Add graph level convolution fusion with post operator

Resolves: COMPMID-4701

Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I8a0d3c2ed4bf84489d94b8ae6641d6041aadaee5
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6557
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

9 files changed, 360 insertions, 3 deletions

diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h
index 85da59022b..4cb601fb02 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 FusedConvolutionWithPostOpNode.
+     *
+     * @param[in] n Node to visit.
+     */
+    virtual void visit(FusedConvolutionWithPostOpNode &n) = 0;
     /** Visit FusedDepthwiseConvolutionBatchNormalizationNode.
      *
      * @param[in] n Node to visit.
@@ -205,6 +210,7 @@ public:
     virtual void visit(FlattenLayerNode &n) override;
     virtual void visit(FullyConnectedLayerNode &n) override;
     virtual void visit(FusedConvolutionBatchNormalizationNode &n) override;
+    virtual void visit(FusedConvolutionWithPostOpNode &n) override;
     virtual void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     virtual void visit(InputNode &n) override;
     virtual void visit(NormalizationLayerNode &n) override;
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index 0bbb4695de..a8a20c9de8 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::FusedConvolutionWithPostOp:
+            os << "FusedConvolutionWithPostOp";
+            break;
         case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer:
             os << "FusedDepthwiseConvolutionBatchNormalizationLayer";
             break;
diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h
index 63a9433fe6..e802e9dc77 100644
--- a/arm_compute/graph/Types.h
+++ b/arm_compute/graph/Types.h
@@ -62,6 +62,7 @@ using arm_compute::PoolingType;
 using arm_compute::PriorBoxLayerInfo;
 using arm_compute::DimensionRoundingType;
 using arm_compute::InterpolationPolicy;
+using arm_compute::experimental::PostOpType;
 
 using GraphID    = unsigned int;
 using TensorID   = unsigned int;
@@ -145,6 +146,55 @@ enum class FastMathHint
     Disabled, /**< Fast math disabled for Convolution layer */
 };
 
+/** Convolution post operator info */
+class ConvPostOpInfo
+{
+public:
+    /** Returns post op type
+     *
+     * @return Post op type
+     */
+    virtual PostOpType type() const = 0;
+    virtual ~ConvPostOpInfo()
+    {
+    }
+};
+
+class ConvPostOpInfoActivation : public ConvPostOpInfo
+{
+public:
+    ConvPostOpInfoActivation(const ActivationLayerInfo &act)
+        : _act(act)
+    {
+    }
+    ~ConvPostOpInfoActivation() override
+    {
+    }
+    PostOpType type() const override
+    {
+        return PostOpType::Activation;
+    }
+    ActivationLayerInfo _act;
+};
+
+class ConvPostOpInfoEltwiseAdd : public ConvPostOpInfo
+{
+public:
+    ConvPostOpInfoEltwiseAdd(int arg_pos, const ConvertPolicy &policy)
+        : _prev_op_dst_pos(arg_pos), _policy(policy)
+    {
+    }
+    PostOpType type() const override
+    {
+        return PostOpType::Eltwise_Add;
+    }
+    ~ConvPostOpInfoEltwiseAdd() override
+    {
+    }
+    int           _prev_op_dst_pos;
+    ConvertPolicy _policy;
+};
+
 /** Supported nodes */
 enum class NodeType
 {
@@ -165,6 +215,7 @@ enum class NodeType
     FlattenLayer,
     FullyConnectedLayer,
     FusedConvolutionBatchNormalizationLayer,
+    FusedConvolutionWithPostOp,
     FusedDepthwiseConvolutionBatchNormalizationLayer,
     GenerateProposalsLayer,
     L2NormalizeLayer,
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 9830290d0f..6bec66a6ff 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -450,7 +450,7 @@ std::unique_ptr<arm_compute::IFunction> create_concatenate_layer(ConcatenateLaye
 /** Create a backend convolution layer function
  *
  * @tparam ConvolutionLayerFunctions Backend convolution functions
- * @tparam TargetInfo              Target-specific information
+ * @tparam TargetInfo                Target-specific information
  *
  * @param[in] node Node to create the backend function for
  * @param[in] ctx  Graph context
@@ -538,6 +538,98 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     return std::move(func);
 }
 
