Experiment on vgg flower, 2 trainers, 1ps.
    more trainer could have more speedup.

    After:
    Pass = 0, Iters = 327, Speed = (7.52) img/s
    Before:
    Pass = 0, Iters = 385, Speed = (6.77) img/s

Add an interface to set the number of threads for math function, and set the default value to 1 for inference.

* test_patch_mkldnn

* use patch file

* follow comments

* fix patch

disable building tests and examples when install mkldnn
skip -Werror=unused-result

* "re-commit "

* "picked up"

* "fix ci"

* "fix pdb hang up issue in cuda 9"

* "fix cuda9 error"

* "change commit id"

* "remote git tag"

* Add USE_OP of all operators and kernels and remove ARCHIVE_START/END in CMakeLists.txt of inference unittests.

* Remove ARCHIVE_START/END when linking inference shared library.

* Disable some fluid related cmake operations for cross-compiling.

* Update source files.

* Update headers

* Update

* Update

* Update

* Update

* Fix a CMake dependency

1. Add option WITH_AMD_GPU.
2. Add cmake/hip.cmake for HIP toolchain.
3. Some external module such as eigen may need HIP port.
4. Add macro hip_library/hip_binary/hip_test to cmake/generic.cmake.
5. Add one HIP source concat.hip.cu as an example. Each .cu may have its corresponding .hip.cu.

1. Add option WITH_AMD_GPU.
2. Add cmake/hip.cmake for HIP toolchain.
3. Some external module such as eigen may need HIP port.
4. Add macro hip_library/hip_binary/hip_test to cmake/generic.cmake.
5. Add one HIP source concat.hip.cu as an example. Each .cu may have its corresponding .hip.cu.