Graph Fusion With Post Ops Fix

- Fusing ConvolutionBatchNormalization Nodes with post ops (activation
or element wise ops)

Resolves: COMPMID-4982
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: I5b2d32cad00f710fd744cb5aa2d59fd7e5c97e0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6766
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>

17 files changed, 734 insertions, 66 deletions

diff --git a/arm_compute/graph/DataLayerVisitor.h b/arm_compute/graph/DataLayerVisitor.h
index 670b9f02b6..ac7f1c84ee 100644
--- a/arm_compute/graph/DataLayerVisitor.h
+++ b/arm_compute/graph/DataLayerVisitor.h
@@ -48,6 +48,7 @@ public:
     void visit(ConvolutionLayerNode &n) override;
     void visit(DepthwiseConvolutionLayerNode &n) override;
     void visit(FusedConvolutionBatchNormalizationNode &n) override;
+    void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
     void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     void visit(OutputNode &n) override;
 
diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h
index 4cb601fb02..97e95336ef 100644
--- a/arm_compute/graph/INodeVisitor.h
+++ b/arm_compute/graph/INodeVisitor.h
@@ -106,6 +106,11 @@ public:
      * @param[in] n Node to visit.
      */
     virtual void visit(FusedConvolutionBatchNormalizationNode &n) = 0;
+    /** Visit FusedConvolutionBatchNormalizationWithPostOpsNode.
+     *
+     * @param[in] n Node to visit.
+     */
+    virtual void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) = 0;
     /** Visit FusedConvolutionWithPostOpNode.
      *
      * @param[in] n Node to visit.
@@ -210,6 +215,7 @@ public:
     virtual void visit(FlattenLayerNode &n) override;
     virtual void visit(FullyConnectedLayerNode &n) override;
     virtual void visit(FusedConvolutionBatchNormalizationNode &n) override;
+    virtual void visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
     virtual void visit(FusedConvolutionWithPostOpNode &n) override;
     virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     virtual void visit(InputNode &n) override;
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index a8a20c9de8..8f97bbf845 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -116,6 +116,9 @@ inline ::std::ostream &operator<<(::std::ostream &os, const NodeType &node_type)
         case NodeType::FusedConvolutionBatchNormalizationLayer:
             os << "FusedConvolutionBatchNormalizationLayer";
             break;
+        case NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer:
+            os << "FusedConvolutionBatchNormalizationLayerWithPostOpsLayer";
+            break;
         case NodeType::FusedConvolutionWithPostOp:
             os << "FusedConvolutionWithPostOp";
             break;
diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h
index e802e9dc77..ff33d5037b 100644
--- a/arm_compute/graph/Types.h
+++ b/arm_compute/graph/Types.h
@@ -216,6 +216,7 @@ enum class NodeType
     FullyConnectedLayer,
     FusedConvolutionBatchNormalizationLayer,
     FusedConvolutionWithPostOp,
+    FusedConvolutionBatchNormalizationLayerWithPostOpsLayer,
     FusedDepthwiseConvolutionBatchNormalizationLayer,
     GenerateProposalsLayer,
     L2NormalizeLayer,
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 1e420a803f..a7e52d4d6d 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -32,6 +32,7 @@
 #include "arm_compute/graph/Types.h"
 #include "arm_compute/graph/Utils.h"
 #include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
+#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h"
 #include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h"
 #include "arm_compute/graph/backends/Utils.h"
 #include "arm_compute/graph/nodes/Nodes.h"
@@ -540,7 +541,7 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     return std::move(func);
 }
 
-/** Create a backend convolution layer function with post opreator
+/** Create a backend convolution layer function with post operator
  *
  * @tparam ConvolutionLayerFunctions Backend convolution functions
  * @tparam TargetInfo                Target-specific information
@@ -629,6 +630,91 @@ std::unique_ptr<IFunction> create_fused_convolution_with_post_op(FusedConvolutio
                                << " Output shape: " << output->info()->tensor_shape()
                                << qss.str()
                                << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+                               << " Post ops" << post_ops;
+                               << std::endl);
+    return std::move(func);
+}
+
+/** Create a backend convolution batch normalization layer function with post operator
+ *
+ * @tparam FusedLayerTypes           Backend convolution functions
+ * @tparam TargetInfo                Target-specific information
+ *
+ * @param[in] node Node to create the backend function for
+ * @param[in] ctx  Graph context
+ *
+ * @return Backend fused convolution with batch normalization layer function
+ */
+template <typename FusedLayerTypes, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_with_post_op(FusedConvolutionBatchNormalizationWithPostOpsNode &node, GraphContext &ctx)
+{
+    validate_node<TargetInfo>(node, 8 /* 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  num_groups = node.num_groups();
+    const bool          fast_math  = node.fast_math_hint() == FastMathHint::Enabled;
+    const float         epsilon    = node.epsilon();
+
+    experimental::PostOpList<typename TargetInfo::TensorType *> post_ops;
+
+    auto &post_op_info_list = node.post_op_info_list();
+    for(const auto &post_op_info : post_op_info_list)
+    {
+        switch(post_op_info->type())
+        {
+            case PostOpType::Activation:
+            {
+                const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get());
+                post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act);
+                break;
+            }
+            case PostOpType::Eltwise_Add:
+            {
+                typename TargetInfo::TensorType *add_input    = get_backing_tensor<TargetInfo>(node.input(3));
+                const auto                       eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get());
+                post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy);
+                break;
+            }
+            default:
+            {
+                ARM_COMPUTE_ERROR("Unsupported PostOpType");
+            }
+        }
+    }
+
+    // Create and configure function (we assume that functions have been validated before creation)
+    std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
+    std::unique_ptr<IFunction>      func;
+    std::string                     func_name;
+
+    using FType = FusedConvolutionBatchNormalizationWithPostOpsFunction<TargetInfo, FusedLayerTypes>;
+
+    // Create and configure function
+    std::tie(func, func_name) = create_named_memory_managed_function<FType>(
+                                    std::string("FusedConvolutionBatchNormalizationLayerWithPostOpsLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, post_ops);
+
+    // 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()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << output->info()->tensor_shape()
+                               << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+                               << " Post Ops:" << post_ops;
                                << std::endl);
     return std::move(func);
 }
diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h
new file mode 100644
index 0000000000..10f2e5c25e
--- /dev/null
+++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h
@@ -0,0 +1,136 @@
+/*
+ * Copyright (c) 2021 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_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H
+#define ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H
+
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/experimental/IPostOp.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}ConvolutionLayer with the modified weights */
+template <typename TargetInfo, typename FusedLayerTypes>
+class FusedConvolutionBatchNormalizationWithPostOpsFunction : public IFunction
+{
+public:
+    using TensorType         = typename TargetInfo::TensorType;
+    using TensorConcreteType = typename TargetInfo::TensorConcreteType;
+
+    FusedConvolutionBatchNormalizationWithPostOpsFunction(std::shared_ptr<IMemoryManager> memory_manager = nullptr)
+        : _conv_layer(memory_manager), _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: 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.
+     * @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]  num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
+     * @param[in]  fast_math  Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
+     *                        available which may introduce a drop of accuracy as well. Default is false
+     * @param[in]  post_ops   A sequence of post operations that are performed after the main operation.
+     *
+     */
+    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 num_groups, bool fast_math,
+                   const arm_compute::experimental::PostOpList<TensorType *> &post_ops = experimental::PostOpList<TensorType *> {})
+    {
+        // 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);
+            bias_to_use = bias;
+        }
+        else
+        {
+            _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon);
+            bias_to_use = &_fused_bias;
+        }
+
+        ActivationLayerInfo fused_act = ActivationLayerInfo(); // Passing an empty ActivationLayerInfo.
+        _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups, post_ops);
+
+        if(!has_bias)
+        {
+            _fused_bias.allocator()->allocate();
+        }
+    }
+
+    // Inherited methods overridden:
+    void run()
+    {
+        prepare();
+        _conv_layer.run();
+    }
+
+    void prepare()
+    {
+        if(!_is_prepared)
+        {
+            _fused_batch_norm_layer.run();
+            _is_prepared = true;
+        }
+    }
+
+private:
+    typename FusedLayerTypes::ConvolutionLayer       _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_CONVOLUTION_BATCH_NORMAZLIZATION_WITH_POST_OPS_FUNCTION_H */
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
index b3661c300f..b0051b1385 100644
--- a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 Arm Limited.
+ * Copyright (c) 2019, 2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -100,7 +100,7 @@ public:
      */
     ConvolutionMethod convolution_method() const;
 
-    /** Sets the fast math fast hint
+    /** Sets the fast math hint
      *
      * @param[in] hint Hint to use for convolution
      */
diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h
new file mode 100644
index 0000000000..a42e06d889
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h
@@ -0,0 +1,127 @@
+/*
+ * Copyright (c) 2021 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_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H
+#define ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_WITH_POST_OPS_NODE_H
+
+#include "arm_compute/graph/INode.h"
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Batch Normalization node */
+class FusedConvolutionBatchNormalizationWithPostOpsNode final : public INode
+{
+public:
+    /** Constructor
+     *
+     * @param[in] epsilon        Epsilon parameter.
+     * @param[in] info           Convolution layer attributes.
+     * @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
+     */
+    FusedConvolutionBatchNormalizationWithPostOpsNode(float epsilon, PadStrideInfo info,
+                                                      unsigned int      num_groups     = 1,
+                                                      ConvolutionMethod method         = ConvolutionMethod::Default,
+                                                      FastMathHint      fast_math_hint = FastMathHint::Disabled);
+
+    /** Epsilon parameter accessor
+     *
+     * @return Epsilon parameter
+     */
+    float epsilon() const;
+
+    /** Computes convolution output descriptor
+     *
+     * @param[in] input_descriptor   Input descriptor
+     * @param[in] weights_descriptor Weights descriptor
+     * @param[in] info               Convolution operation attributes
+     *
+     * @return Output descriptor
+     */
+    static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+                                                      const TensorDescriptor &weights_descriptor,
+                                                      const PadStrideInfo    &info);
+
+    /** Sets the convolution layer method to use
+     *
+     * @param[in] method Method to use for convolution
+     */
+    void set_convolution_method(ConvolutionMethod method);
+
+    /** Number of groups in convolution accessor
+     *
+     * @return Number of groups in convolution
+     */
+    unsigned int num_groups() const;
+
+    /** Convolution layer method accessor
+     *
+     * @note This is an indication on which convolution layer implementation to use,
+     *       if it fails to be created the library's heuristic approach will be used
+     *
+     * @return Convolution layer method to be used by the node
+     */
+    ConvolutionMethod convolution_method() const;
+
+    /** Sets the fast math hint
+     *
+     * @param[in] hint Hint to use for convolution
+     */
+    void set_fast_math_hint(FastMathHint hint);
+
+    /** Fast math hint accessor
+     *
+     * @return Fast math hint to be used by the node
+     */
+    FastMathHint fast_math_hint() 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::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer;
+
+private:
+    float _epsilon;
+
+    PadStrideInfo     _info;
+    unsigned int      _num_groups;
+    ConvolutionMethod _method;
+    FastMathHint      _fast_math_hint;
+};
+
+} // namespace graph
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_GRAPH_BATCH_NORMALIZATION_LAYER_NODE_H */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index fb0eb155fd..3887eaeac6 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -43,6 +43,7 @@
 #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/FusedConvolutionBatchNormalizationWithPostOpsNode.h"
 #include "arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h"
 #include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
 #include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 6393b1d5d4..f1576d6336 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -51,6 +51,7 @@ class FullyConnectedLayerNode;
 class FusedConvolutionBatchNormalizationNode;
 class FusedConvolutionWithPostOpNode;
 class FusedDepthwiseConvolutionBatchNormalizationNode;
+class FusedConvolutionBatchNormalizationWithPostOpsNode;
 class GenerateProposalsLayerNode;
 class InputNode;
 class L2NormalizeLayerNode;
diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h
index 42a20678fe..63b89272f4 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(FusedConvolutionBatchNormalizationWithPostOpsNode &n) override;
     void visit(FusedConvolutionWithPostOpNode &n) override;
     void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     void visit(NormalizationLayerNode &n) override;
diff --git a/src/graph/DataLayerVisitor.cpp b/src/graph/DataLayerVisitor.cpp
index 3d5d9578ca..85d24b4654 100644
--- a/src/graph/DataLayerVisitor.cpp
+++ b/src/graph/DataLayerVisitor.cpp
@@ -131,6 +131,14 @@ void DataLayerVisitor::visit(FusedConvolutionBatchNormalizationNode &n)
     add_convolution_layer_method<FusedConvolutionBatchNormalizationNode>(_layer_data, n);
 }
 
