arm_compute v18.02

Change-Id: I7207aa488e5470f235f39b6c188b4678dc38d1a6
diff --git a/src/graph/Graph.cpp b/src/graph/Graph.cpp
index ac5316f..b6c6822 100644
--- a/src/graph/Graph.cpp
+++ b/src/graph/Graph.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -62,6 +62,7 @@
     std::unique_ptr<INode>                      _current_node{ nullptr };
     ITensorObject                              *_current_output{ nullptr };
     bool                                        _info_enabled{ false };
+    CLTuner                                     _tuner{};
 
 private:
     ITensorObject *_current_input{ nullptr };
@@ -76,10 +77,22 @@
 Graph::Graph()
     : _pimpl{ new Private() }
 {
+    graph_init();
+}
+
+void Graph::graph_init(const bool use_cl_tuner)
+{
     // Check if OpenCL is available and initialize the scheduler
     if(opencl_is_available())
     {
-        arm_compute::CLScheduler::get().default_init();
+        if(use_cl_tuner)
+        {
+            arm_compute::CLScheduler::get().default_init(&_pimpl->_tuner);
+        }
+        else
+        {
+            arm_compute::CLScheduler::get().default_init();
+        }
     }
 }
 
@@ -119,6 +132,11 @@
         _previous_hints = _current_hints; // For the first node just assume the previous node was of the same type as this one
     }
 
+    if(_current_node->supports_in_place())
+    {
+        _current_output = _current_input;
+    }
+
     //Automatic output configuration ?
     if(_current_output == nullptr)
     {
@@ -140,8 +158,12 @@
     _ctx.hints()                                 = _current_hints;
     std::unique_ptr<arm_compute::IFunction> func = _current_node->instantiate_node(_ctx, _current_input, _current_output);
 
-    // Allocate current input
-    _current_input->allocate();
+    // If the operation is done in-place, do not allocate or it will prevent following layers from performing the configuration
+    if(!_current_node->supports_in_place())
+    {
+        // Allocate current input
+        _current_input->allocate();
+    }
 
     // Map input if needed
     if(_current_input->target() == TargetHint::OPENCL)
@@ -215,11 +237,25 @@
         _pimpl->_graph_output->allocate();
     }
 }
+
 bool Graph::opencl_is_available()
 {
     return arm_compute::opencl_is_available();
 }
 
+arm_compute::GPUTarget Graph::gpu_target()
+{
+    // Check if OpenCL is available before returning the GPU target
+    if(opencl_is_available())
+    {
+        return arm_compute::CLScheduler::get().target();
+    }
+    else
+    {
+        return GPUTarget::MIDGARD;
+    }
+}
+
 void Graph::set_temp(TensorInfo &&tmp)
 {
     ARM_COMPUTE_ERROR_ON(_pimpl->_graph_input == nullptr);
diff --git a/src/graph/INode.cpp b/src/graph/INode.cpp
index 582f936..c753f66 100644
--- a/src/graph/INode.cpp
+++ b/src/graph/INode.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,6 +39,14 @@
     ARM_COMPUTE_ERROR_ON(target_hint == TargetHint::OPENCL && !opencl_is_available());
     return target_hint;
 }
+bool INode::supports_in_place() const
+{
+    return _supports_in_place;
+}
+void INode::set_supports_in_place(bool value)
+{
+    _supports_in_place = value;
+}
 GraphHints INode::node_override_hints(GraphHints hints) const
 {
     TargetHint target_hint = hints.target_hint();
diff --git a/src/graph/SubGraph.cpp b/src/graph/SubGraph.cpp
index 8ba2af6..4065e1d 100644
--- a/src/graph/SubGraph.cpp
+++ b/src/graph/SubGraph.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -66,6 +66,10 @@
     }
     graph->add_tensor_object(std::move(_input));
 
+    // Make sure first and last nodes of the subgraph always do operations out-of-place
+    _nodes.front()->set_supports_in_place(false);
+    _nodes.back()->set_supports_in_place(false);
+
     // Construct nodes
     for(auto &node : _nodes)
     {
diff --git a/src/graph/nodes/ActivationLayer.cpp b/src/graph/nodes/ActivationLayer.cpp
index 54f30ef..546c42a 100644
--- a/src/graph/nodes/ActivationLayer.cpp
+++ b/src/graph/nodes/ActivationLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -33,6 +33,7 @@
 ActivationLayer::ActivationLayer(const ActivationLayerInfo activation_info)
     : _activation_info(activation_info)
 {
+    set_supports_in_place(true);
 }
 
 std::unique_ptr<arm_compute::IFunction> ActivationLayer::instantiate_node(GraphContext &ctx, ITensorObject *input, ITensorObject *output)
diff --git a/src/graph/nodes/BatchNormalizationLayer.cpp b/src/graph/nodes/BatchNormalizationLayer.cpp
index 7851aa5..24287ac 100644
--- a/src/graph/nodes/BatchNormalizationLayer.cpp
+++ b/src/graph/nodes/BatchNormalizationLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -77,6 +77,7 @@
     node_ctx.add_input(_gamma.tensor());
     node_ctx.add_output(out);
     node_ctx.add_parameter<float>("epsilon", _epsilon);
+    node_ctx.add_parameter<ActivationLayerInfo>("act_info", _act_info);
 
     // Configure operation
     auto func = OperationRegistry::get().find_operation(OperationType::BatchNormalizationLayer, _target_hint)->configure(node_ctx);
diff --git a/src/graph/nodes/DepthwiseConvolutionLayer.cpp b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
index 1209d03..e5101cc 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayer.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
@@ -40,10 +40,8 @@
 
