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>

12 files changed, 63 insertions, 12 deletions

diff --git a/arm_compute/graph/IDeviceBackend.h b/arm_compute/graph/IDeviceBackend.h
index 3a0a7e634e..f84aac0ae0 100644
--- a/arm_compute/graph/IDeviceBackend.h
+++ b/arm_compute/graph/IDeviceBackend.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019,2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -118,6 +118,8 @@ public:
      * @return Weights manager
      */
     virtual std::shared_ptr<arm_compute::IWeightsManager> create_weights_manager() = 0;
+    /** Synchronize kernels execution on the backend. On GPU, this results in a blocking call waiting for all kernels to be completed. */
+    virtual void sync() = 0;
 };
 } // namespace backends
 } // namespace graph
diff --git a/arm_compute/graph/ITensorAccessor.h b/arm_compute/graph/ITensorAccessor.h
index 96bd499573..a8818be2e5 100644
--- a/arm_compute/graph/ITensorAccessor.h
+++ b/arm_compute/graph/ITensorAccessor.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019,2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,6 +45,14 @@ public:
      * @return True if access is successful else false
      */
     virtual bool access_tensor(ITensor &tensor) = 0;
+    /** Returns true if the tensor data is being accessed
+     *
+     * @return True if the tensor data is being accessed by the accessor. False otherwise
+     */
+    virtual bool access_tensor_data()
+    {
+        return true;
+    }
 };
 
 using ITensorAccessorUPtr = std::unique_ptr<ITensorAccessor>;
diff --git a/arm_compute/graph/Utils.h b/arm_compute/graph/Utils.h
index 6c3b76839e..43a8eca9e3 100644
--- a/arm_compute/graph/Utils.h
+++ b/arm_compute/graph/Utils.h
@@ -107,6 +107,8 @@ void setup_requested_backend_context(GraphContext &ctx, Target target);
  * @param[in,out] ctx Graph Context
  */
 void release_default_graph_context(GraphContext &ctx);
+/** Synchronize kernels execution on the backends. On GPU, this results in a blocking call waiting for all kernels to be completed. */
+void sync_backends();
 /** Get size of a tensor's given dimension depending on its layout
  *
  * @param[in] descriptor            Descriptor
diff --git a/arm_compute/graph/backends/CL/CLDeviceBackend.h b/arm_compute/graph/backends/CL/CLDeviceBackend.h
index 278a8e5031..63674ad794 100644
--- a/arm_compute/graph/backends/CL/CLDeviceBackend.h
+++ b/arm_compute/graph/backends/CL/CLDeviceBackend.h
@@ -28,8 +28,8 @@
 
 #include "arm_compute/runtime/CL/CLBufferAllocator.h"
 #include "arm_compute/runtime/CL/CLGEMMHeuristicsHandle.h"
-#include "arm_compute/runtime/CL/CLTypes.h"
 #include "arm_compute/runtime/CL/CLTuner.h"
+#include "arm_compute/runtime/CL/CLTypes.h"
 
 namespace arm_compute
 {
@@ -70,6 +70,7 @@ public:
     Status validate_node(INode &node) override;
     std::shared_ptr<arm_compute::IMemoryManager> create_memory_manager(MemoryManagerAffinity affinity) override;
     std::shared_ptr<arm_compute::IWeightsManager> create_weights_manager() override;
+    void                                          sync() override;
 
 private:
     int                                _context_count;   /**< Counts how many contexts are currently using the backend */
diff --git a/arm_compute/graph/backends/NEON/NEDeviceBackend.h b/arm_compute/graph/backends/NEON/NEDeviceBackend.h
index 0b343c0d67..9cb37d4553 100644
--- a/arm_compute/graph/backends/NEON/NEDeviceBackend.h
+++ b/arm_compute/graph/backends/NEON/NEDeviceBackend.h
@@ -52,6 +52,7 @@ public:
     Status validate_node(INode &node) override;
     std::shared_ptr<arm_compute::IMemoryManager> create_memory_manager(MemoryManagerAffinity affinity) override;
     std::shared_ptr<arm_compute::IWeightsManager> create_weights_manager() override;
+    void                                          sync() override;
 
 private:
     Allocator _allocator; /**< Backend allocator */
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
diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp
index 20e1369e73..c3f71299f6 100644
--- a/utils/GraphUtils.cpp
+++ b/utils/GraphUtils.cpp
@@ -162,6 +162,11 @@ DummyAccessor::DummyAccessor(unsigned int maximum)
 {
 }
 
+bool DummyAccessor::access_tensor_data()
+{
+    return false;
+}
+
 bool DummyAccessor::access_tensor(ITensor &tensor)
 {
     ARM_COMPUTE_UNUSED(tensor);
diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h
index d2b05f27c6..80055acc0f 100644
--- a/utils/GraphUtils.h
+++ b/utils/GraphUtils.h
@@ -135,6 +135,7 @@ public:
     DummyAccessor(DummyAccessor &&) = default;
 
     // Inherited methods overriden:
+    bool access_tensor_data() override;
     bool access_tensor(ITensor &tensor) override;
 
 private: