Check failed: status == CUBLAS_STATUS_SUCCESS (13 vs. 0) CUBLAS_STATUS_EXECUTION_FAILED

[ RUN      ] PowerLayerTest/3.TestPowerOneGradient
F0319 15:50:19.414253 22426 math_functions.cu:92] Check failed: status == CUBLAS_STATUS_SUCCESS (13 vs. 0)  CUBLAS_STATUS_EXECUTION_FAILED
*** Check failure stack trace: ***
    @     0x7f574fe7e78d  google::LogMessage::Fail()
    @     0x7f574fe80d43  google::LogMessage::SendToLog()
    @     0x7f574fe7e31b  google::LogMessage::Flush()
    @     0x7f574fe7fc8e  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f574e348e4e  caffe::caffe_gpu_scal<>()
    @     0x7f574e336906  caffe::PowerLayer<>::Forward_gpu()
    @           0x455772  caffe::Layer<>::Forward()
    @           0x496300  caffe::GradientChecker<>::CheckGradientSingle()
    @           0x50ea94  caffe::GradientChecker<>::CheckGradientEltwise()
    @           0x6b69b5  caffe::PowerLayerTest<>::TestBackward()
    @           0x7ac6b3  testing::internal::HandleExceptionsInMethodIfSupported<>()
    @           0x7a5cca  testing::Test::Run()
    @           0x7a5e18  testing::TestInfo::Run()
    @           0x7a5ef5  testing::TestCase::Run()
    @           0x7a71cf  testing::internal::UnitTestImpl::RunAllTests()
    @           0x7a74f3  testing::UnitTest::Run()
    @           0x44bfa9  main
    @     0x7f574d5a8830  __libc_start_main
    @           0x451c39  _start
Makefile:478: recipe for target 'runtest' failed
make: *** [runtest] 已放棄 (core dumped)

 首先明確，這是make runtest的錯誤，所以一定不能是代碼問題。一定是我的配置問題。雖然是segnet作者改得caffe，但是應該沒有問題。不過我還是打算用官方的caffe跑一下試試。

我之前那篇教程，安裝教程http://www.cnblogs.com/SweetBeens/p/8525131.html，提到如果ubuntu16.04裝9,可能會有問題，那么問題來了。官網給出的doc提到了這個問題：

The GPU program failed to execute. This is often caused by a launch failure of the kernel on the GPU, which can be caused by multiple reasons.

To correct: check that the hardware, an appropriate version of the driver, and the cuBLAS library are correctly installed.

github上有個比較火的討論貼：

https://github.com/BVLC/caffe/issues/2417

大家的解決方法是裝cuda8.0,還有說不用cudnn。

但由於我不到黃河心不死，不想用cuda8

那么有如下思路：

1,跑官網caffe，看看是不是因為segnet_caffe版本太低或者什么的

2.改下gcc版本

3.更換driver和cuda版本

等我試試。而且為什么別人的有一些可用16.04+9.1，奇怪了。

擬解決過程

1.我跑了官網的caffe，出現如下錯誤：

[  FAILED  ] EmbedLayerTest/3.TestForward, where TypeParam = caffe::GPUDevice<double> (1 ms)
[ RUN      ] EmbedLayerTest/3.TestGradient
[       OK ] EmbedLayerTest/3.TestGradient (101 ms)
[ RUN      ] EmbedLayerTest/3.TestSetUp
[       OK ] EmbedLayerTest/3.TestSetUp (0 ms)
[ RUN      ] EmbedLayerTest/3.TestForwardWithBias
F0319 17:26:25.959848 30839 math_functions.cu:42] Check failed: status == CUBLAS_STATUS_SUCCESS (13 vs. 0)  CUBLAS_STATUS_EXECUTION_FAILED
*** Check failure stack trace: ***
    @     0x7f787028d78d  google::LogMessage::Fail()
    @     0x7f787028fd43  google::LogMessage::SendToLog()
    @     0x7f787028d31b  google::LogMessage::Flush()
    @     0x7f787028ec8e  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f786e105672  caffe::caffe_gpu_gemm<>()
    @     0x7f786e13d75a  caffe::EmbedLayer<>::Forward_gpu()
    @           0x476522  caffe::Layer<>::Forward()
    @           0x4f5ef6  caffe::EmbedLayerTest_TestForwardWithBias_Test<>::TestBody()
    @           0x90b393  testing::internal::HandleExceptionsInMethodIfSupported<>()
    @           0x9049aa  testing::Test::Run()
    @           0x904af8  testing::TestInfo::Run()
    @           0x904bd5  testing::TestCase::Run()
    @           0x905eaf  testing::internal::UnitTestImpl::RunAllTests()
    @           0x9061d3  testing::UnitTest::Run()
    @           0x469fed  main
    @     0x7f786d2ec830  __libc_start_main
    @           0x471a69  _start
Makefile:532: recipe for target 'runtest' failed
make: *** [runtest] 已放棄 (core dumped)

 問題解決

cuda版本問題。在我的電腦配置環境下，必須是ubuntu 16.04+cuda8。不可以用9.1.

那么借鑒這個思路，請各位看下自己的版本有沒有不按要求整的。基本限制就是cuda，python。

我簡化的安裝步驟，更方便查漏補缺：

http://www.cnblogs.com/SweetBeens/p/8652083.html

詳細的安裝版本制約:

http://www.cnblogs.com/SweetBeens/p/8525131.html

以及利用anaconda2隨意配置多個版本：

http://www.cnblogs.com/SweetBeens/p/8650460.html

卸載cuda9.1安裝cuda8

http://www.cnblogs.com/SweetBeens/p/8616797.html