     if(_weights.tensor() == nullptr)
     {
-        TensorShape shape = in->info()->tensor_shape();
-        shape.set(Window::DimX, _conv_width);
-        shape.set(Window::DimY, _conv_height);
-        TensorInfo info = TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position());
+        TensorShape weights_shape(_conv_width, _conv_height, input->tensor()->info()->tensor_shape().z());
+        TensorInfo  info = TensorInfo(TensorShape(weights_shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position());
         info.set_quantization_info(_quant_info);
         _weights.set_info(std::move(info));
     }
diff --git a/src/graph/nodes/FullyConnectedLayer.cpp b/src/graph/nodes/FullyConnectedLayer.cpp
index 219e0f9..3742150 100644
--- a/src/graph/nodes/FullyConnectedLayer.cpp
+++ b/src/graph/nodes/FullyConnectedLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/graph/nodes/ResidualLayer.cpp b/src/graph/nodes/ResidualLayer.cpp
new file mode 100644
index 0000000..87404f9
--- /dev/null
+++ b/src/graph/nodes/ResidualLayer.cpp
@@ -0,0 +1,199 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/nodes/ResidualLayer.h"
+
+#include "arm_compute/graph/Error.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/NodeContext.h"
+#include "arm_compute/graph/OperationRegistry.h"
+#include "arm_compute/graph/SubGraph.h"
+#include "arm_compute/graph/Tensor.h"
+#include "arm_compute/runtime/IFunction.h"
+#include "support/ToolchainSupport.h"
+#include "utils/Utils.h"
+
+#include <memory>
+#include <tuple>
+#include <vector>
+
+using namespace arm_compute::graph;
+
+/** Residual function */
+class ResidualFunction final : public arm_compute::IFunction
+{
+public:
+    /** Default Constructor */
+    ResidualFunction(GraphContext &ctx, ITensorObject *output)
+        : _ctx(ctx), _input(nullptr), _output(output), _func(nullptr), _graphs(), _graph_outputs()
+    {
+    }
+
+    /** Prevent instances from being copy constructed */
+    ResidualFunction(const ResidualFunction &) = delete;
+    /** Prevent instances from being copy assigned */
+    const ResidualFunction &operator=(const ResidualFunction &) = delete;
+    /** Prevent instances from being move constructed */
+    ResidualFunction(ResidualFunction &&) = delete;
+    /** Prevent instances from being move assigned */
+    ResidualFunction &operator=(ResidualFunction &&) = delete;
+    /** Default destructor */
+    ~ResidualFunction() override = default;
+
+    /** Set the input (when using only one sub graph)
+     *
+     * @param[in] input Input to set
+     */
+    void set_input(std::unique_ptr<ITensorObject> input)
+    {
+        _input = std::move(input);
+    }
+
+    /** Registers graph to be executed by the residual function
+     *
+     * @param[in] graph  Graph to register
+     * @param[in] output Output to register
+     */
+    void register_graph(std::unique_ptr<Graph> graph, std::unique_ptr<ITensorObject> output)
+    {
+        _graphs.push_back(std::move(graph));
+        _graph_outputs.push_back(std::move(output));
+    }
+
+    /** Configure the function */
+    void configure()
+    {
+        ARM_COMPUTE_ERROR_ON(_graphs.size() < 1 || _graphs.size() > 2);
+        TargetHint target_hint = _ctx.hints().target_hint();
+
+        // Create node context
+        NodeContext node_ctx(OperationType::ArithmeticAddition);
+        node_ctx.set_target(target_hint);
+
+        if(_graphs.size() == 1)
+        {
+            arm_compute::ITensor *in = _input->tensor();
+            node_ctx.add_input(in);
+        }
+
+        for(auto &o : _graph_outputs)
+        {
+            arm_compute::ITensor *in = o->tensor();
+            node_ctx.add_input(in);
+        }
+
+        arm_compute::ITensor *out = _output->tensor();
+        auto_init_if_empty(*out->info(), *_graph_outputs[0]->tensor()->info());
+        node_ctx.add_output(out);
+
+        _func = OperationRegistry::get().find_operation(OperationType::ArithmeticAddition, target_hint)->configure(node_ctx);
+
+        for(auto &o : _graph_outputs)
+        {
+            o->allocate();
+        }
+    }
+
+    // Inherited methods overriden:
+    void run() override
+    {
+        ARM_COMPUTE_ERROR_ON(_graphs.size() < 1 || _graphs.size() > 2);
+
+        for(auto &g : _graphs)
+        {
+            ARM_COMPUTE_ERROR_ON(g.get() == nullptr);
+            g->run();
+        }
+
+        _func->run();
+    }
+
+private:
+    GraphContext                                _ctx;
+    std::unique_ptr<ITensorObject>              _input;
+    ITensorObject                              *_output;
+    std::unique_ptr<arm_compute::IFunction>     _func;
+    std::vector<std::unique_ptr<Graph>>         _graphs;
+    std::vector<std::unique_ptr<ITensorObject>> _graph_outputs;
+};
+
+std::unique_ptr<arm_compute::IFunction> ResidualLayer::instantiate_node(GraphContext &ctx, ITensorObject *input, ITensorObject *output)
+{
+    ARM_COMPUTE_ERROR_ON_UNALLOCATED_TENSOR_OBJECT(input, output);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<Tensor *>(input) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<Tensor *>(output) == nullptr);
+
+    // Create residual function
+    auto func = arm_compute::support::cpp14::make_unique<ResidualFunction>(ctx, output);
+
+    if(_sub_graphs.size() == 1)
+    {
+        std::unique_ptr<ITensorObject> original_in;
+        original_in = arm_compute::support::cpp14::make_unique<SubTensor>(*dynamic_cast<Tensor *>(input),
+                                                                          input->tensor()->info()->tensor_shape(),
+                                                                          Coordinates());
+        func->set_input(std::move(original_in));
+    }
+
+    // Constuct all sub-graphs given the input/output
+    for(auto &sg : _sub_graphs)
+    {
+        ARM_COMPUTE_ERROR_ON(sg.get() == nullptr);
+
+        // IO buffers
+        std::unique_ptr<ITensorObject> in;
+        std::unique_ptr<ITensorObject> out;
+        std::unique_ptr<ITensorObject> func_in;
+
+        // Create input sub-tensor
+        if(!sg->has_input())
+        {
+            in = arm_compute::support::cpp14::make_unique<SubTensor>(*dynamic_cast<Tensor *>(input),
+                                                                     input->tensor()->info()->tensor_shape(),
+                                                                     Coordinates());
+        }
+
+        // Create output sub-tensor
+        if(!sg->has_output())
+        {
+            ITensorInfo *info = input->tensor()->info();
+            func_in           = arm_compute::support::cpp14::make_unique<Tensor>(TensorInfo(info->num_channels(), info->data_type(), info->fixed_point_position()));
+            func_in->set_target(ctx.hints().target_hint());
+            out = arm_compute::support::cpp14::make_unique<SubTensor>(func_in->tensor(),
+                                                                      TensorShape(),
+                                                                      Coordinates(0, 0, 0),
+                                                                      func_in->target(),
+                                                                      true);
+        }
+
+        // Construct sub_graph
+        auto g = sg->construct(ctx, std::move(in), std::move(out));
+
+        // Register graph to function
+        func->register_graph(std::move(g), std::move(func_in));
+    }
+
+    func->configure();
+
+    return std::move(func);
+}
diff --git a/src/graph/operations/CLSimpleOperations.cpp b/src/graph/operations/CLSimpleOperations.cpp
index 61315e7..fe56122 100644
--- a/src/graph/operations/CLSimpleOperations.cpp
+++ b/src/graph/operations/CLSimpleOperations.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -66,6 +66,34 @@
     return std::move(activation);
 }
 