+void DataLayerVisitor::visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n)
+{
+    _layer_data.clear();
+    add_generic_layer_data<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n);
+    add_convolution_layer_data<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n);
+    add_convolution_layer_method<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n);
+}
+
 void DataLayerVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n)
 {
     _layer_data.clear();
diff --git a/src/graph/INodeVisitor.cpp b/src/graph/INodeVisitor.cpp
index ceaaeae7f2..f067d618bd 100644
--- a/src/graph/INodeVisitor.cpp
+++ b/src/graph/INodeVisitor.cpp
@@ -85,6 +85,10 @@ void DefaultNodeVisitor::visit(FusedConvolutionBatchNormalizationNode &n)
 {
     default_visit(n);
 }
+void DefaultNodeVisitor::visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n)
+{
+    default_visit(n);
+}
 void DefaultNodeVisitor::visit(FusedConvolutionWithPostOpNode &n)
 {
     default_visit(n);
@@ -147,4 +151,4 @@ void DefaultNodeVisitor::visit(StackLayerNode &n)
 }
 #endif /* DOXYGEN_SKIP_THIS */
 } // namespace graph
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index 838977e75c..c67f6a538b 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -318,6 +318,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_stack_layer<CLStackLayer, CLTargetInfo>(*polymorphic_downcast<StackLayerNode *>(node));
         case NodeType::StridedSliceLayer:
             return detail::create_strided_slice_layer<CLStridedSlice, CLTargetInfo>(*polymorphic_downcast<StridedSliceLayerNode *>(node));
+        case NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer:
+            return detail::create_fused_convolution_batch_normalization_with_post_op<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationWithPostOpsNode *>(node), ctx);
         default:
             return nullptr;
     }
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 78f275f3bb..5284fce806 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -28,6 +28,7 @@
 #include "arm_compute/graph/Utils.h"
 #include "arm_compute/graph/backends/BackendRegistry.h"
 #include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h"
+#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.h"
 #include "arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h"
 #include "arm_compute/graph/nodes/Nodes.h"
 
@@ -420,88 +421,91 @@ void fuse_convolution_with_post_op(Graph &g, INode *fused_node, std::list<INode
     }
 }
 
-std::list<INode *> get_post_op_list(Graph &g, int &eltwise_operand_id, int &prev_op_dst_pos, int conv_node_id, const std::set<Activation> &supported_fused_activations)
+std::list<INode *> get_post_op_list(Graph &g, int &eltwise_operand_id, int &prev_op_dst_pos, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations)
 {
     std::list<INode *> post_op_node_list    = {};
     NodeID             prev_op_dst_id       = conv_node_id;
     NodeType           post_op_type_list[3] = { NodeType::Dummy, NodeType::Dummy, NodeType::Dummy };
     int                post_op_idx          = 0;
-    for(unsigned int i = conv_node_id + 1; i < g.nodes().size(); ++i)
+
+    // Get list of the connected nodes
+    auto current_node = g.node(conv_node_id);
+
+    while(post_op_node_list.size() < 3)
     {
-        auto post_op_node    = g.node(i);
+        // This convolution node must have only one output edge, otherwise this function would not have been called
+
+        auto current_output_edge_id = current_node->output_edges().begin();
+        auto current_output_edge    = g.edge(*current_output_edge_id);
+        auto post_op_node           = current_output_edge->consumer();
+
         bool fusable_post_op = false;
         if(post_op_node != nullptr && post_op_node->output_edges().size() > 0)
         {
-            const auto post_op_output_edge_id = *post_op_node->output_edges().begin();
-            const auto post_op_output_edge    = g.edge(post_op_output_edge_id);
-
-            if(post_op_output_edge != nullptr)
+            switch(post_op_node->type())
             {
-                switch(post_op_output_edge->producer()->type())
+                case EltwiseLayerNode::node_type:
                 {
-                    case EltwiseLayerNode::node_type:
+                    auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast<EltwiseLayerNode *>(post_op_node);
+                    ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr);
+                    if(eltwise_node->output(0)->accessor() == nullptr)
                     {
-                        auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast<EltwiseLayerNode *>(post_op_output_edge->producer());
-                        ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr);
-                        if(eltwise_node->output(0)->accessor() == nullptr)
+                        post_op_node_list.push_back(post_op_node);
+                        fusable_post_op                  = true;
+                        post_op_type_list[post_op_idx++] = eltwise_node->type();
+
+                        // Extract elementwise inputs
+                        const auto eltwise_input_id_0 = eltwise_node->input_edge(0)->producer_id();
+                        const auto eltwise_input_id_1 = eltwise_node->input_edge(1)->producer_id();
+                        if(eltwise_input_id_0 == prev_op_dst_id)
                         {
-                            post_op_node_list.push_back(post_op_output_edge->producer());
-                            fusable_post_op                  = true;
-                            post_op_type_list[post_op_idx++] = eltwise_node->type();
-
-                            // Extract elementwise inputs
-                            const auto eltwise_input_id_0 = eltwise_node->input_edge(0)->producer_id();
-                            const auto eltwise_input_id_1 = eltwise_node->input_edge(1)->producer_id();
-                            if(eltwise_input_id_0 == prev_op_dst_id)
-                            {
-                                eltwise_operand_id = eltwise_input_id_1;
-                                prev_op_dst_pos    = 0;
-                            }
-                            else if(eltwise_input_id_1 == prev_op_dst_id)
-                            {
-                                eltwise_operand_id = eltwise_input_id_0;
-                                prev_op_dst_pos    = 1;
-                            }
+                            eltwise_operand_id = eltwise_input_id_1;
+                            prev_op_dst_pos    = 0;
                         }
-                        else
+                        else if(eltwise_input_id_1 == prev_op_dst_id)
                         {
-                            ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with elementwise due to the presence of an output accessor\n");
+                            eltwise_operand_id = eltwise_input_id_0;
+                            prev_op_dst_pos    = 1;
                         }
-                        break;
                     }
-                    case ActivationLayerNode::node_type:
+                    else
                     {
-                        auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(post_op_output_edge->producer());
-                        ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr);
-                        // Check if activation is supported for fusion
-                        if(supported_fused_activations.count(act_node->activation_info().activation()) == 0)
-                        {
-                            break;
-                        }
-                        if(act_node->output(0)->accessor() == nullptr)
-                        {
-                            post_op_node_list.push_back(post_op_output_edge->producer());
-                            fusable_post_op                  = true;
-                            post_op_type_list[post_op_idx++] = act_node->type();
-                            prev_op_dst_id                   = act_node->id();
-                        }
-                        else
-                        {
-                            ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with activation due to the presence of an output accessor\n");
-                        }
-                        break;
+                        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with elementwise due to the presence of an output accessor\n");
                     }
-                    default:
+                    break;
+                }
+                case ActivationLayerNode::node_type:
+                {
+                    auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(post_op_node);
+                    ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr);
+                    // Check if activation is supported for fusion
+                    if(supported_fused_activations.count(act_node->activation_info().activation()) == 0)
                     {
                         break;
                     }
+                    if(act_node->output(0)->accessor() == nullptr)
+                    {
+                        post_op_node_list.push_back(post_op_node);
+                        fusable_post_op                  = true;
+                        post_op_type_list[post_op_idx++] = act_node->type();
+                        prev_op_dst_id                   = act_node->id();
+                    }
+                    else
+                    {
+                        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to the presence of an output accessor\n");
+                    }
+                    break;
+                }
+                default:
+                {
+                    break;
                 }
             }
 
