Compiler warnings fix 2 (#1801)

* Changes to device code

* Changes to src/misc

* Changes to graph

* src/include changes

* src/transport changes

* changes in init, enqueue, proxy

* Changes to CMakeLists.txt

* Additional changes to device code

* Additional changes to net.cc

* adding 'compiler warning' tag to ease upstream merge'

* typo correction

* Addessing comments

* Additional changes for new commits

[ROCm/rccl commit: 3f8cac388e]

projects/rccl/src/init.cc

@@ -709,7 +709,7 @@ static ncclResult_t devCommSetup(ncclComm_t comm) {
   int nRanks = comm->nRanks;
   struct ncclDevCommAndChannels tmpCommAndChans;
   struct ncclDevCommAndChannels *devCommAndChans = NULL;
-  struct ncclNvmlCCStatus ccStatus;
+  //struct ncclNvmlCCStatus ccStatus; //unused variable - compiler warning
   bool ccEnable = false;
   cudaStream_t deviceStream;
 
@@ -1882,9 +1882,6 @@ fail:
 static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
   struct ncclCommInitRankAsyncJob* job = (struct ncclCommInitRankAsyncJob*)job_;
   ncclComm_t comm = job->comm;
-#ifdef ENABLE_MSCCLPP
-  ncclUniqueId origUniqueId = *job->commId;
-#endif
   ncclResult_t res = ncclSuccess;
   int archMajor, archMinor;
   size_t maxLocalSizeBytes = 0;
@@ -1895,8 +1892,10 @@ static ncclResult_t ncclCommInitRankFunc(struct ncclAsyncJob* job_) {
   double sum_timers = 0;
   uint64_t timers[TIMERS_INIT_COUNT] = {0};
   unsigned long long commIdHash;
+  #ifdef USE_INDIRECT_FUNCTION_CALL
   int64_t stackSize;
   hipDeviceProp_t devProp;
+  #endif
 
   timers[TIMER_INIT_TOTAL] = clockNano();
   CUDACHECKGOTO(cudaSetDevice(cudaDev), res, fail);