From 5e98b012597c13dc0acb048bd10a84a2a397a346 Mon Sep 17 00:00:00 2001
From: Matthew Bentham <matthew.bentham@arm.com>
Date: Fri, 24 Jan 2020 23:11:43 +0000
Subject: Clean up header usage a bit in NEON backend

Including NEFunctions.h is unnecessary and adds about a second
to compile time each translation unit in which it appears,
so we should use just the header file with the arm compute function
declarations that we need.

Signed-off-by: Matthew Bentham <matthew.bentham@arm.com>
Change-Id: I605d0eb82ccf2aafa35381a5d9d54337d3fe17a7
---
 src/backends/neon/workloads/NeonDequantizeWorkload.cpp |  9 ++++++---
 src/backends/neon/workloads/NeonDequantizeWorkload.hpp |  5 +++--
 src/backends/neon/workloads/NeonSplitterWorkload.cpp   | 10 ++++++----
 src/backends/neon/workloads/NeonSplitterWorkload.hpp   |  5 +++--
 4 files changed, 18 insertions(+), 11 deletions(-)

diff --git a/src/backends/neon/workloads/NeonDequantizeWorkload.cpp b/src/backends/neon/workloads/NeonDequantizeWorkload.cpp
index aa454c921b..8b229a1cda 100644
--- a/src/backends/neon/workloads/NeonDequantizeWorkload.cpp
+++ b/src/backends/neon/workloads/NeonDequantizeWorkload.cpp
@@ -7,6 +7,8 @@
 
 #include "NeonWorkloadUtils.hpp"
 
+#include <arm_compute/runtime/NEON/functions/NEDequantizationLayer.h>
+
 #include <aclCommon/ArmComputeTensorUtils.hpp>
 #include <backendsCommon/CpuTensorHandle.hpp>
 #include <neon/NeonTensorHandle.hpp>
@@ -33,9 +35,10 @@ NeonDequantizeWorkload::NeonDequantizeWorkload(const DequantizeQueueDescriptor&
     arm_compute::ITensor& input = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ITensor& output = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
-    m_Layer.reset(new arm_compute::NEDequantizationLayer());
-    m_Layer->configure(&input, &output);
-    m_Layer->prepare();
+    std::unique_ptr<arm_compute::NEDequantizationLayer> layer(new arm_compute::NEDequantizationLayer());
+    layer->configure(&input, &output);
+    layer->prepare();
+    m_Layer.reset(layer.release());
 }
 
 void NeonDequantizeWorkload::Execute() const
diff --git a/src/backends/neon/workloads/NeonDequantizeWorkload.hpp b/src/backends/neon/workloads/NeonDequantizeWorkload.hpp
index a6616376e7..63ea783474 100644
--- a/src/backends/neon/workloads/NeonDequantizeWorkload.hpp
+++ b/src/backends/neon/workloads/NeonDequantizeWorkload.hpp
@@ -7,7 +7,8 @@
 
 #include <backendsCommon/Workload.hpp>
 
-#include <arm_compute/runtime/NEON/NEFunctions.h>
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/IFunction.h>
 
 #include <functional>
 
@@ -25,7 +26,7 @@ public:
     void Execute() const override;
 
 private:
-    mutable std::unique_ptr<arm_compute::NEDequantizationLayer> m_Layer;
+    mutable std::unique_ptr<arm_compute::IFunction> m_Layer;
 };
 
 } //namespace armnn
diff --git a/src/backends/neon/workloads/NeonSplitterWorkload.cpp b/src/backends/neon/workloads/NeonSplitterWorkload.cpp
index 5b5614662d..224e97af2d 100644
--- a/src/backends/neon/workloads/NeonSplitterWorkload.cpp
+++ b/src/backends/neon/workloads/NeonSplitterWorkload.cpp
@@ -5,13 +5,14 @@
 
 #include "NeonSplitterWorkload.hpp"
 
-#include "NeonWorkloadUtils.hpp"
+#include <arm_compute/runtime/NEON/functions/NESplit.h>
 
 #include <aclCommon/ArmComputeTensorUtils.hpp>
 #include <aclCommon/ArmComputeUtils.hpp>
 #include <backendsCommon/CpuTensorHandle.hpp>
 #include <neon/NeonTensorHandle.hpp>
 
+#include "NeonWorkloadUtils.hpp"
 
 namespace armnn
 {
@@ -83,7 +84,7 @@ NeonSplitterWorkload::NeonSplitterWorkload(const SplitterQueueDescriptor& descri
     }
 
     // Create the layer function
-    m_Layer.reset(new arm_compute::NESplit());
+    std::unique_ptr<arm_compute::NESplit> layer(new arm_compute::NESplit());
 
     // Configure input and output tensors
     std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape());
@@ -93,10 +94,11 @@ NeonSplitterWorkload::NeonSplitterWorkload(const SplitterQueueDescriptor& descri
     }
 
     unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin());
-    m_Layer->configure(&input, aclOutputs, aclAxis);
+    layer->configure(&input, aclOutputs, aclAxis);
 
     // Prepare
-    m_Layer->prepare();
+    layer->prepare();
+    m_Layer.reset(layer.release());
 }
 
 void NeonSplitterWorkload::Execute() const
diff --git a/src/backends/neon/workloads/NeonSplitterWorkload.hpp b/src/backends/neon/workloads/NeonSplitterWorkload.hpp
index f9025663ca..66349f947f 100644
--- a/src/backends/neon/workloads/NeonSplitterWorkload.hpp
+++ b/src/backends/neon/workloads/NeonSplitterWorkload.hpp
@@ -7,7 +7,8 @@
 
 #include <backendsCommon/Workload.hpp>
 
-#include <arm_compute/runtime/NEON/NEFunctions.h>
+#include <arm_compute/core/Error.h>
+#include <arm_compute/runtime/IFunction.h>
 
 #include <functional>
 
@@ -26,7 +27,7 @@ public:
     void Execute() const override;
 
 private:
-    mutable std::unique_ptr<arm_compute::NESplit> m_Layer;
+    mutable std::unique_ptr<arm_compute::IFunction> m_Layer;
 };
 
 } //namespace armnn
-- 
cgit v1.2.1