Remove map/unmap overhead for input/output accessor when using DummyAccessor

Don't map/unmap when the tensor data is not accessed in the input or output accessor.
This is so to avoid measuring CPU overhead when benchmarking on the GPU backend.

Resolve COMPMID-4712

Change-Id: I5baba1b93e7a51fe13525bcce6c0cfdecb14493e
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6140
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 40 insertions, 9 deletions

diff --git a/src/graph/Tensor.cpp b/src/graph/Tensor.cpp
index f69d49d016..3d4723430f 100644
--- a/src/graph/Tensor.cpp
+++ b/src/graph/Tensor.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019,2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -80,20 +80,28 @@ bool Tensor::call_accessor()
         return false;
     }
 
-    // Map tensor
-    _handle->map(true);
+    const bool access_data = _accessor->access_tensor_data();
 
-    // Return in case of null backend buffer
-    if(_handle->tensor().buffer() == nullptr)
+    if(access_data)
     {
-        return false;
+        // Map tensor
+        _handle->map(true);
+
+        // Return in case of null backend buffer
+        if(_handle->tensor().buffer() == nullptr)
+        {
+            return false;
+        }
     }
 
     // Call accessor
     bool retval = _accessor->access_tensor(_handle->tensor());
 
-    // Unmap tensor
-    _handle->unmap();
+    if(access_data)
+    {
+        // Unmap tensor
+        _handle->unmap();
+    }
 
     return retval;
 }
diff --git a/src/graph/Utils.cpp b/src/graph/Utils.cpp
index 37af1bff93..ceb5ba95de 100644
--- a/src/graph/Utils.cpp
+++ b/src/graph/Utils.cpp
@@ -116,6 +116,17 @@ void release_default_graph_context(GraphContext &ctx)
     }
 }
 
+void sync_backends()
+{
+    for(const auto &backend : backends::BackendRegistry::get().backends())
+    {
+        if(backend.second->is_backend_supported())
+        {
+            backend.second->sync();
+        }
+    }
+}
+
 void setup_requested_backend_context(GraphContext &ctx, Target target)
 {
     if(backends::BackendRegistry::get().contains(target))
diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp
index b6b25cc7d0..01e5ab1730 100644
--- a/src/graph/backends/CL/CLDeviceBackend.cpp
+++ b/src/graph/backends/CL/CLDeviceBackend.cpp
@@ -221,6 +221,11 @@ std::shared_ptr<arm_compute::IWeightsManager> CLDeviceBackend::create_weights_ma
     auto weights_mgr = std::make_shared<IWeightsManager>();
     return weights_mgr;
 }
+
+void CLDeviceBackend::sync()
+{
+    CLScheduler::get().sync();
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
diff --git a/src/graph/backends/NEON/NEDeviceBackend.cpp b/src/graph/backends/NEON/NEDeviceBackend.cpp
index 9efa3ac0c8..47ef2c2394 100644
--- a/src/graph/backends/NEON/NEDeviceBackend.cpp
+++ b/src/graph/backends/NEON/NEDeviceBackend.cpp
@@ -175,6 +175,11 @@ std::shared_ptr<arm_compute::IWeightsManager> NEDeviceBackend::create_weights_ma
     auto weights_mgr = std::make_shared<IWeightsManager>();
     return weights_mgr;
 }
+
+void NEDeviceBackend::sync()
+{
+    // nop
+}
 } // namespace backends
 } // namespace graph
 } // namespace arm_compute
diff --git a/src/graph/detail/ExecutionHelpers.cpp b/src/graph/detail/ExecutionHelpers.cpp
index 5be3706cfe..ac800df76c 100644
--- a/src/graph/detail/ExecutionHelpers.cpp
+++ b/src/graph/detail/ExecutionHelpers.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -265,6 +265,8 @@ bool call_all_output_node_accessors(ExecutionWorkload &workload)
         is_valid          = is_valid && valid_output;
     });
 
+    sync_backends();
+
     return is_valid;
 }
 } // namespace detail