+/** Create a backend convolution layer function with post opreator
+ *
+ * @tparam ConvolutionLayerFunctions 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 convolution layer function
+ */
+template <typename ConvolutionLayerFunctions, typename TargetInfo>
+std::unique_ptr<IFunction> create_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node, GraphContext &ctx)
+{
+    validate_node<TargetInfo>(node, 4 /* 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 *output  = get_backing_tensor<TargetInfo>(node.output(0));
+
+    const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+
+    if(is_quantized)
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
+    const PadStrideInfo       conv_info  = node.convolution_info();
+    const unsigned int        num_groups = node.num_groups();
+    const ActivationLayerInfo fused_act  = node.fused_activation();
+
+    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;
+
+    std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::GEMMConvolutionLayer>(
+                                    std::string("GEMMConvolutionLayer"), mm,
+                                    input, weights, biases, output, conv_info,
+                                    WeightsInfo(), Size2D(1U, 1U), fused_act, num_groups, post_ops);
+
+    // Log info
+    std::ostringstream qss;
+    if(is_quantized)
+    {
+        qss << " Input QuantInfo: " << input->info()->quantization_info()
+            << " Weights QuantInfo: " << weights->info()->quantization_info()
+            << " Output QuantInfo: " << output->info()->quantization_info();
+    }
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << func_name
+                               << " Target: " << TargetInfo::TargetType
+                               << " Data Type: " << input->info()->data_type()
+                               << " Groups: " << num_groups
+                               << " Input shape: " << input->info()->tensor_shape()
+                               << " Weights shape: " << weights->info()->tensor_shape()
+                               << " Output shape: " << output->info()->tensor_shape()
+                               << qss.str()
+                               << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
+                               << std::endl);
+    return std::move(func);
+}
+
 /** Create a backend deconvolution layer function
  *
  * @tparam DeconvolutionLayerFunction Backend deconvolution function
diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h
index 93d547b036..89dccd88b7 100644
--- a/arm_compute/graph/backends/ValidateHelpers.h
+++ b/arm_compute/graph/backends/ValidateHelpers.h
@@ -183,6 +183,42 @@ Status validate_convolution_layer(ConvolutionLayerNode &node)
     return status;
 }
 
+/** Validates a Convolution layer node
+ *
+ * @tparam GEMMConvolutionLayer      GEMM Convolution layer function type
+ *
+ * @param[in] node Node to validate
+ *
+ * @return Status
+ */
+template <typename GEMMConvolutionLayer>
+Status validate_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node)
+{
+    ARM_COMPUTE_LOG_GRAPH_VERBOSE("Validating fused ConvolutionLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl);
+    ARM_COMPUTE_RETURN_ERROR_ON(node.num_inputs() != 4);
+    ARM_COMPUTE_RETURN_ERROR_ON(node.num_outputs() != 1);
+
+    // Extract IO and info
+    arm_compute::ITensorInfo *input   = get_backing_tensor_info(node.input(0));
+    arm_compute::ITensorInfo *weights = get_backing_tensor_info(node.input(1));
+    arm_compute::ITensorInfo *biases  = get_backing_tensor_info(node.input(2));
+    arm_compute::ITensorInfo *output  = get_backing_tensor_info(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->data_type()))
+    {
+        biases->set_data_type(DataType::S32);
+    }
+
+    const PadStrideInfo conv_info = node.convolution_info();
+    //const ConvolutionMethod conv_algorithm = node.convolution_method();
+    //const bool              fast_math      = node.fast_math_hint() == FastMathHint::Enabled;
+    const unsigned int num_groups = node.num_groups();
+
+    // Validate function
+    return GEMMConvolutionLayer::validate(input, weights, biases, output, conv_info,
+                                          WeightsInfo(), Size2D(1, 1), ActivationLayerInfo(), num_groups);
+}
+
 /** Validates a Depthwise Convolution layer node
  *
  * @tparam DepthwiseConvolutionLayer    Default Depthwise Convolution layer type
diff --git a/arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h b/arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h
new file mode 100644
index 0000000000..0bb3d1b7ef
--- /dev/null
+++ b/arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h
@@ -0,0 +1,143 @@
+/*
+ * 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_WITH_POST_OP_NODE_H
+#define ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_WITH_POST_OP_NODE_H
+
+#include "arm_compute/graph/INode.h"
+
+#include <list>
+
+namespace arm_compute
+{
+namespace graph
+{
+/** Convolution node */
+class FusedConvolutionWithPostOpNode final : public INode
+{
+public:
+    /** Constructor
+     *
+     * @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
+     * @param[in] out_quant_info (Optional) Output quantization info
+     */
+    FusedConvolutionWithPostOpNode(PadStrideInfo     info,
+                                   unsigned int      num_groups     = 1,
+                                   ConvolutionMethod method         = ConvolutionMethod::Default,
+                                   FastMathHint      fast_math_hint = FastMathHint::Disabled,
+                                   QuantizationInfo  out_quant_info = QuantizationInfo());
+    /** Sets the convolution layer method to use
+     *
+     * @param[in] method Method to use for convolution
+     */
+    void set_convolution_method(ConvolutionMethod method);
+    /** 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 fast 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;
+    /** Post operator info list
+     *
+     * @return Post operator info list
+     */
+    const std::list<std::unique_ptr<ConvPostOpInfo>> &post_op_info_list() const;
+    /** Post operator info list
+     *
+     * @return Post operator info list
+     */
+    std::list<std::unique_ptr<ConvPostOpInfo>> &post_op_info_list();
+    /** Convolution metadata accessor
+     *
+     * @return Convolution information
+     */
+    PadStrideInfo convolution_info() const;
+    /** Number of groups in convolution accessor
+     *
+     * @return Number of groups in convolution
+     */
+    unsigned int num_groups() 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);
+    /** Sets convolution info
+     *
+     * @param[in] info Convolution info to set
+     */
+    void set_convolution_info(PadStrideInfo info);
+    /** 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);
+
+    // 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::FusedConvolutionWithPostOp;
+
+private:
+    PadStrideInfo                              _info;
+    unsigned int                               _num_groups;
+    ConvolutionMethod                          _method;
+    FastMathHint                               _fast_math_hint;
+    QuantizationInfo                           _out_quant_info;
+    ActivationLayerInfo                        _fused_activation;
+    std::list<std::unique_ptr<ConvPostOpInfo>> _post_op_info_list;
+};
+} // namespace graph
+} // namespace arm_compute
+
+#endif /* ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_WITH_POST_OP_NODE_H */
diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h
index edb1876722..fb0eb155fd 100644
--- a/arm_compute/graph/nodes/Nodes.h
+++ b/arm_compute/graph/nodes/Nodes.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -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/FusedConvolutionWithPostOpNode.h"
 #include "arm_compute/graph/nodes/FusedDepthwiseConvolutionBatchNormalizationNode.h"
 #include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h"
 #include "arm_compute/graph/nodes/InputNode.h"
diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h
index 485361296a..6393b1d5d4 100644
--- a/arm_compute/graph/nodes/NodesFwd.h
+++ b/arm_compute/graph/nodes/NodesFwd.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -49,6 +49,7 @@ class EltwiseLayerNode;
 class FlattenLayerNode;
 class FullyConnectedLayerNode;
 class FusedConvolutionBatchNormalizationNode;
+class FusedConvolutionWithPostOpNode;
 class FusedDepthwiseConvolutionBatchNormalizationNode;
 class GenerateProposalsLayerNode;
 class InputNode;
diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h
index 3c9de8cba9..41607e831a 100644
--- a/arm_compute/graph/printers/DotGraphPrinter.h
+++ b/arm_compute/graph/printers/DotGraphPrinter.h
@@ -1,4 +1,27 @@
 /*
+ * 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.
+ */
+/*
  * Copyright (c) 2018-2019,2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
@@ -57,6 +80,7 @@ public:
     void visit(DepthwiseConvolutionLayerNode &n) override;
     void visit(EltwiseLayerNode &n) override;
     void visit(FusedConvolutionBatchNormalizationNode &n) override;
+    void visit(FusedConvolutionWithPostOpNode &n) override;
     void visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) override;
     void visit(NormalizationLayerNode &n) override;
     void visit(PoolingLayerNode &n) override;