COMPMID-2273: Fuse Batch Normalization with Depthwise Convolution layer at graph level (only for CL)

Change-Id: I1d941c6e66722f39583bf68148c980bb28ff89a1
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1423
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

17 files changed, 578 insertions, 23 deletions

diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h
index be43b57e48..5c5b777ac9 100644
--- a/arm_compute/graph/INodeVisitor.h
+++ b/arm_compute/graph/INodeVisitor.h
@@ -96,6 +96,11 @@ public:
      * @param[in] n Node to visit.
      */
     virtual void visit(FusedConvolutionBatchNormalizationNode &n) = 0;
+    /** Visit FusedDepthwiseConvolutionBatchNormalizationNode.
+     *
+     * @param[in] n Node to visit.
+     */
+    virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) = 0;
     /** Visit InputNode.
      *
      * @param[in] n Node to visit.
@@ -214,6 +219,10 @@ public:
     {
         default_visit();
     }
+    virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override
+    {
+        default_visit();
+    }
     virtual void visit(InputNode &n) override
     {
         default_visit();
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index 4fb5b73333..9da0e6157c 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -101,6 +101,9 @@ inline ::std::ostream &operator<<(::std::ostream &os, const NodeType &node_type)
         case NodeType::FusedConvolutionBatchNormalizationLayer:
             os << "FusedConvolutionBatchNormalizationLayer";
             break;
+        case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer:
+            os << "FusedDepthwiseConvolutionBatchNormalizationLayer";
+            break;
         case NodeType::GenerateProposalsLayer:
             os << "GenerateProposalsLayer";
             break;
diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h
index 2f09abbbab..9f962425b3 100644
--- a/arm_compute/graph/Types.h
+++ b/arm_compute/graph/Types.h
@@ -141,6 +141,7 @@ enum class NodeType
     FlattenLayer,
     FullyConnectedLayer,
     FusedConvolutionBatchNormalizationLayer,
+    FusedDepthwiseConvolutionBatchNormalizationLayer,
     GenerateProposalsLayer,
     NormalizationLayer,
     NormalizePlanarYUVLayer,
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 5ac4fdaed9..ed5b35c0d1 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -30,6 +30,7 @@
 #include "arm_compute/graph/Types.h"
 #include "arm_compute/graph/Utils.h"
 #include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
+#include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h"
 #include "arm_compute/graph/backends/Utils.h"
 #include "arm_compute/graph/nodes/Nodes.h"
 
@@ -197,12 +198,6 @@ std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(Fu
     const ActivationLayerInfo fused_act  = node.fused_activation();
     const float               epsilon    = node.epsilon();
 
-    const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
-    if(is_quantized && biases != nullptr)
-    {
-        biases->info()->set_data_type(DataType::S32);
-    }
-
     // Create and configure function
     auto func = support::cpp14::make_unique<FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>();
     func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act);
@@ -210,7 +205,55 @@ std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(Fu
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
                                << node.name()
-                               << " Type: " << node.name()
+                               << " Type: " << node.type()
+                               << " Target: " << TargetInfo::TargetType
+                               << " Data Type: " << input->info()->data_type()
+                               << " Input shape: " << input->info()->tensor_shape()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << output->info()->tensor_shape()
+                               << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+                               << std::endl);
+    return std::move(func);
+}
+
+/** Create a backend fused depthwise convolution batch normalization layer function
+ *
+ * @tparam FusedLayerTypes             Fused layer types
+ * @tparam TargetInfo                  Target-specific information
+ *
+ * @param[in] node Node to create the backend function for
+ *
+ * @return Backend fused depthwise convolution batch normalization layer function
+ */
+template <typename FusedLayerTypes, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node)
+{
+    validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */);
+
+    // Extract IO and info
+    typename TargetInfo::TensorType *input   = get_backing_tensor<TargetInfo>(node.input(0));
+    typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
+    typename TargetInfo::TensorType *biases  = get_backing_tensor<TargetInfo>(node.input(2));
+    typename TargetInfo::TensorType *mean    = get_backing_tensor<TargetInfo>(node.input(3));
+    typename TargetInfo::TensorType *var     = get_backing_tensor<TargetInfo>(node.input(4));
+    typename TargetInfo::TensorType *beta    = get_backing_tensor<TargetInfo>(node.input(5));
+    typename TargetInfo::TensorType *gamma   = get_backing_tensor<TargetInfo>(node.input(6));
+
+    typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
+
+    const PadStrideInfo       conv_info        = node.convolution_info();
+    const unsigned int        depth_multiplier = node.depth_multiplier();
+    const ActivationLayerInfo fused_act        = node.fused_activation();
+    const float               epsilon          = node.epsilon();
+
+    // Create and configure function
+    auto func = support::cpp14::make_unique<FusedDepthwiseConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>>();
+    func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, depth_multiplier, fused_act);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << node.type()
                                << " Target: " << TargetInfo::TargetType
                                << " Data Type: " << input->info()->data_type()
                                << " Input shape: " << input->info()->tensor_shape()
diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
index 92af17b227..a6da76bb06 100644
--- a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
+++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h
@@ -54,7 +54,7 @@ public:
      *                        Data types supported: QASYMM8/F16/F32.
      * @param[in]  weights    Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
      * @param[in]  bias       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
-     *                        Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
+     *                        Data type supported: Should match @p input data type.
      * @param[out] output     Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
      *                        Data types supported: Same as @p input.
      * @param[in]  mean       Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
diff --git a/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
new file mode 100644
index 0000000000..6f70d3c3a0
--- /dev/null
+++ b/arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+
+#ifndef __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+#define __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/IFunction.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+namespace backends
+{
+/** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}DepthwiseConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedDepthwiseConvolutionBatchNormalizationFunction : public IFunction
+{
+public:
+    using TensorType         = typename TargetInfo::TensorType;
+    using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+    FusedDepthwiseConvolutionBatchNormalizationFunction()
+        : _depth_conv_layer(), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false)
+    {
+    }
+
+    /** Set the input and output tensors.
+     *
+     * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                              while every optional dimension from 4 and above represent a batch of inputs.
+     *                              Data types supported: F16/F32.
+     * @param[in]  weights          Weights tensor.  These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
+     * @param[in]  bias             Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [IFM].
+     *                              Data type supported: Should match @p input data type.
+     * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                              Data types supported: Same as @p input.
+     * @param[in]  mean             Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in]  var              Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
+     * @param[in]  beta             Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input
+     * @param[in]  gamma            Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input
+     * @param[in]  epsilon          Small value to avoid division with zero. Default value is 0.001f.
+     * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in]  fused_act        Activation layer information in case of a fused activation.
+     *
+     */
+    void configure(TensorType       *input,
+                   TensorType       *weights,
+                   TensorType       *bias,
+                   TensorType       *output,
+                   const TensorType *mean,
+                   const TensorType *var,
+                   const TensorType *beta,
+                   const TensorType *gamma,
+                   float epsilon, const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo const &fused_act)
+    {
+        // We don't run any validate, as we assume that the layers have been already validated
+        const bool        has_bias = (bias != nullptr);
+        const TensorType *bias_to_use;
+
+        // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one
+        // as batch normalization might end up with a bias != 0
+        if(has_bias)
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+            bias_to_use = bias;
+        }
+        else
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon, FuseBatchNormalizationType::DEPTHWISECONVOLUTION);
+            bias_to_use = &_fused_bias;
+        }
+
+        _depth_conv_layer.configure(input, weights, bias_to_use, output, conv_info, depth_multiplier, fused_act.enabled() ? fused_act : ActivationLayerInfo());
+
+        if(!has_bias)
+        {
+            _fused_bias.allocator()->allocate();
+        }
+    }
+
+    // Inherited methods overridden:
+    void run()
+    {
+        prepare();
+        _depth_conv_layer.run();
+    }
+
+    void prepare()
+    {
+        if(!_is_prepared)
+        {
+            _fused_batch_norm_layer.run();
+            _is_prepared = true;
+        }
+    }
+
+private:
+    typename FusedLayerTypes::DepthwiseConvolutionLayer _depth_conv_layer;
+    typename FusedLayerTypes::FuseBatchNormalization    _fused_batch_norm_layer;
+    TensorConcreteType                                  _fused_bias;
+    bool                                                _is_prepared;
+};
+} // namespace backends
+} // namespace graph
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_FUNCTION_H__ */
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
index 9b0f5b7ade..c124c982cd 100644
--- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
@@ -41,14 +41,13 @@ public:
      * @param[in] num_groups       (Optional) Number of groups (Defaults to 1)
      * @param[in] method           (Optional) Convolution method to use
      * @param[in] fast_math_hint   (Optional) Fast math hint
-     * @param[in] out_quant_info   (Optional) Output quantization info
      * @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
      */
     FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info,
                                            unsigned int      num_groups     = 1,
                                            ConvolutionMethod method         = ConvolutionMethod::Default,
                                            FastMathHint      fast_math_hint = FastMathHint::Disabled,
-                                           QuantizationInfo out_quant_info = QuantizationInfo(), ActivationLayerInfo fused_activation = ActivationLayerInfo());
+                                           ActivationLayerInfo fused_activation = ActivationLayerInfo());
 
     /** Epsilon parameter accessor
      *
@@ -135,7 +134,6 @@ private:
     unsigned int        _num_groups;
     ConvolutionMethod   _method;
     FastMathHint        _fast_math_hint;
-    QuantizationInfo    _out_quant_info;
     ActivationLayerInfo _fused_activation;
 };
 
diff --git a/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
new file mode 100644
index 0000000000..a2241ef64d
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2019 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.
+ */
+#ifndef __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+#define __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__
+
+#include "arm_compute/graph/INode.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Fused Depthwise Convolution Batch Normalization node */
+class FusedDepthwiseConvolutionBatchNormalizationNode final : public INode
+{
+public:
+    /** Constructor
+     *
+     * @param[in] epsilon          Epsilon parameter.
+     * @param[in] info             Convolution layer attributes.
+     * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+     * @param[in] method           (Optional) Convolution method to use
+     * @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified
+     */
+    FusedDepthwiseConvolutionBatchNormalizationNode(float                      epsilon,
+                                                    PadStrideInfo              info,
+                                                    unsigned int               depth_multiplier,
+                                                    DepthwiseConvolutionMethod method,
+                                                    ActivationLayerInfo        fused_activation = ActivationLayerInfo());
+
+    /** Sets the depthwise convolution layer method to use
+     *
+     * @param[in] method Method to use for depthwise convolution
+     */
+    void set_depthwise_convolution_method(DepthwiseConvolutionMethod method);
+
+    /** Depthwise convolution layer method accessor
+     *
+     * @note This is an indication on which depthwise convolution layer implementation to use,
+     *       if it fails to be created the library's heuristic approach will be used
+     *
+     * @return Depthwise convolution layer method to be used by the node
+     */
+    DepthwiseConvolutionMethod depthwise_convolution_method() const;
+
+    /** Epsilon parameter accessor
+     *
+     * @return Epsilon parameter
+     */
+    float epsilon() const;
+
+    /** Returns fused activation
+     *
+     * @return Fused activation
+     */
+    ActivationLayerInfo fused_activation() const;
+
+    /** Sets fused activation
+     *
+     * @param[in] fused_activation Fused activation to set
+     */
+    void set_fused_activation(ActivationLayerInfo fused_activation);
+
+    /** Computes convolution output descriptor
+     *
+     * @param[in] input_descriptor   Input descriptor
+     * @param[in] weights_descriptor Weights descriptor
+     * @param[in] info               Convolution operation attributes
+     * @param[in] depth_multiplier   Depth multiplier
+     *
+     * @return Output descriptor
+     */
+    static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+                                                      const TensorDescriptor &weights_descriptor,
+                                                      const PadStrideInfo    &info,
+                                                      int                     depth_multiplier);
+
+    /** Sets the convolution layer method to use
+     *
+     * @param[in] method Method to use for convolution
+     */
+    void set_convolution_method(ConvolutionMethod method);
+
+    /** Depth multiplier accessor
+     *
+     * @return Depth multiplier
+     */
+    unsigned int depth_multiplier() const;
+
+    /** Convolution metadata accessor
+     *
+     * @return Convolution information
+     */
+    PadStrideInfo convolution_info() const;
+
+    // Inherited overridden methods:
+    NodeType         type() const override;
+    bool             forward_descriptors() override;
+    TensorDescriptor configure_output(size_t idx) const override;
+    void accept(INodeVisitor &v) override;
+
+public:
+    static constexpr NodeType node_type = NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer;
+
+private:
+    float _epsilon;
+
+    PadStrideInfo              _info;
+    unsigned int               _depth_multiplier;
+    DepthwiseConvolutionMethod _method;
+    ActivationLayerInfo        _fused_activation;
+};
+
+} // namespace graph
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_GRAPH_FUSED_DEPTHWISE_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__ */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index c891bc2ca2..52e2f88528 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -39,6 +39,7 @@
 #include "arm_compute/graph/nodes/FlattenLayerNode.h"
 #include "arm_compute/graph/nodes/FullyConnectedLayerNode.h"
 #include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h"
+#include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
 #include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
 #include "arm_compute/graph/nodes/InputNode.h"
 #include "arm_compute/graph/nodes/NormalizationLayerNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 0f3450b08f..2c89679902 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -45,6 +45,7 @@ class EltwiseLayerNode;
 class FlattenLayerNode;
 class FullyConnectedLayerNode;
 class FusedConvolutionBatchNormalizationNode;
+class FusedDepthwiseConvolutionBatchNormalizationNode;
 class GenerateProposalsLayerNode;
 class InputNode;
 class NormalizationLayerNode;
diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h
index 9d2ea46fde..c28a17b21a 100644
--- a/arm_compute/graph/printers/DotGraphPrinter.h
+++ b/arm_compute/graph/printers/DotGraphPrinter.h
@@ -57,6 +57,7 @@ public:
     void visit(DepthwiseConvolutionLayerNode &n) override;
     void visit(EltwiseLayerNode &n) override;
     void visit(FusedConvolutionBatchNormalizationNode &n) override;
+    void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     void visit(NormalizationLayerNode &n) override;
     void visit(PoolingLayerNode &n) override;
     void default_visit() override;
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index c14100ab42..b9f22f6199 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -74,8 +74,9 @@ struct CLEltwiseFunctions
 /** Function and tensor types to be used inside a CL fused convolution/batch normalization layer */
 struct CLFusedLayerTypes
 {
-    using ConvolutionLayer       = CLConvolutionLayer;
-    using FuseBatchNormalization = CLFuseBatchNormalization;
+    using ConvolutionLayer          = CLConvolutionLayer;
+    using DepthwiseConvolutionLayer = CLDepthwiseConvolutionLayer;
+    using FuseBatchNormalization    = CLFuseBatchNormalization;
 };
 
 // TODO (isagot01): Remove once we support heterogeneous scheduling at function level
@@ -203,6 +204,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx);
         case NodeType::FusedConvolutionBatchNormalizationLayer:
             return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node));
+        case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer:
+            return detail::create_fused_depthwise_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return detail::create_generate_proposals_layer<CLGenerateProposalsLayer, CLTargetInfo>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node), ctx);
         case NodeType::NormalizationLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index d4892f53a6..ad96240a4b 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -80,8 +80,9 @@ struct NEEltwiseFunctions
 /** Function and tensor types to be used inside a NEON fused convolution/batch normalization layer */
 struct NEFusedLayerTypes
 {
-    using ConvolutionLayer       = NEConvolutionLayer;
-    using FuseBatchNormalization = NEFuseBatchNormalization;
+    using ConvolutionLayer          = NEConvolutionLayer;
+    using DepthwiseConvolutionLayer = NEDepthwiseConvolutionLayer;
+    using FuseBatchNormalization    = NEFuseBatchNormalization;
 };
 
 namespace detail
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 427d7b5095..83177a8431 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -64,7 +64,6 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
         // Extract conv inputs
         const auto   conv_input_id   = conv_node->input_edge(0)->producer_id();
         const auto   conv_weights_id = conv_node->input_edge(1)->producer_id();
-        const auto   out_quant_info  = conv_node->output(0)->desc().quant_info;
         const auto   conv_info       = conv_node->convolution_info();
         const auto   conv_method     = conv_node->convolution_method();
         const auto   num_groups      = conv_node->num_groups();
@@ -79,7 +78,7 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
         const auto epsilon     = bn_node->epsilon();
 
         // Create the fused node
-        const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint, out_quant_info, act_info);
+        const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint, act_info);
 
         if(conv_node->input_edge(2) != nullptr)
         {
@@ -125,6 +124,83 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge
     }
 }
 
+void fuse_depthwise_convolution_with_batch_normalization(Graph &g, const Edge *output_edge)
+{
+    ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
+
+    auto *depth_conv_node = arm_compute::utils::cast::polymorphic_downcast<DepthwiseConvolutionLayerNode *>(output_edge->producer());
+    auto *bn_node         = arm_compute::utils::cast::polymorphic_downcast<BatchNormalizationLayerNode *>(output_edge->consumer());
+
+    ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing depthwise convolution node with ID : " << output_edge->producer_id()
+                                  << " with BatchNormalization Layer node with ID : " << output_edge->consumer_id() << std::endl);
+
+    // Prevent fusion if fused node has an output accessor
+    if(depth_conv_node->output(0)->accessor() == nullptr)
+    {
+        const Target assigned_target = depth_conv_node->assigned_target();
+
+        // Extract conv inputs
+        const auto depth_conv_input_id = depth_conv_node->input_edge(0)->producer_id();
+        const auto conv_weights_id     = depth_conv_node->input_edge(1)->producer_id();
+        const auto conv_info           = depth_conv_node->convolution_info();
+        const auto depth_conv_method   = depth_conv_node->depthwise_convolution_method();
+        const auto depth_multiplier    = depth_conv_node->depth_multiplier();
+        const auto act_info            = bn_node->fused_activation();
+
+        // Extract bn inputs
+        const auto bn_mean_id  = bn_node->input_edge(1)->producer_id();
+        const auto bn_var_id   = bn_node->input_edge(2)->producer_id();
+        const auto bn_beta_id  = bn_node->input_edge(3)->producer_id();
+        const auto bn_gamma_id = bn_node->input_edge(4)->producer_id();
+        const auto epsilon     = bn_node->epsilon();
+
+        // Create the fused node
+        const NodeID fused_id = g.add_node<FusedDepthwiseConvolutionBatchNormalizationNode>(epsilon, conv_info, depth_multiplier, depth_conv_method, act_info);
+
+        if(depth_conv_node->input_edge(2) != nullptr)
+        {
+            const auto conv_bias_id = depth_conv_node->input_edge(2)->producer_id();
+            g.add_connection(conv_bias_id, 0, fused_id, 2);
+        }
+
+        // Add connections from the conv/batch_norm inputs to the fused node
+        g.add_connection(depth_conv_input_id, 0, fused_id, 0);
+        g.add_connection(conv_weights_id, 0, fused_id, 1);
+        g.add_connection(bn_mean_id, 0, fused_id, 3);
+        g.add_connection(bn_var_id, 0, fused_id, 4);
+        g.add_connection(bn_beta_id, 0, fused_id, 5);
+        g.add_connection(bn_gamma_id, 0, fused_id, 6);
+
+        auto                     fused_node       = g.node(fused_id);
+        std::vector<NodeIdxPair> bn_driving_nodes = get_driving_nodes(*bn_node);
+
+        // Extract batch normalization node accessor if any
+        auto bn_node_accessor = bn_node->output(0)->extract_accessor();
+        auto bn_node_name     = bn_node->name();
+
+        // Remove batch normalization node
+        g.remove_node(bn_node->id());
+
+        // Get driving nodes of batch normalization node
+        for(auto &driving_node : bn_driving_nodes)
+        {
+            g.add_connection(fused_id, 0, driving_node.node_id, driving_node.index);
+            configure_tensor(fused_node->output(0));
+        }
+        // Update fused node outputs
+        fused_node->output(0)->set_accessor(std::move(bn_node_accessor));
+        fused_node->set_assigned_target(assigned_target);
+        fused_node->set_common_node_parameters(NodeParams{ depth_conv_node->name() + "+" + bn_node_name, assigned_target });
+
+        // Remove convolution node
+        g.remove_node(depth_conv_node->id());
+    }
+    else
+    {
+        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of depthwise convolution with batch normalization due to the presence of an output accessor\n");
+    }
+}
+
 template <typename N>
 void fuse_node_with_activation(Graph &g, const Edge *output_edge, const std::set<Activation> &supported_fused_activations)
 {
@@ -224,6 +300,8 @@ void NodeFusionMutator::mutate(Graph &g)
         return (output_qasymm8 && same_qinfo) || !output_qasymm8;
     };
 
+    Target target = g.nodes()[0].get()->output(0)->desc().target;
+
     // Fusion mutations
     detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations);
     detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations);
@@ -231,6 +309,11 @@ void NodeFusionMutator::mutate(Graph &g)
 
     // TODO (COMPMID-2055): re-enable once we fuse bias and activations to convolution
     // detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
+    if(target == Target::CL)
+    {
+        //Depthwise Convolution and Batch Normalization Fusion active only for CL
+        detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
+    }
 }
 } // namespace graph
 } // namespace arm_compute
diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
index 6496a71251..0a0c0c50e8 100644
--- a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
+++ b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp
@@ -36,8 +36,8 @@ FusedConvolutionBatchNormalizationNode::FusedConvolutionBatchNormalizationNode(f
                                                                                unsigned int      num_groups,
                                                                                ConvolutionMethod method,
                                                                                FastMathHint      fast_math_hint,
-                                                                               QuantizationInfo out_quant_info, ActivationLayerInfo fused_activation)
-    : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(std::move(out_quant_info)), _fused_activation(fused_activation)
+                                                                               ActivationLayerInfo fused_activation)
+    : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _fused_activation(fused_activation)
 {
     _input_edges.resize(7, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -132,10 +132,6 @@ TensorDescriptor FusedConvolutionBatchNormalizationNode::configure_output(size_t
     ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
 
     TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info);
-    if(!_out_quant_info.empty())
-    {
-        output_info.quant_info = _out_quant_info;
-    }
 
     return output_info;
 }
diff --git a/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp
new file mode 100644
index 0000000000..a04d75407a
--- /dev/null
+++ b/src/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.cpp
@@ -0,0 +1,140 @@
+/*
+ * Copyright (c) 2019 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/FusedDepthwiseConvolutionBatchNormalizationNode.h"
+
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/INodeVisitor.h"
+#include "arm_compute/graph/Utils.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+FusedDepthwiseConvolutionBatchNormalizationNode::FusedDepthwiseConvolutionBatchNormalizationNode(float                      epsilon,
+                                                                                                 PadStrideInfo              info,
+                                                                                                 unsigned int               depth_multiplier,
+                                                                                                 DepthwiseConvolutionMethod method,
+                                                                                                 ActivationLayerInfo        fused_activation)
+    : _epsilon(epsilon), _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _fused_activation(fused_activation)
+{
+    _input_edges.resize(7, EmptyEdgeID);
+    _outputs.resize(1, NullTensorID);
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::set_depthwise_convolution_method(DepthwiseConvolutionMethod method)
+{
+    _method = method;
+}
+
+DepthwiseConvolutionMethod FusedDepthwiseConvolutionBatchNormalizationNode::depthwise_convolution_method() const
+{
+    return _method;
+}
+
+float FusedDepthwiseConvolutionBatchNormalizationNode::epsilon() const
+{
+    return _epsilon;
+}
+
+PadStrideInfo FusedDepthwiseConvolutionBatchNormalizationNode::convolution_info() const
+{
+    return _info;
+}
+
+unsigned int FusedDepthwiseConvolutionBatchNormalizationNode::depth_multiplier() const
+{
+    return _depth_multiplier;
+}
+
+ActivationLayerInfo FusedDepthwiseConvolutionBatchNormalizationNode::fused_activation() const
+{
+    return _fused_activation;
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::set_fused_activation(ActivationLayerInfo fused_activation)
+{
+    _fused_activation = fused_activation;
+}
+
+TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
+                                                                                            const TensorDescriptor &weights_descriptor,
+                                                                                            const PadStrideInfo    &info,
+                                                                                            int                     depth_multiplier)
+{
+    unsigned int output_width  = 0;
+    unsigned int output_height = 0;
+
+    const unsigned int input_width    = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH);
+    const unsigned int input_height   = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT);
+    const unsigned int input_channels = get_dimension_size(input_descriptor, DataLayoutDimension::CHANNEL);
+    const unsigned int kernel_width   = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH);
+    const unsigned int kernel_height  = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT);
+
+    std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info);
+
+    TensorDescriptor output_descriptor = input_descriptor;
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::WIDTH), output_width);
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::HEIGHT), output_height);
+    output_descriptor.shape.set(get_dimension_idx(output_descriptor.layout, DataLayoutDimension::CHANNEL), input_channels * depth_multiplier);
+
+    return output_descriptor;
+}
+
+bool FusedDepthwiseConvolutionBatchNormalizationNode::forward_descriptors()
+{
+    if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (output_id(0) != NullTensorID))
+    {
+        Tensor *dst = output(0);
+        ARM_COMPUTE_ERROR_ON(dst == nullptr);
+        dst->desc() = configure_output(0);
+        return true;
+    }
+    return false;
+}
+
+TensorDescriptor FusedDepthwiseConvolutionBatchNormalizationNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_UNUSED(idx);
+    const Tensor *src     = input(0);
+    const Tensor *weights = input(1);
+
+    ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
+
+    TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+
+    return output_info;
+}
+
+NodeType FusedDepthwiseConvolutionBatchNormalizationNode::type() const
+{
+    return FusedDepthwiseConvolutionBatchNormalizationNode::node_type;
+}
+
+void FusedDepthwiseConvolutionBatchNormalizationNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
diff --git a/src/graph/printers/DotGraphPrinter.cpp b/src/graph/printers/DotGraphPrinter.cpp
index c939de1b64..46f6ee828e 100644
--- a/src/graph/printers/DotGraphPrinter.cpp
+++ b/src/graph/printers/DotGraphPrinter.cpp
@@ -85,6 +85,14 @@ void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n)
     _info = ss.str();
 }
 
+void DotGraphVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n)
+{
+    ARM_COMPUTE_UNUSED(n);
+    std::stringstream ss;
+    ss << "FusedDepthwiseConvolutionBatchNormalizationNode";
+    _info = ss.str();
+}
+
 void DotGraphVisitor::visit(NormalizationLayerNode &n)
 {
     std::stringstream ss;