added rocm7 support to EnablePeerAccess (#347)

* added rocm7 support to EnablePeerAccess

* use wrap namespace, clang-format and add comments

Author: amd-songpiao

xla/stream_executor/rocm/rocm_driver_wrapper.h

@@ -51,7 +51,7 @@ namespace wrap {
     static FuncPtrT loaded = []() -> FuncPtrT {                             \
       static const char *kName = TO_STR(hipSymbolName);                     \
       void *f;                                                              \
-      auto s = tsl::Env::Default()->GetSymbolFromLibrary(                   \
+      auto s = tsl::Env::Default() -> GetSymbolFromLibrary(                 \
           tsl::internal::CachedDsoLoader::GetHipDsoHandle().value(), kName, \
           &f);                                                              \
       CHECK(s.ok()) << "could not find " << kName                           \
@@ -100,6 +100,7 @@ namespace wrap {
   __macro(hipGetDeviceCount)                        \
   __macro(hipGetDeviceProperties)                   \
   __macro(hipGetErrorString)                        \
+  __macro(hipGetLastError)                          \
   __macro(hipGraphAddKernelNode)                    \
   __macro(hipGraphAddChildGraphNode)                \
   __macro(hipGraphAddEmptyNode)                     \

xla/stream_executor/rocm/rocm_executor.cc

@@ -378,11 +378,16 @@ absl::Status EnablePeerAccess(Context* from, Context* to) {
   hipError_t result =
       wrap::hipDeviceEnablePeerAccess(to->device_ordinal(), 0 /* = flags */);
 
-  if (result != hipSuccess && result != hipErrorPeerAccessAlreadyEnabled) {
+  if (result == hipErrorPeerAccessAlreadyEnabled) {
+    // hipGetLastError is used to reset per thread error state,
+    // as hipGetLastError would get the recent error code since rocm7 even the
+    // last call is successful.
+    (void)wrap::hipGetLastError();
+  } else if (result != hipSuccess) {
     return absl::InternalError(
         absl::StrFormat("failed to enable peer access from %d to %d: %s",
                         from->device_ordinal(), to->device_ordinal(),
-                        ToString(result).c_str()));
+                        wrap::hipGetErrorString(result)));
   }
 
   return absl::OkStatus();