COMPMID-1849: Implement CPPDetectionPostProcessLayer

* Add DetectionPostProcessLayer
* Add DetectionPostProcessLayer at the graph

Change-Id: I7e56f6cffc26f112d26dfe74853085bb8ec7d849
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1639
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 200 insertions, 0 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 54bd066712..228f2d211a 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -393,6 +393,36 @@ NodeID GraphBuilder::add_detection_output_node(Graph &g, NodeParams params, Node
     return detect_nid;
 }
 
+NodeID GraphBuilder::add_detection_post_process_node(Graph &g, NodeParams params, NodeIdxPair input_box_encoding, NodeIdxPair input_class_prediction, const DetectionPostProcessLayerInfo &detect_info,
+                                                     ITensorAccessorUPtr anchors_accessor, const QuantizationInfo &anchor_quant_info)
+{
+    check_nodeidx_pair(input_box_encoding, g);
+    check_nodeidx_pair(input_class_prediction, g);
+
+    // Get input tensor descriptor
+    const TensorDescriptor input_box_encoding_tensor_desc = get_tensor_descriptor(g, g.node(input_box_encoding.node_id)->outputs()[0]);
+
+    // Calculate anchor descriptor
+    TensorDescriptor anchor_desc = input_box_encoding_tensor_desc;
+    if(!anchor_quant_info.empty())
+    {
+        anchor_desc.quant_info = anchor_quant_info;
+    }
+
+    // Create anchors nodes
+    auto anchors_nid = add_const_node_with_name(g, params, "Anchors", anchor_desc, std::move(anchors_accessor));
+
+    // Create detection_output node and connect
+    NodeID detect_nid = g.add_node<DetectionPostProcessLayerNode>(detect_info);
+    g.add_connection(input_box_encoding.node_id, input_box_encoding.index, detect_nid, 0);
+    g.add_connection(input_class_prediction.node_id, input_class_prediction.index, detect_nid, 1);
+    g.add_connection(anchors_nid, 0, detect_nid, 2);
+
+    set_node_params(g, detect_nid, params);
+
+    return detect_nid;
+}
+
 NodeID GraphBuilder::add_dummy_node(Graph &g, NodeParams params, NodeIdxPair input, TensorShape shape)
 {
     return create_simple_single_input_output_node<DummyNode>(g, params, input, shape);
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index b9f22f6199..82b6dd6a54 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -166,6 +166,62 @@ std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer
 
     return std::move(wrap_function);
 }
+template <>
+std::unique_ptr<IFunction> create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(DetectionPostProcessLayerNode &node)
+{
+    validate_node<CLTargetInfo>(node, 3 /* expected inputs */, 4 /* expected outputs */);
+
+    // Extract IO and info
+    CLTargetInfo::TensorType           *input0      = get_backing_tensor<CLTargetInfo>(node.input(0));
+    CLTargetInfo::TensorType           *input1      = get_backing_tensor<CLTargetInfo>(node.input(1));
+    CLTargetInfo::TensorType           *input2      = get_backing_tensor<CLTargetInfo>(node.input(2));
+    CLTargetInfo::TensorType           *output0     = get_backing_tensor<CLTargetInfo>(node.output(0));
+    CLTargetInfo::TensorType           *output1     = get_backing_tensor<CLTargetInfo>(node.output(1));
+    CLTargetInfo::TensorType           *output2     = get_backing_tensor<CLTargetInfo>(node.output(2));
+    CLTargetInfo::TensorType           *output3     = get_backing_tensor<CLTargetInfo>(node.output(3));
+    const DetectionPostProcessLayerInfo detect_info = node.detection_post_process_info();
+
+    ARM_COMPUTE_ERROR_ON(input0 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input1 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input2 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output0 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output1 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output2 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output3 == nullptr);
+
+    // Create and configure function
+    auto func = support::cpp14::make_unique<CPPDetectionPostProcessLayer>();
+    func->configure(input0, input1, input2, output0, output1, output2, output3, detect_info);
+
+    // Log info
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << node.type()
+                               << " Target: " << CLTargetInfo::TargetType
+                               << " Data Type: " << input0->info()->data_type()
+                               << " Input0 shape: " << input0->info()->tensor_shape()
+                               << " Input1 shape: " << input1->info()->tensor_shape()
+                               << " Input2 shape: " << input2->info()->tensor_shape()
+                               << " Output0 shape: " << output0->info()->tensor_shape()
+                               << " Output1 shape: " << output1->info()->tensor_shape()
+                               << " Output2 shape: " << output2->info()->tensor_shape()
+                               << " Output3 shape: " << output3->info()->tensor_shape()
+                               << " DetectionPostProcessLayer info: " << detect_info
+                               << std::endl);
+
+    auto wrap_function = support::cpp14::make_unique<CPPWrapperFunction>();
+
+    wrap_function->register_function(std::move(func));
+    wrap_function->register_tensor(input0);
+    wrap_function->register_tensor(input1);
+    wrap_function->register_tensor(input2);
+    wrap_function->register_tensor(output0);
+    wrap_function->register_tensor(output1);
+    wrap_function->register_tensor(output2);
+    wrap_function->register_tensor(output3);
+
+    return std::move(wrap_function);
+}
 } // namespace detail
 
 std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &ctx)
@@ -196,6 +252,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_depthwise_convolution_layer<CLDepthwiseConvolutionLayerFunctions, CLTargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, CLTargetInfo>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::EltwiseLayer:
             return detail::create_eltwise_layer<CLEltwiseFunctions, CLTargetInfo>(*polymorphic_downcast<EltwiseLayerNode *>(node));
         case NodeType::FlattenLayer:
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index 78771102e8..40ec508767 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -62,6 +62,8 @@ Status CLNodeValidator::validate(INode *node)
                    CLDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return detail::validate_generate_proposals_layer<CLGenerateProposalsLayer>(*polymorphic_downcast<GenerateProposalsLayerNode *>(node));
         case NodeType::NormalizePlanarYUVLayer:
diff --git a/src/graph/backends/GLES/GCNodeValidator.cpp b/src/graph/backends/GLES/GCNodeValidator.cpp
index a767d7b107..9cbb9a12ef 100644
--- a/src/graph/backends/GLES/GCNodeValidator.cpp
+++ b/src/graph/backends/GLES/GCNodeValidator.cpp
@@ -113,6 +113,8 @@ Status GCNodeValidator::validate(INode *node)
             return validate_depthwise_convolution_layer(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : DetectionOutputLayer");
+        case NodeType::DetectionPostProcessLayer:
+            return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : DetectionPostProcessLayer");
         case NodeType::FlattenLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : FlattenLayer");
         case NodeType::GenerateProposalsLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index b808ef81f9..852de549fa 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -215,6 +215,8 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
             return detail::create_depthwise_convolution_layer<NEDepthwiseConvolutionLayerFunctions, NETargetInfo>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::create_detection_output_layer<CPPDetectionOutputLayer, NETargetInfo>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::create_detection_post_process_layer<CPPDetectionPostProcessLayer, NETargetInfo>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::EltwiseLayer:
             return detail::create_eltwise_layer<NEEltwiseFunctions, NETargetInfo>(*polymorphic_downcast<EltwiseLayerNode *>(node));
         case NodeType::FlattenLayer:
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index 3b1d2aa59c..734b3401f7 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -62,6 +62,8 @@ Status NENodeValidator::validate(INode *node)
                    NEDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
         case NodeType::DetectionOutputLayer:
             return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(node));
+        case NodeType::DetectionPostProcessLayer:
+            return detail::validate_detection_post_process_layer<CPPDetectionPostProcessLayer>(*polymorphic_downcast<DetectionPostProcessLayerNode *>(node));
         case NodeType::GenerateProposalsLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : GenerateProposalsLayer");
         case NodeType::NormalizePlanarYUVLayer:
diff --git a/src/graph/nodes/DetectionPostProcessLayerNode.cpp b/src/graph/nodes/DetectionPostProcessLayerNode.cpp
new file mode 100644
index 0000000000..4a5df1ac4e
--- /dev/null
+++ b/src/graph/nodes/DetectionPostProcessLayerNode.cpp
@@ -0,0 +1,104 @@
+/*
+ * 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/DetectionPostProcessLayerNode.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
+{
+DetectionPostProcessLayerNode::DetectionPostProcessLayerNode(DetectionPostProcessLayerInfo detection_info)
+    : _info(detection_info)
+{
+    _input_edges.resize(3, EmptyEdgeID);
+    _outputs.resize(4, NullTensorID);
+}
+
+DetectionPostProcessLayerInfo DetectionPostProcessLayerNode::detection_post_process_info() const
+{
+    return _info;
+}
+
+bool DetectionPostProcessLayerNode::forward_descriptors()
+{
+    if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (input_id(2) != NullTensorID) && (output_id(0) != NullTensorID) && (output_id(1) != NullTensorID)
+       && (output_id(2) != NullTensorID) && (output_id(3) != NullTensorID))
+    {
+        for(unsigned int i = 0; i < 4; ++i)
+        {
+            Tensor *dst = output(i);
+            ARM_COMPUTE_ERROR_ON(dst == nullptr);
+            dst->desc() = configure_output(i);
+        }
+        return true;
+    }
+    return false;
+}
+
+TensorDescriptor DetectionPostProcessLayerNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_UNUSED(idx);
+    ARM_COMPUTE_ERROR_ON(idx >= _outputs.size());
+
+    TensorDescriptor   output_desc;
+    const unsigned int num_detected_box = _info.max_detections() * _info.max_classes_per_detection();
+
+    switch(idx)
+    {
+        case 0:
+            // Configure boxes output
+            output_desc.shape = TensorShape(kNumCoordBox, num_detected_box, kBatchSize);
+            break;
+        case 1:
+        case 2:
+            // Configure classes or scores output
+            output_desc.shape = TensorShape(num_detected_box, kBatchSize);
+            break;
+        case 3:
+            // Configure num_detection
+            output_desc.shape = TensorShape(1);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Unsupported output index");
+    }
+    output_desc.data_type = DataType::F32;
+
+    return output_desc;
+}
+
+NodeType DetectionPostProcessLayerNode::type() const
+{
+    return NodeType::DetectionPostProcessLayer;
+}
+
+void DetectionPostProcessLayerNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute
\ No newline at end of file