COMPMID-2588: Optimize the output detection kernel required by MobileNet-SSD (~27% improvement)

Change-Id: Ic6ce570af3878a0666ec680e0efabba3fcfd1222
Signed-off-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2160
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 107 insertions, 9 deletions

diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index d8b0ae92ea..12f44e303e 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -210,7 +210,7 @@ std::unique_ptr<IFunction> NEFunctionFactory::create(INode *node, GraphContext &
         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));
+            return detail::create_detection_post_process_layer<NEDetectionPostProcessLayer, 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 0b53657c42..f17b116892 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -62,7 +62,7 @@ Status NENodeValidator::validate(INode *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));
+            return detail::validate_detection_post_process_layer<NEDetectionPostProcessLayer>(*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/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
index 0addb0ead3..bc88f71af4 100644
--- a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
+++ b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
@@ -42,7 +42,7 @@ Status validate_arguments(const ITensorInfo *input_box_encoding, const ITensorIn
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input_box_encoding, input_class_score, input_anchors);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_box_encoding, 1, DataType::F32, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_box_encoding, input_class_score, input_anchors);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_box_encoding, input_anchors);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(input_box_encoding->num_dimensions() > 3, "The location input tensor shape should be [4, N, kBatchSize].");
     if(input_box_encoding->num_dimensions() > 2)
     {
@@ -183,8 +183,8 @@ void SaveOutputs(const Tensor *decoded_boxes, const std::vector<int> &result_idx
 
 CPPDetectionPostProcessLayer::CPPDetectionPostProcessLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _nms(), _input_box_encoding(nullptr), _input_scores(nullptr), _input_anchors(nullptr), _output_boxes(nullptr), _output_classes(nullptr),
-      _output_scores(nullptr), _num_detection(nullptr), _info(), _num_boxes(), _num_classes_with_background(), _num_max_detected_boxes(), _decoded_boxes(), _decoded_scores(), _selected_indices(),
-      _class_scores(), _input_scores_to_use(nullptr)
+      _output_scores(nullptr), _num_detection(nullptr), _info(), _num_boxes(), _num_classes_with_background(), _num_max_detected_boxes(), _dequantize_scores(false), _decoded_boxes(), _decoded_scores(),
+      _selected_indices(), _class_scores(), _input_scores_to_use(nullptr)
 {
 }
 
@@ -214,6 +214,7 @@ void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding,
     _info                        = info;
     _num_boxes                   = input_box_encoding->info()->dimension(1);
     _num_classes_with_background = _input_scores->info()->dimension(0);
+    _dequantize_scores           = (info.dequantize_scores() && is_data_type_quantized(input_box_encoding->info()->data_type()));
 
     auto_init_if_empty(*_decoded_boxes.info(), TensorInfo(TensorShape(_kNumCoordBox, _input_box_encoding->info()->dimension(1), _kBatchSize), 1, DataType::F32));
     auto_init_if_empty(*_decoded_scores.info(), TensorInfo(TensorShape(_input_scores->info()->dimension(0), _input_scores->info()->dimension(1), _kBatchSize), 1, DataType::F32));
@@ -221,7 +222,7 @@ void CPPDetectionPostProcessLayer::configure(const ITensor *input_box_encoding,
     const unsigned int num_classes_per_box = std::min(info.max_classes_per_detection(), info.num_classes());
     auto_init_if_empty(*_class_scores.info(), TensorInfo(info.use_regular_nms() ? TensorShape(_num_boxes) : TensorShape(_num_boxes * num_classes_per_box), 1, DataType::F32));
 
-    _input_scores_to_use = is_data_type_quantized(input_box_encoding->info()->data_type()) ? &_decoded_scores : _input_scores;
+    _input_scores_to_use = _dequantize_scores ? &_decoded_scores : _input_scores;
 
     // Manage intermediate buffers
     _memory_group.manage(&_decoded_boxes);
@@ -261,7 +262,7 @@ void CPPDetectionPostProcessLayer::run()
     DecodeCenterSizeBoxes(_input_box_encoding, _input_anchors, _info, &_decoded_boxes);
 
     // Decode scores if necessary
-    if(is_data_type_quantized(_input_box_encoding->info()->data_type()))
+    if(_dequantize_scores)
     {
         for(unsigned int idx_c = 0; idx_c < _num_classes_with_background; ++idx_c)
         {
@@ -365,7 +366,6 @@ void CPPDetectionPostProcessLayer::run()
 
