COMPMID-1808: Add Detection Output Layer to the GraphAPI

COMPMID-1710: Integrate Detection ouput in MobilenetSSD graph example

Change-Id: I384d1eb492ef14ece58f2023ad7bbc16f834450b
Reviewed-on: https://review.mlplatform.org/356
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

7 files changed, 211 insertions, 1 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 3fc258d8bd..d09002d69b 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -362,6 +362,22 @@ NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params,
 
     return conv_nid;
 }
+NodeID GraphBuilder::add_detection_output_node(Graph &g, NodeParams params, NodeIdxPair input_loc, NodeIdxPair input_conf, NodeIdxPair input_priorbox, DetectionOutputLayerInfo detect_info)
+{
+    CHECK_NODEIDX_PAIR(input_loc, g);
+    CHECK_NODEIDX_PAIR(input_conf, g);
+    CHECK_NODEIDX_PAIR(input_priorbox, g);
+
+    // Create detection_output node and connect
+    NodeID detect_nid = g.add_node<DetectionOutputLayerNode>(detect_info);
+    g.add_connection(input_loc.node_id, input_loc.index, detect_nid, 0);
+    g.add_connection(input_conf.node_id, input_conf.index, detect_nid, 1);
+    g.add_connection(input_priorbox.node_id, input_priorbox.index, 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)
 {
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index c37a137cf7..5b329c04be 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/graph/Graph.h"
 #include "arm_compute/graph/backends/FunctionHelpers.h"
 #include "arm_compute/runtime/CL/CLFunctions.h"
+#include "arm_compute/runtime/CPP/CPPFunctions.h"
 
 using namespace arm_compute::utils::cast;
 
@@ -68,6 +69,94 @@ struct CLEltwiseFunctions
     using Subtraction    = CLArithmeticSubtraction;
     using Multiplication = CLPixelWiseMultiplication;
 };
+// TODO (isagot01): Remove once we support heterogeneous scheduling at function level
+/** Wrapper for the CPP Function in the OpenCL backend **/
+class CPPWrapperFunction : public IFunction
+{
+public:
+    /* Default constructor */
+    CPPWrapperFunction()
+        : _tensors(), _func(nullptr)
+    {
+    }
+
+    void run() override
+    {
+        for(auto &tensor : _tensors)
+        {
+            tensor->map(CLScheduler::get().queue());
+        }
+        _func->run();
+
+        for(auto &tensor : _tensors)
+        {
+            tensor->unmap(CLScheduler::get().queue());
+        }
+    }
+
+    void register_tensor(ICLTensor *tensor)
+    {
+        _tensors.push_back(tensor);
+    }
+
+    void register_function(std::unique_ptr<IFunction> function)
+    {
+        _func = std::move(function);
+    }
+
+private:
+    std::vector<arm_compute::ICLTensor *> _tensors;
+    std::unique_ptr<IFunction>            _func;
+};
+
+namespace detail
+{
+// Specialized functions
+template <>
+std::unique_ptr<IFunction> create_detection_output_layer<CPPDetectionOutputLayer, CLTargetInfo>(DetectionOutputLayerNode &node)
+{
+    validate_node<CLTargetInfo>(node, 3 /* expected inputs */, 1 /* 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      *output      = get_backing_tensor<CLTargetInfo>(node.output(0));
+    const DetectionOutputLayerInfo detect_info = node.detection_output_info();
+
+    ARM_COMPUTE_ERROR_ON(input0 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input1 == nullptr);
+    ARM_COMPUTE_ERROR_ON(input2 == nullptr);
+    ARM_COMPUTE_ERROR_ON(output == nullptr);
+
+    // Create and configure function
+    auto func = support::cpp14::make_unique<CPPDetectionOutputLayer>();
+    func->configure(input0, input1, input2, output, 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()
+                               << " Output shape: " << output->info()->tensor_shape()
+                               << " DetectionOutputLayer 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(output);
+
+    return std::move(wrap_function);
+}
+} // namespace detail
 
 std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &ctx)
 {
@@ -95,6 +184,8 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext &
             return detail::create_concatenate_layer<CLConcatenateLayer, CLTargetInfo>(*polymorphic_downcast<ConcatenateLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
             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::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 a070973fd4..85ac1f59c6 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -28,6 +28,7 @@
 
 #include "arm_compute/core/utils/misc/Cast.h"
 #include "arm_compute/runtime/CL/CLFunctions.h"
+#include "arm_compute/runtime/CPP/CPPFunctions.h"
 
 using namespace arm_compute::utils::cast;
 
