Adding custom op instructions lacking the GPU building instructions

[panfengli]

Tensorflow Add a new op tutorial never mention the building instructions for GPU ops.

I find one here, but it met the similar issue with Building custom op instructions out of date .
tensorflow.python.framework.errors_impl.NotFoundError: ./libcuda_op.so: undefined symbol: _ZTIN10tensorflow8OpKernelE

Tensorflow really needs to give some suggestions on how to build GPU ops.

[panfengli]

The one from the above link:

cmake_minimum_required(VERSION 3.5)

set(CMAKE_C_COMPILER /opt/cuda/bin/gcc)
set(CMAKE_CXX_COMPILER /opt/cuda/bin/g++)

execute_process(COMMAND python3 -c "import tensorflow; print(tensorflow.sysconfig.get_include())" OUTPUT_VARIABLE Tensorflow_INCLUDE_DIRS)

find_package(CUDA)

set(CMAKE_CXX_FLAGS "-std=c++11 -O2 ${CMAKE_CXX_FLAGS}")

if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 5.0 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 5.0)
  set(CMAKE_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=0 ${CMAKE_CXX_FLAGS}")
endif()

SET(CUDA_PROPAGATE_HOST_FLAGS ON)

include_directories(${Tensorflow_INCLUDE_DIRS})

cuda_add_library(
    add_one SHARED
    cuda_op_kernel.cu
add_one.cc)

Currently Tensorflow only provides instructions for CPU ops as:

TF_CFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_compile_flags()))') )
TF_LFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_link_flags()))') )
g++ -std=c++11 -shared zero_out.cc -o zero_out.so -fPIC ${TF_CFLAGS[@]} ${TF_LFLAGS[@]} -O2

[ppwwyyxx]

The tutorial does mention the build instructions for GPU ops: https://www.tensorflow.org/extend/adding_an_op#gpu_support

[panfengli]

Thanks! @ppwwyyxx