rocr: Fixes memory allocation issue

Fixes a bug in memory allocation in which dmabuf export only works on
GPU 0 in a multi-GPU environment.


[ROCm/ROCR-Runtime commit: 7b8a6f8ca2]

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/amd_gpu_agent.cpp

@@ -648,10 +648,13 @@ hsa_status_t GpuAgent::VisitRegion(bool include_peer,
                                    void* data) const {
   if (include_peer) {
     // Only expose system, local, and LDS memory of the blit agent.
-    if (this->node_id() == core::Runtime::runtime_singleton_->region_gpu()->node_id()) {
+    const auto& gpu_ids = core::Runtime::runtime_singleton_->gpu_ids();
-      hsa_status_t stat = VisitRegion(regions_, callback, data);
+    for (auto& gpu_id : gpu_ids) {
-      if (stat != HSA_STATUS_SUCCESS) {
+      if (this->node_id() == gpu_id) {
-        return stat;
+        hsa_status_t stat = VisitRegion(regions_, callback, data);
+        if (stat != HSA_STATUS_SUCCESS) {
+          return stat;
+        }
       }
     }