@@ -59,6 +60,8 @@ Status CLNodeValidator::validate(INode *node)
         case NodeType::DepthwiseConvolutionLayer:
             return detail::validate_depthwise_convolution_layer<CLDepthwiseConvolutionLayer,
                    CLDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+        case NodeType::DetectionOutputLayer:
+            return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(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 fe69c7a9ee..95bb44f5cc 100644
--- a/src/graph/backends/GLES/GCNodeValidator.cpp
+++ b/src/graph/backends/GLES/GCNodeValidator.cpp
@@ -111,6 +111,8 @@ Status GCNodeValidator::validate(INode *node)
             return validate_convolution_layer(*polymorphic_downcast<ConvolutionLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
             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::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 ca8d485f8b..dc987dd86e 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -31,6 +31,7 @@
 #include "arm_compute/graph/backends/FunctionHelpers.h"
 #include "arm_compute/graph/backends/Utils.h"
 #include "arm_compute/graph/nodes/Nodes.h"
+#include "arm_compute/runtime/CPP/CPPFunctions.h"
 #include "arm_compute/runtime/NEON/NEFunctions.h"
 #include "support/ToolchainSupport.h"
 
@@ -77,7 +78,7 @@ struct NEEltwiseFunctions
 
 namespace detail
 {
-// Specialize functions
+// Specialized functions
 template <>
 std::unique_ptr<IFunction> create_convolution_layer<NEConvolutionLayerFunctions, NETargetInfo>(ConvolutionLayerNode &node,
                                                                                                GraphContext &ctx)
@@ -201,6 +202,8 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
             return detail::create_concatenate_layer<NEConcatenateLayer, NETargetInfo>(*polymorphic_downcast<ConcatenateLayerNode *>(node));
         case NodeType::DepthwiseConvolutionLayer:
             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::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 a2abc8330c..db6af5eab7 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/graph/nodes/Nodes.h"
 
 #include "arm_compute/core/utils/misc/Cast.h"
+#include "arm_compute/runtime/CPP/CPPFunctions.h"
 #include "arm_compute/runtime/NEON/NEFunctions.h"
 
 using namespace arm_compute::utils::cast;
@@ -59,6 +60,8 @@ Status NENodeValidator::validate(INode *node)
         case NodeType::DepthwiseConvolutionLayer:
             return detail::validate_depthwise_convolution_layer<NEDepthwiseConvolutionLayer,
                    NEDepthwiseConvolutionLayer3x3>(*polymorphic_downcast<DepthwiseConvolutionLayerNode *>(node));
+        case NodeType::DetectionOutputLayer:
+            return detail::validate_detection_output_layer<CPPDetectionOutputLayer>(*polymorphic_downcast<DetectionOutputLayerNode *>(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/DetectionOutputLayerNode.cpp b/src/graph/nodes/DetectionOutputLayerNode.cpp
new file mode 100644
index 0000000000..c2d9f2446f
--- /dev/null
+++ b/src/graph/nodes/DetectionOutputLayerNode.cpp
@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/nodes/DetectionOutputLayerNode.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
+{
+DetectionOutputLayerNode::DetectionOutputLayerNode(DetectionOutputLayerInfo detection_info)
+    : _info(detection_info)
+{
+    _input_edges.resize(3, EmptyEdgeID);
+    _outputs.resize(1, NullTensorID);
+}
+
+DetectionOutputLayerInfo DetectionOutputLayerNode::detection_output_info() const
+{
+    return _info;
+}
+
+TensorDescriptor DetectionOutputLayerNode::compute_output_descriptor(const TensorDescriptor         &input_descriptor,
+                                                                     const DetectionOutputLayerInfo &info)
+{
+    const unsigned int max_size = info.keep_top_k() * ((input_descriptor.shape.num_dimensions() > 1) ? input_descriptor.shape[1] : 1);
+
+    TensorDescriptor output_descriptor = input_descriptor;
+    output_descriptor.shape.set(0, detection_size);
+    output_descriptor.shape.set(1, max_size);
+
+    return output_descriptor;
+}
+
+bool DetectionOutputLayerNode::forward_descriptors()
+{
+    if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (input_id(2) != 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 DetectionOutputLayerNode::configure_output(size_t idx) const
+{
+    ARM_COMPUTE_UNUSED(idx);
+    ARM_COMPUTE_ERROR_ON(idx >= _outputs.size());
+
+    const Tensor *input0 = input(0);
+    ARM_COMPUTE_ERROR_ON(input0 == nullptr);
+
+    return compute_output_descriptor(input0->desc(), _info);
+}
+
+NodeType DetectionOutputLayerNode::type() const
+{
+    return NodeType::DetectionOutputLayer;
+}
+
+void DetectionOutputLayerNode::accept(INodeVisitor &v)
+{
+    v.visit(*this);
+}
+} // namespace graph
+} // namespace arm_compute