+/* Arithmetic addition */
+REGISTER_SIMPLE_OPERATION(CLArithmeticAdditionOperation, OPENCL, OperationType::ArithmeticAddition)
+{
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2);
+    ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.input(1)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0)) == nullptr);
+
+    // Extract IO and info
+    auto *in1 = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0));
+    auto *in2 = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(1));
+    auto *out = dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0));
+
+    auto addition = arm_compute::support::cpp14::make_unique<arm_compute::CLArithmeticAddition>();
+    addition->configure(in1, in2, out, ConvertPolicy::SATURATE);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLArithmeticAddition"
+                               << " Data Type: " << in1->info()->data_type()
+                               << " Input 1 shape: " << in1->info()->tensor_shape()
+                               << " Input 2 shape: " << in2->info()->tensor_shape()
+                               << " Output shape: " << out->info()->tensor_shape()
+                               << std::endl);
+
+    return std::move(addition);
+}
+
 /* Batch Normalization Layer */
 REGISTER_SIMPLE_OPERATION(CLBatchNormalizationLayerOperation, OPENCL, OperationType::BatchNormalizationLayer)
 {
@@ -79,17 +107,18 @@
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0)) == nullptr);
 
     // Extract IO and info
-    auto      *in      = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0));
-    auto      *mean    = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(1));
-    auto      *var     = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(2));
-    auto      *beta    = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(3));
-    auto      *gamma   = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(4));
-    auto      *out     = dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0));
-    const auto epsilon = ctx.parameter<float>("epsilon");
+    auto      *in       = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(0));
+    auto      *mean     = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(1));
+    auto      *var      = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(2));
+    auto      *beta     = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(3));
+    auto      *gamma    = dynamic_cast<arm_compute::ICLTensor *>(ctx.input(4));
+    auto      *out      = dynamic_cast<arm_compute::ICLTensor *>(ctx.output(0));
+    const auto epsilon  = ctx.parameter<float>("epsilon");
+    const auto act_info = ctx.parameter<ActivationLayerInfo>("act_info");
 
     // Create and configure function
     auto batch_norm = arm_compute::support::cpp14::make_unique<arm_compute::CLBatchNormalizationLayer>();
-    batch_norm->configure(in, out, mean, var, beta, gamma, epsilon);
+    batch_norm->configure(in, out, mean, var, beta, gamma, epsilon, act_info);
 
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLBatchNormalizationLayer"
@@ -101,6 +130,9 @@
                                << " Beta shape: " << beta->info()->tensor_shape()
                                << " Gamma shape: " << gamma->info()->tensor_shape()
                                << " Epsilon: " << epsilon
+                               << " Activation function: " << act_info.activation()
+                               << " a: " << act_info.a()
+                               << " b: " << act_info.b()
                                << std::endl);
 
     return std::move(batch_norm);
@@ -460,4 +492,4 @@
                                << std::endl);
 
     return std::move(smx);
-}
\ No newline at end of file
+}
diff --git a/src/graph/operations/NESimpleOperations.cpp b/src/graph/operations/NESimpleOperations.cpp
index 49adbe9..4154b9a 100644
--- a/src/graph/operations/NESimpleOperations.cpp
+++ b/src/graph/operations/NESimpleOperations.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -66,6 +66,34 @@
     return std::move(activation);
 }
 
