[perf]hipMalloc performance optimization

Change-Id: I6e8a918cc1c4cafad197b09e10755cd180e11ead

rocclr/platform/command.cpp

@@ -595,9 +595,22 @@ bool TransferBufferFileCommand::validateMemory() {
 
 bool CopyMemoryP2PCommand::validateMemory() {
   amd::Device* queue_device = &queue()->device();
+
   // Rocr backend maps memory from different devices by default and runtime doesn't need to track
   // extra memory objects. Also P2P staging buffer always allocated
   if (queue_device->settings().rocr_backend_) {
+    // Explicit allow access is needed for P2P access
+    const std::vector<Device*>& srcDevices = memory1_->getContext().devices();
+    const std::vector<Device*>& dstDevices = memory2_->getContext().devices();
+    if (srcDevices.size() == 1 && dstDevices.size() == 1) {
+      device::Memory* mem2 = memory2_->getDeviceMemory(*dstDevices[0]);
+      if (!mem2->getAllowedPeerAccess()) {
+        void* dst = mem2->owner()->getSvmPtr();
+        bool status = dstDevices[0]->deviceAllowAccess(dst);
+        mem2->setAllowedPeerAccess(true);
+        return status;
+      }
+    }
     return true;
   }