Add note on ifdefing based on CUDA_VERSION for ROCm path (pytorch#62850)

Summary:
CUDA_VERSION and HIP_VERSION follow very unrelated versioning schemes, so it does not make sense to use CUDA_VERSION to determine the ROCm path. This note explicitly addresses it.

Pull Request resolved: pytorch#62850

Reviewed By: mruberry

Differential Revision: D30547562

Pulled By: malfet

fbshipit-source-id: 02990fa66a88466c2330ab85f446b25b78545150

docs/source/notes/hip.rst

@@ -119,6 +119,27 @@ torch.distributed backends
 
 Currently, only the "nccl" and "gloo" backends for torch.distributed are supported on ROCm.
 
+.. _cuda-api-to_hip-api-mappings:
+
+CUDA API to HIP API mappings in C++
+-----------------------------------
+
+Please refer: https://rocmdocs.amd.com/en/latest/Programming_Guides/HIP_API_Guide.html
+
+NOTE: The CUDA_VERSION macro, cudaRuntimeGetVersion and cudaDriverGetVersion APIs do not
+semantically map to the same values as HIP_VERSION macro, hipRuntimeGetVersion and
+hipDriverGetVersion APIs. Please do not use them interchangeably when doing version checks.
+
+Eg: Instead of
+#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
+If it is desired to not take the code path for ROCm/HIP:
+#if defined(CUDA_VERSION) && CUDA_VERSION >= 11000 && !defined(USE_ROCM)
+If it is desired to take the code path for ROCm/HIP:
+#if (defined(CUDA_VERSION) && CUDA_VERSION >= 11000) || defined(USE_ROCM)
+If it is desired to take the code path for ROCm/HIP only for specific HIP versions:
+#if (defined(CUDA_VERSION) && CUDA_VERSION >= 11000) || (defined(USE_ROCM) && ROCM_VERSION >= 40300)
+
+
 Refer to CUDA Semantics doc
 ---------------------------