additional check for fine grain support in p2pCanConnect (#146)

[ROCm/rccl commit: 5a502955c9]

projects/rccl/src/init.cc

@@ -171,12 +171,15 @@ void parseHsaForceFineGrainVramPcie() {
     errno = 0;
     int64_t v = strtoll(str, NULL, 0);
     if (errno || (v != 0 && v != 1)) {
-      INFO(NCCL_ALL,"Invalid value %s for %s, using default %u.", str, "HSA_FORCE_FINE_GRAIN_PCIE", useFineGrainVramPcie); \
+      INFO(NCCL_ALL,"Invalid value %s for %s, using default %u.", str, "HSA_FORCE_FINE_GRAIN_PCIE", useFineGrainVramPcie);
     } else {
       useFineGrainVramPcie = v;
-      INFO(NCCL_ALL,"%s set by environment to %u.", "HSA_FORCE_FINE_GRAIN_PCIE", useFineGrainVramPcie);  \
+      INFO(NCCL_ALL,"%s set by environment to %u.", "HSA_FORCE_FINE_GRAIN_PCIE", useFineGrainVramPcie);
     }
   }
+  else {
+    INFO(NCCL_ALL,"%s not set by environment.", "HSA_FORCE_FINE_GRAIN_PCIE");
+  }
 }
 
 pthread_mutex_t initLock = PTHREAD_MUTEX_INITIALIZER;

projects/rccl/src/transport/p2p.cc

@@ -112,6 +112,9 @@ ncclResult_t p2pCanConnect(ncclTvalue_t* ret, struct ncclPeerInfo* myInfo, struc
 
   // Do not detect topology if we're on the same GPU. Note this is not really supported.
   if (myInfo->cudaDev == peerCudaDev) {
+#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
+    if (!useFineGrainVramPcie) return ncclSuccess;
+#endif
     *ret = 1 + PATH_SYS;
     return ncclSuccess;
   }