TensorFlow 1.6 with Cuda8.0 cuDNN 7.5 Python2.7

[InzamamAnwar]

Hi, I would like to compile TensorFlow 1.6 with CUDA 8.0 and cuDNN 7.5 with Python 2.7 for Windows. Can you please provide build instructions or provide link is there is any.

[fo40225]

• use anaconda2, you're halfway there.

• you need those change fo40225/tensorflow@b0aac83

• the build command is nothing different from python 2 to 3

If you get the error msg, post it and I will tell you how to overcome them.

[InzamamAnwar]

I have started building the TF with the following options:

TensorFlow 1.6, CUDA 8.0, cuDNN 7.2, Anaconda2 5.2.0, Python2.7.15, Microsoft Visual Studio 2015 (Community) and made changes as you suggested.

Following error is occurred when I start compiling TF

Can you please suggest any remedy!

[InzamamAnwar]

I commented the following piece of code in CMakeLists.txt

Test compatibility of compiler on CUDA

try_compile(CUDA_TEST_COMPILE_C
${CMAKE_CURRENT_BINARY_DIR}/tests/cuda
${CMAKE_CURRENT_SOURCE_DIR}/tests/cuda/compatibility_test.c
CMAKE_FLAGS -DINCLUDE_DIRECTORIES=${CUDA_INCLUDE_DIRS})
try_compile(CUDA_TEST_COMPILE_CXX
${CMAKE_CURRENT_BINARY_DIR}/tests/cuda
${CMAKE_CURRENT_SOURCE_DIR}/tests/cuda/compatibility_test.cc
CMAKE_FLAGS -DINCLUDE_DIRECTORIES=${CUDA_INCLUDE_DIRS})
if(NOT (CUDA_TEST_COMPILE_C AND CUDA_TEST_COMPILE_CXX))
message(FATAL_ERROR "Selected compiler (or version) is not supported for CUDA")
endif()

After this I have encountered the following errors

[fo40225]

use VS2015 x64 Native Tools prompt

no need to change CMakeLists.txt

[InzamamAnwar]

I have used the VS2015 x64 Native Tools prompt and uncommented the code but again it shows the same error of "Selected Compiler (or version) is not supported for CUDA"

[fo40225]

C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin\x86_amd64\cl.exe is VS2015 x86 x64 Cross Tools not VS2015 x64 Native Tools.

[InzamamAnwar]

I have used "VS2015 x64 Native Tools Command Prompt". Before that I used the command call "C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\vcvarsall.bat" amd64. Both gives the same result as shown in previous comment.

[fo40225]

I had used your environment and got the same error.

I think it should be a compatibility issue between version 1.6 and the old version of CUDA.

If you need CUDA 8.0, use tf 1.4.

If you need tf 1.6, use cuda 9.0.

or you can fix the build break yourself