         // Run Non-maxima Suppression
         _nms.run();
-
         std::vector<unsigned int> selected_indices;
         for(unsigned int i = 0; i < max_detections; ++i)
         {
@@ -384,4 +384,4 @@ void CPPDetectionPostProcessLayer::run()
                     num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection);
     }
 }
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp b/src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp
new file mode 100644
index 0000000000..d1d13432a1
--- /dev/null
+++ b/src/runtime/NEON/functions/NEDetectionPostProcessLayer.cpp
@@ -0,0 +1,98 @@
+/*
+ * 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/runtime/NEON/functions/NEDetectionPostProcessLayer.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Validate.h"
+#include "support/ToolchainSupport.h"
+
+#include <cstddef>
+#include <ios>
+#include <list>
+
+namespace arm_compute
+{
+NEDetectionPostProcessLayer::NEDetectionPostProcessLayer(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _dequantize(), _detection_post_process(), _decoded_scores(), _run_dequantize(false)
+{
+}
+
+void NEDetectionPostProcessLayer::configure(const ITensor *input_box_encoding, const ITensor *input_scores, const ITensor *input_anchors,
+                                            ITensor *output_boxes, ITensor *output_classes, ITensor *output_scores, ITensor *num_detection, DetectionPostProcessLayerInfo info)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, output_scores);
+    ARM_COMPUTE_ERROR_THROW_ON(NEDetectionPostProcessLayer::validate(input_box_encoding->info(), input_scores->info(), input_anchors->info(), output_boxes->info(), output_classes->info(),
+                                                                     output_scores->info(),
+                                                                     num_detection->info(), info));
+
+    const ITensor                *input_scores_to_use = input_scores;
+    DetectionPostProcessLayerInfo info_to_use         = info;
+    _run_dequantize                                   = is_data_type_quantized(input_box_encoding->info()->data_type());
+
+    if(_run_dequantize)
+    {
+        _memory_group.manage(&_decoded_scores);
+
+        _dequantize.configure(input_scores, &_decoded_scores);
+
+        input_scores_to_use = &_decoded_scores;
+
+        // Create a new info struct to avoid dequantizing in the CPP layer
+        std::array<float, 4> scales_values{ info.scale_value_y(), info.scale_value_x(), info.scale_value_h(), info.scale_value_w() };
+        DetectionPostProcessLayerInfo info_quantized(info.max_detections(), info.max_classes_per_detection(), info.nms_score_threshold(), info.iou_threshold(), info.num_classes(),
+                                                     scales_values, info.use_regular_nms(), info.detection_per_class(), false);
+        info_to_use = info_quantized;
+    }
+
+    _detection_post_process.configure(input_box_encoding, input_scores_to_use, input_anchors, output_boxes, output_classes, output_scores, num_detection, info_to_use);
+    _decoded_scores.allocator()->allocate();
+}
+
+Status NEDetectionPostProcessLayer::validate(const ITensorInfo *input_box_encoding, const ITensorInfo *input_scores, const ITensorInfo *input_anchors,
+                                             ITensorInfo *output_boxes, ITensorInfo *output_classes, ITensorInfo *output_scores, ITensorInfo *num_detection, DetectionPostProcessLayerInfo info)
+{
+    bool run_dequantize = is_data_type_quantized(input_box_encoding->data_type());
+    if(run_dequantize)
+    {
+        TensorInfo decoded_classes_info = input_scores->clone()->set_is_resizable(true).set_data_type(DataType::F32);
+        ARM_COMPUTE_RETURN_ON_ERROR(NEDequantizationLayer::validate(input_scores, &decoded_classes_info));
+    }
+    ARM_COMPUTE_RETURN_ON_ERROR(CPPDetectionPostProcessLayer::validate(input_box_encoding, input_scores, input_anchors, output_boxes, output_classes, output_scores, num_detection, info));
+
+    return Status{};
+}
+
+void NEDetectionPostProcessLayer::run()
+{
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    // Decode scores if necessary
+    if(_run_dequantize)
+    {
+        _dequantize.run();
+    }
+    _detection_post_process.run();
+}
+} // namespace arm_compute