             // Check if the node is not a branching node and current node is fusable
-            if(post_op_node->output_edges().size() == 1 && fusable_post_op == true && post_op_node_list.size() < 3)
+            if(post_op_node->output_edges().size() == 1 && fusable_post_op == true)
             {
-                continue;
+                current_node = post_op_node;
             }
             else
             {
@@ -534,12 +538,11 @@ std::list<INode *> get_post_op_list(Graph &g, int &eltwise_operand_id, int &prev
  *
  * Notes: currently, only GEMM supports fusion with post operator
 */
-template <typename N>
-void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const std::set<Activation> &supported_fused_activations)
+void fuse_convolution_with_post_ops(Graph &g, const Edge *output_edge, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations)
 {
     ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
 
-    auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<N *>(output_edge->producer());
+    auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<ConvolutionLayerNode *>(output_edge->producer());
     ARM_COMPUTE_ERROR_ON(conv_node->output(0) == nullptr);
 
     const ConvolutionMethod conv_algorithm = conv_node->convolution_method();
@@ -552,14 +555,14 @@ void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const
     // Prevent fusion if fused node has an output accessor
     if(conv_node->output(0)->accessor() == nullptr)
     {
-        // If data type is FP32/FP16, data layout is NHWC, and filter size if 1x1, fuse convolution with post op, as Conv1x1 always leads to GEMM.
+        // If data type is FP32/FP16, data layout is NHWC, and filter size is 1x1, fuse convolution with post op, as Conv1x1 always leads to GEMM.
         const Edge *input_edge = conv_node->input_edge(1);
         if(input_edge != nullptr && input_edge->tensor() != nullptr)
         {
             const DataLayout  data_layout  = input_edge->tensor()->desc().layout;
             const DataType    data_type    = input_edge->tensor()->desc().data_type;
             const TensorShape tensor_shape = input_edge->tensor()->desc().shape;
-            if(data_layout != DataLayout::NHWC || is_data_type_float(data_type) == false || tensor_shape.y() != 1 || tensor_shape.z() != 1)
+            if((data_layout != DataLayout::NHWC) || (is_data_type_float(data_type) == false) || (tensor_shape.y() != 1) || (tensor_shape.z() != 1))
             {
                 ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n");
                 return;
@@ -603,7 +606,18 @@ void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const
                 auto conv_bias_id = conv_node->input_edge(2)->producer_id();
                 g.add_connection(conv_bias_id, 0, fused_id, 2);
             }
-            g.add_connection(eltwise_operand_id, 0, fused_id, 3);
+            // Adding the Element wise operand in case the post op is element wise operation
+            auto it = std::find_if(post_op_node_list.begin(),
+                                   post_op_node_list.end(),
+                                   [&](const INode * nd)
+            {
+                return (nd->type() == graph::NodeType::EltwiseLayer);
+            });
+
+            if(it != post_op_node_list.end())
+            {
+                g.add_connection(eltwise_operand_id, 0, fused_id, 3);
+            }
             g.remove_node(conv_node->id());
 
             // Update fused node outputs
@@ -623,6 +637,136 @@ void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const
     }
 }
 
+void fuse_convolution_batch_normalization_with_post_ops(Graph &g, const Edge *output_edge, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations)
+{
+    ARM_COMPUTE_ERROR_ON(output_edge == nullptr);
+
+    auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(output_edge->producer());
+    ARM_COMPUTE_ERROR_ON(conv_node->output(0) == nullptr);
+    const ConvolutionMethod conv_algorithm = conv_node->convolution_method();
+    if(conv_algorithm != ConvolutionMethod::GEMM)
+    {
+        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n");
+        return;
+    }
+
+    // Prevent fusion if fused node has an output accessor
+    if(conv_node->output(0)->accessor() == nullptr)
+    {
+        // If data type is FP32/FP16, data layout is NHWC, and filter size is 1x1, fuse convolution with post op, as Conv1x1 always leads to GEMM.
+        const Edge *input_edge = conv_node->input_edge(1);
+        if(input_edge != nullptr && input_edge->tensor() != nullptr)
+        {
+            const DataLayout  data_layout  = input_edge->tensor()->desc().layout;
+            const DataType    data_type    = input_edge->tensor()->desc().data_type;
+            const TensorShape tensor_shape = input_edge->tensor()->desc().shape;
+            if((data_layout != DataLayout::NHWC) || (is_data_type_float(data_type) == false) || (tensor_shape.y() != 1) || (tensor_shape.z() != 1))
+            {
+                ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n");
+                return;
+            }
+        }
+        else
+        {
+            return;
+        }
+
+        // Get post op list
+        int                eltwise_operand_id = 0;
+        int                prev_op_dst_pos    = 0; // Previous operator dst's postion in current operator
+        std::list<INode *> post_op_node_list  = get_post_op_list(g, eltwise_operand_id, prev_op_dst_pos, conv_node_id, supported_fused_activations);
+
+        if(post_op_node_list.size() == 0)
+        {
+            return;
+        }
+        else // Do convolution fusion with post op if there're one(elementwise), two or more operators
+        {
+            const Target assigned_target = conv_node->assigned_target();
+
+            // 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   bn_mean_id      = conv_node->input_edge(3)->producer_id();
+            const auto   bn_var_id       = conv_node->input_edge(4)->producer_id();
+            const auto   conv_info       = conv_node->convolution_info();
+            const auto   conv_method     = conv_node->convolution_method();
+            const auto   num_groups      = conv_node->num_groups();
+            FastMathHint fast_math_hint  = conv_node->fast_math_hint();
+
+            // Create the fused node
+
+            const float  epsilon  = conv_node->epsilon();
+            const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationWithPostOpsNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint);
+
+            ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing FusedConvolutionBatchNormalization node with ID : " << conv_node->id());
+
+            // Add connections from the conv inputs to the fused node
+            g.add_connection(conv_input_id, 0, fused_id, 0);
+            g.add_connection(conv_weights_id, 0, fused_id, 1);
+
+            if(conv_node->input_edge(2) != nullptr)
+            {
+                auto conv_bias_id = conv_node->input_edge(2)->producer_id();
+                g.add_connection(conv_bias_id, 0, fused_id, 2);
+            }
+            g.add_connection(bn_mean_id, 0, fused_id, 3);
+            g.add_connection(bn_var_id, 0, fused_id, 4);
+
+            // Move connections of old FusedConvolutionBatchNormalization to the fused node
+            if(conv_node->input_edge(5) != nullptr)
+            {
+                const auto bn_beta_id = conv_node->input_edge(5)->producer_id();
+                g.add_connection(bn_beta_id, 0, fused_id, 5);
+            }
+
+            if(conv_node->input_edge(6) != nullptr)
+            {
+                const auto bn_gamma_id = conv_node->input_edge(6)->producer_id();
+                g.add_connection(bn_gamma_id, 0, fused_id, 6);
+            }
+
+            // Adding the Element wise operand in case the post op is element wise operation
+            auto it = std::find_if(post_op_node_list.begin(),
+                                   post_op_node_list.end(),
+                                   [&](const INode * nd)
+            {
+                return (nd->type() == graph::NodeType::EltwiseLayer);
+            });
+
+            if(it != post_op_node_list.end())
+            {
+                g.add_connection(eltwise_operand_id, 0, fused_id, 7);
+            }
+
+            // Update fused node outputs
+            auto fused_node = g.node(fused_id);
+            fused_node->set_assigned_target(assigned_target);
+
+            auto conv_node_name = conv_node->name();
+
+            // collect the post ops names
+            std::string post_ops_name = "";
+            for(auto &post_op : post_op_node_list)
+            {
+                post_ops_name += post_op->name();
+            }
+            fused_node->set_common_node_parameters(NodeParams{ conv_node->name() + "+" + post_ops_name, assigned_target });
+
+            // Fuse convolution with post op
+            fuse_convolution_with_post_op(g, fused_node, post_op_node_list, prev_op_dst_pos);
+
+            post_op_node_list.clear();
+            g.remove_node(conv_node->id());
+            ARM_COMPUTE_LOG_GRAPH_VERBOSE(std::endl);
+        }
+    }
+    else
+    {
+        ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to the presence of an output accessor\n");
+    }
+}
+
 template <typename N1, typename F, typename... Args>
 void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse_fcn, Args &&... optional_arguments)
 {
@@ -697,7 +841,7 @@ void NodeFusionMutator::mutate(Graph &g)
     // The fusion of PostOps to ConvolutionLayer:
     // It must occur after the fusion of PadLayer into ConvolutionLayer
     // It must occur before the fusion of normal ActivationLayer into ConvolutionLayer as it takes precedence
-    detail::fuse_layer<ConvolutionLayerNode>(g, cl_target_prec, detail::fuse_convolution<ConvolutionLayerNode>, supported_fused_activations);
+    detail::fuse_layer<ConvolutionLayerNode>(g, cl_target_prec, detail::fuse_convolution_with_post_ops, supported_fused_activations);
     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);
     detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>(g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations);
@@ -706,6 +850,7 @@ void NodeFusionMutator::mutate(Graph &g)
     // The fusion of BatchNormalizationLayer must occur after the fusion of ActivationLayer. Because FusedConvolutionBatchNormalizationNode assumes the BatchNormalization is already fused with activation, if any
     detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization);
     detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization);
+    detail::fuse_layer<FusedConvolutionBatchNormalizationNode>(g, cl_target_prec, detail::fuse_convolution_batch_normalization_with_post_ops, supported_fused_activations);
 }
 } // namespace graph
 } // namespace arm_compute
diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp
new file mode 100644
index 0000000000..af81f0369a
--- /dev/null
+++ b/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2021 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/FusedConvolutionBatchNormalizationWithPostOpsNode.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
+{
+FusedConvolutionBatchNormalizationWithPostOpsNode::FusedConvolutionBatchNormalizationWithPostOpsNode(float epsilon, PadStrideInfo info,
+                                                                                                     unsigned int      num_groups,
+                                                                                                     ConvolutionMethod method,
+                                                                                                     FastMathHint      fast_math_hint)
+    : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint)
+{
+    _input_edges.resize(8, EmptyEdgeID);
+    _outputs.resize(1, NullTensorID);
+}
+
+void FusedConvolutionBatchNormalizationWithPostOpsNode::set_convolution_method(ConvolutionMethod method)
+{
+    _method = method;
+}
+
+float FusedConvolutionBatchNormalizationWithPostOpsNode::epsilon() const
+{
+    return _epsilon;
+}
+
+ConvolutionMethod FusedConvolutionBatchNormalizationWithPostOpsNode::convolution_method() const
+{
+    return _method;
+}
+
+void FusedConvolutionBatchNormalizationWithPostOpsNode::set_fast_math_hint(FastMathHint hint)
+{
+    _fast_math_hint = hint;
+}
+
+FastMathHint FusedConvolutionBatchNormalizationWithPostOpsNode::fast_math_hint() const
+{
+    return _fast_math_hint;
+}
+
+PadStrideInfo FusedConvolutionBatchNormalizationWithPostOpsNode::convolution_info() const
+{
+    return _info;
+}
+
+unsigned int FusedConvolutionBatchNormalizationWithPostOpsNode::num_groups() const
+{
+    return _num_groups;
+}
+
+TensorDescriptor FusedConvolutionBatchNormalizationWithPostOpsNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
+                                                                                              const TensorDescriptor &weights_descriptor,
+                                                                                              const PadStrideInfo    &info)
+{
+    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 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);
+
+    const DataLayout data_layout       = input_descriptor.layout;
+    TensorDescriptor output_descriptor = input_descriptor;
+    output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width);
+    output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height);
+    output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]);
+
+    return output_descriptor;
+}
+
+bool FusedConvolutionBatchNormalizationWithPostOpsNode::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 FusedConvolutionBatchNormalizationWithPostOpsNode::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);
+
+    return output_info;
+}
+
+NodeType FusedConvolutionBatchNormalizationWithPostOpsNode::type() const
+{
+    return FusedConvolutionBatchNormalizationWithPostOpsNode::node_type;
+}
+
+void FusedConvolutionBatchNormalizationWithPostOpsNode::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 47371e34d5..1071d50197 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(FusedConvolutionBatchNormalizationWithPostOpsNode &n)
+{
+    ARM_COMPUTE_UNUSED(n);
+    std::stringstream ss;
+    ss << "FusedConvolutionBatchNormalizationWithPostOpsNode";
+    _info = ss.str();
+}
+
 void DotGraphVisitor::visit(FusedConvolutionWithPostOpNode &n)
 {
     ARM_COMPUTE_UNUSED(n);