commit d985378af0c9a4db6a483634dd40526cd4031dee
author Giuseppe Rossini <giuseppe.rossini@arm.com> Fri Oct 25 11:11:44 2019 +0100
committer Giuseppe Rossini <giuseppe.rossini@arm.com> Sat Oct 26 00:46:43 2019 +0000
tree 241ccb49da6b1908ec82138ab7e683d91e5908d8
parent 279814bfdc3e2ec3ed6c4e248356b4e0b2b2abc0

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>

src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp

diff --git a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
index 0addb0e..bc88f71 100644
--- a/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
+++ b/src/runtime/CPP/functions/CPPDetectionPostProcessLayer.cpp
@@ -42,7 +42,7 @@
 {
     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 @@
 
 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 @@
     _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 @@
     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 @@
     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 @@
 
         // 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 @@
                     num_output, max_detections, _output_boxes, _output_classes, _output_scores, _num_detection);
     }
 }
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute