported<>()

    @           0x7c7f21  testing::Test::Run()

    @           0x7c8006  testing::TestInfo::Run()

    @           0x7c8147  testing::TestCase::Run()

    @           0x7c849e  testing::internal::UnitTestImpl::RunAllTests()

    @           0x7cfb0d  testing::internal::HandleExceptionsInMethodIfSupported<>()

    @           0x7c757e  testing::UnitTest::Run()

    @           0x44f832  main

    @     0x2b0f3d168ec5  (unknown)

    @           0x455664  (unknown)

    @              (nil)  (unknown)

make: *** [runtest] Aborted (core dumped)

 

I am using a Quadro K2200 card with NVIDIA driver version 340.93 on an Ubuntu 14.04.2 machine.

 

My Caffe Makefile defines the following CUDA architecture, although I have tried changing it in various ways to no avail (with 50 I am unable to build Caffe at all):

CUDA_ARCH := -gencode arch=compute_35,code=sm_35 \

                                -gencode arch=compute_35,code=compute_35

 

I have tried everything I could find online, but haven't been able to solve the problem. I would greatly appreciate some help. Thank you!

Tiferet Gazit

unread,
Nov 22, 2015, 9:37:18 AM11/22/15
to Caffe Users
Updating the solution I finally figured out: 
Apparently, in addition to the correct CUDA driver, I also had an older driver installed on my computer, and the older one was the one being used in practice. Once I removed the older driver I was able to build Caffe with the CUDA architecture defined in the original example Makefile, and was then able to run all the tests, as well as my own networks.


Reply all
Reply to author
Forward
0 new messages