caffe opencl branch : test case (GPUMathFunctionsTest, TestCopy)

[Ravishekhar Banger]

I posted this question on the Caffe Users forum and reposting it here for the right group. I am sorry for the churn. 

This issue is regarding the testcase GPUMathFunctionsTest/0.TestCopy defined in src/caffe/test/test_math_functions.cpp 

In this test case we are creating a new opencl context when we call viennacl::ocl::get_context () at line 272 

Below are my observations. Function viennacl::ocl::get_context() is being called twice 

1 . The function caffe::device::Init() which is in the src/caffe/device.cpp file 

2 . In the test case file src/caffe/test/test_math_functions.cpp:272

There will be two different instances of the static variable viennacl::ocl::backend::initialized_ and it may result in a new context being created. I am seeing this issue with the compiler aarch64-linux-android-g++ (gcc version 4.9). Though not able to reproduce the same issue with the gcc x86 compiler (version 4.8.4)

The issues: 

1. Is the x86 compiler having an issue. 

2. ViennaCL library declaring a static variable in a header file. (backend.hpp). A design issue. 

3. Or is the test case incorrect. In this testcase greentea_copy() calls clEnqueueCopyBuffer() and the context for the cl_mem objects are different than the actual context passed to this OpenCL API.  Ideally the context should be obtained using the cl_mem object. 

Is my understanding correct? Please correct me if I am wrong. 

Regards,

Banger.