+/* Arithmetic addition */
+REGISTER_SIMPLE_OPERATION(NEArithmeticAdditionOperation, NEON, OperationType::ArithmeticAddition)
+{
+    ARM_COMPUTE_ERROR_ON(ctx.num_inputs() != 2);
+    ARM_COMPUTE_ERROR_ON(ctx.num_outputs() != 1);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.input(0)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.input(1)) == nullptr);
+    ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.output(0)) == nullptr);
+
+    // Extract IO and info
+    auto *in1 = dynamic_cast<arm_compute::ITensor *>(ctx.input(0));
+    auto *in2 = dynamic_cast<arm_compute::ITensor *>(ctx.input(1));
+    auto *out = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
+
+    auto addition = arm_compute::support::cpp14::make_unique<arm_compute::NEArithmeticAddition>();
+    addition->configure(in1, in2, out, ConvertPolicy::SATURATE);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEArithmeticAddition"
+                               << " Data Type: " << in1->info()->data_type()
+                               << " Input 1 shape: " << in1->info()->tensor_shape()
+                               << " Input 2 shape: " << in2->info()->tensor_shape()
+                               << " Output shape: " << out->info()->tensor_shape()
+                               << std::endl);
+
+    return std::move(addition);
+}
+
 /* Batch Normalization Layer */
 REGISTER_SIMPLE_OPERATION(NEBatchNormalizationLayerOperation, NEON, OperationType::BatchNormalizationLayer)
 {
@@ -79,17 +107,18 @@
     ARM_COMPUTE_ERROR_ON(dynamic_cast<arm_compute::ITensor *>(ctx.output(0)) == nullptr);
 
     // Extract IO and info
-    auto      *in      = dynamic_cast<arm_compute::ITensor *>(ctx.input(0));
-    auto      *mean    = dynamic_cast<arm_compute::ITensor *>(ctx.input(1));
-    auto      *var     = dynamic_cast<arm_compute::ITensor *>(ctx.input(2));
-    auto      *beta    = dynamic_cast<arm_compute::ITensor *>(ctx.input(3));
-    auto      *gamma   = dynamic_cast<arm_compute::ITensor *>(ctx.input(4));
-    auto      *out     = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
-    const auto epsilon = ctx.parameter<float>("epsilon");
+    auto      *in       = dynamic_cast<arm_compute::ITensor *>(ctx.input(0));
+    auto      *mean     = dynamic_cast<arm_compute::ITensor *>(ctx.input(1));
+    auto      *var      = dynamic_cast<arm_compute::ITensor *>(ctx.input(2));
+    auto      *beta     = dynamic_cast<arm_compute::ITensor *>(ctx.input(3));
+    auto      *gamma    = dynamic_cast<arm_compute::ITensor *>(ctx.input(4));
+    auto      *out      = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
+    const auto epsilon  = ctx.parameter<float>("epsilon");
+    const auto act_info = ctx.parameter<ActivationLayerInfo>("act_info");
 
     // Create and configure function
     auto batch_norm = arm_compute::support::cpp14::make_unique<arm_compute::NEBatchNormalizationLayer>();
-    batch_norm->configure(in, out, mean, var, beta, gamma, epsilon);
+    batch_norm->configure(in, out, mean, var, beta, gamma, epsilon, act_info);
 
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEBatchNormalizationLayer"
@@ -101,6 +130,9 @@
                                << " Beta shape: " << beta->info()->tensor_shape()
                                << " Gamma shape: " << gamma->info()->tensor_shape()
                                << " Epsilon: " << epsilon
+                               << " Activation function: " << act_info.activation()
+                               << " a: " << act_info.a()
+                               << " b: " << act_info.b()
                                << std::endl);
 
     return std::move(batch_norm);
@@ -149,12 +181,23 @@
     auto      *biases    = ctx.num_inputs() == 3 ? dynamic_cast<arm_compute::ITensor *>(ctx.input(2)) : nullptr;
     auto      *out       = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
     const auto conv_info = ctx.parameter<PadStrideInfo>("ConvolutionInfo");
+    const auto opt3x3    = ctx.parameter<bool>("Optimized3x3");
 
     // Create and configure function
     std::unique_ptr<arm_compute::IFunction> func;
-    auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
-    depwthwise_conv->configure(in, weights, biases, out, conv_info);
-    func = std::move(depwthwise_conv);
+    bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
+    if(run_3x3_opt)
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
+    else
+    {
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
+        depwthwise_conv->configure(in, weights, biases, out, conv_info);
+        func = std::move(depwthwise_conv);
+    }
 
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDepthwiseConvolutionLayer"