Add ROCm specific documentation

Ported from TensorFlow 1.3 ROCm port. Still needs to be revised to reflect the
status of TensorFlow 1.8.

Author: whchung

README.ROCm.md

@@ -0,0 +1,53 @@
+# Tensorflow ROCm port #
+
+## Introduction ##
+
+This repository hosts the port of [Tensorflow](https://github.com/tensorflow/tensorflow) on ROCm platform. It uses various technologies on ROCm platform such as HIP and MIOpen. For details on HIP, please refer [here](https://github.com/GPUOpen-ProfessionalCompute-Tools/HIP). Optimized DNN library calls (via [MIOpen](https://github.com/ROCmSoftwarePlatform/MIOpen)) are also supported within this codebase.
+
+## Installation ##
+
+For further background information on ROCm, refer [here](https://github.com/RadeonOpenCompute/ROCm/blob/master/README.md).
+
+The project is derived from TensorFlow 1.3.0 and has been verified to work with the latest ROCm 1.7.1 release.
+
+For details on installation and usage, see these links:
+* [Basic installation](rocm_docs/tensorflow-install-basic.md)
+* [Building from source](rocm_docs/tensorflow-build-from-source.md)
+* [Quickstart guide](rocm_docs/tensorflow-quickstart.md)
+
+
+## Technical details ##
+* [Overview of ROCm port](rocm_docs/rocm-port-overview.md)
+* [List of supported operators on ROCm](rocm_docs/core_kernels.md)
+
+
+## Known Issues / Workarounds
+
+### X freezes under load
+ROCm 1.7.1 a kernel parameter `noretry` has been set to 1 to improve overall system performance. However it has been proven to bring instability to graphics driver shipped with Ubuntu. This is an ongoing issue and we are looking into it.
+
+Before that, please try apply this change by changing `noretry` bit to 0.
+
+```
+echo 0 | sudo tee /sys/module/amdkfd/parameters/noretry
+```
+
+Files under `/sys` won't be preserved after reboot so you'll need to do it every time.
+
+One way to keep `noretry=0` is to change `/etc/modprobe.d/amdkfd.conf` and make it be:
+
+```
+options amdkfd noretry=0
+```
+
+Once it's done, run `sudo update-initramfs -u`. Reboot and verify `/sys/module/amdkfd/parameters/noretry` stays as 0.
+
+For more information please check this [issue report](https://github.com/ROCmSoftwarePlatform/tensorflow/issues/13)
+
+### tensorflow/benchmarks workaround for TF v1.3
+Since our current port of TF supports v1.3, we can't use some of the newest commits in `tensorflow/benchmarks`.  One specific error you could observe is a `ImportError: cannot import name batching`.  "Batching" was introduced in this [commit](https://github.com/tensorflow/benchmarks/commit/82dd0539c76afa8491e50d8f796e686b4d97b988). Also RCCL, a ROCm version of NCCL, is under implementation. Therefore we have to drop back to an earlier point in the commit history, and disable NCCL.
+
+```
+git checkout -b sep7 6a33b4a4b5bda950bb7e45faf13120115cbfdb2f
+sed -i 's|from tensorflow.contrib import nccl|#from tensorflow.contrib import nccl|g' ./scripts/tf_cnn_benchmarks/variable_mgr.py
+```