Dependency declaration issue in mpi.so build

[aburden5]

Hi, I am new to bazel, attempting to build TensorFlow-AllReduce w/o CUDA from source.

(git rev-parse HEAD: 44302f0961e76002989523c2f424a011e54fd806)

I am using Bazel 0.4.5, but have also gotten same error when using 0.4.3. I am in RHEL 4.8.5-4.

when I try "bazel build -c opt :mpi.so" I get the following error:

ERROR: /home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/contrib/mpi/BUILD:13:1: undeclared inclusion(s) in rule '//tensorflow/contrib/mpi:mpi.so': this rule is missing dependency declarations for the following files included by 'tensorflow/contrib/mpi/mpi_ops.cc': '/home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/stream_executor/lib/statusor.h' '/home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/stream_executor/platform/port.h' '/home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/stream_executor/lib/error.h' '/home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/stream_executor/lib/status.h' '/home_nfs/aburdenx/apps/tensorflow-allreduce/tensorflow/stream_executor/platform/logging.h'.

I have tried several fixes, but am not having any success, including:

-> Tried adding a dependency declaration in tensorflow-allreduce/tensorflow/contrib/mpi/BUILD, as follows:

tf_custom_op_library( name = "mpi.so", srcs = ["mpi_ops.cc", "ring.cc", "ring.h"], gpu_srcs = ["ring.cu.cc", "ring.h"], deps = [ "//third_party/mpi:mpi", ":mpi_message_proto_cc", "//tensorflow/stream_executor:stream_executor", ], )

but I get this error:

tensorflow/stream_executor:stream_executor cannot depend on tensorflow/core:lib.

-> Tried adding --copt="-Itensorflow/stream_executor/lib" to bazel build with no luck

The "undeclared inclusion(s)" error seems to crop up mostly in CUDA builds due to missing path to cuda include dir, but this is clearly different issue that I haven't seen solutions for.

I will greatly appreciate any help moving forward.

Thanks!

[rhdong]

Maybe ,you should remove the "tensorflow/stream_executor:stream_executor" and add "--config=cuda" to your bazeel build opt.: bazel build -c opt --config=cuda tensorflow/tools/pip_package:build_pip_package