IVGCVSW-2971 Support QSymm16 for DetectionPostProcess workloads

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I8af45afe851a9ccbf8bce54727147fcd52ac9a1f

9 files changed, 161 insertions, 164 deletions

diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index cdca22da31..25d4b28366 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -60,10 +60,8 @@ list(APPEND armnnRefBackendWorkloads_sources
     RefDepthwiseConvolution2dWorkload.hpp
     RefDequantizeWorkload.cpp
     RefDequantizeWorkload.hpp
-    RefDetectionPostProcessUint8Workload.cpp
-    RefDetectionPostProcessUint8Workload.hpp
-    RefDetectionPostProcessFloat32Workload.cpp
-    RefDetectionPostProcessFloat32Workload.hpp
+    RefDetectionPostProcessWorkload.cpp
+    RefDetectionPostProcessWorkload.hpp
     RefFakeQuantizationFloat32Workload.cpp
     RefFakeQuantizationFloat32Workload.hpp
     RefFloorWorkload.cpp
diff --git a/src/backends/reference/workloads/DetectionPostProcess.cpp b/src/backends/reference/workloads/DetectionPostProcess.cpp
index d3790f22ab..d475dd8ac0 100644
--- a/src/backends/reference/workloads/DetectionPostProcess.cpp
+++ b/src/backends/reference/workloads/DetectionPostProcess.cpp
@@ -13,7 +13,7 @@
 #include <algorithm>
 #include <numeric>
 
-namespace
+namespace armnn
 {
 
 std::vector<unsigned int> GenerateRangeK(unsigned int k)
@@ -48,9 +48,12 @@ float IntersectionOverUnion(const float* boxI, const float* boxJ)
     return areaIntersection / areaUnion;
 }
 
-std::vector<unsigned int> NonMaxSuppression(unsigned int numBoxes, const std::vector<float>& boxCorners,
-                                            const std::vector<float>& scores, float nmsScoreThreshold,
-                                            unsigned int maxDetection, float nmsIouThreshold)
+std::vector<unsigned int> NonMaxSuppression(unsigned int numBoxes,
+                                            const std::vector<float>& boxCorners,
+                                            const std::vector<float>& scores,
+                                            float nmsScoreThreshold,
+                                            unsigned int maxDetection,
+                                            float nmsIouThreshold)
 {
     // Select boxes that have scores above a given threshold.
     std::vector<float> scoresAboveThreshold;
@@ -67,7 +70,7 @@ std::vector<unsigned int> NonMaxSuppression(unsigned int numBoxes, const std::ve
     // Sort the indices based on scores.
     unsigned int numAboveThreshold = boost::numeric_cast<unsigned int>(scoresAboveThreshold.size());
     std::vector<unsigned int> sortedIndices = GenerateRangeK(numAboveThreshold);
-    TopKSort(numAboveThreshold,sortedIndices.data(), scoresAboveThreshold.data(), numAboveThreshold);
+    TopKSort(numAboveThreshold, sortedIndices.data(), scoresAboveThreshold.data(), numAboveThreshold);
 
     // Number of output cannot be more than max detections specified in the option.
     unsigned int numOutput = std::min(maxDetection, numAboveThreshold);
@@ -98,10 +101,17 @@ std::vector<unsigned int> NonMaxSuppression(unsigned int numBoxes, const std::ve
     return outputIndices;
 }
 
-void AllocateOutputData(unsigned int numOutput, unsigned int numSelected, const std::vector<float>& boxCorners,
-                        const std::vector<unsigned int>& outputIndices, const std::vector<unsigned int>& selectedBoxes,
-                        const std::vector<unsigned int>& selectedClasses, const std::vector<float>& selectedScores,
-                        float* detectionBoxes, float* detectionScores, float* detectionClasses, float* numDetections)
+void AllocateOutputData(unsigned int numOutput,
+                        unsigned int numSelected,
+                        const std::vector<float>& boxCorners,
+                        const std::vector<unsigned int>& outputIndices,
+                        const std::vector<unsigned int>& selectedBoxes,
+                        const std::vector<unsigned int>& selectedClasses,
+                        const std::vector<float>& selectedScores,
+                        float* detectionBoxes,
+                        float* detectionScores,
+                        float* detectionClasses,
+                        float* numDetections)
 {
     for (unsigned int i = 0; i < numOutput; ++i)
         {
@@ -129,11 +139,6 @@ void AllocateOutputData(unsigned int numOutput, unsigned int numSelected, const
         numDetections[0] = boost::numeric_cast<float>(numSelected);
 }
 
-} // anonymous namespace
-
-namespace armnn
-{
-
 void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
                           const TensorInfo& scoresInfo,
                           const TensorInfo& anchorsInfo,
@@ -142,9 +147,9 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
                           const TensorInfo& detectionScoresInfo,
                           const TensorInfo& numDetectionsInfo,
                           const DetectionPostProcessDescriptor& desc,
-                          const float* boxEncodings,
-                          const float* scores,
-                          const float* anchors,
+                          Decoder<float>& boxEncodings,
+                          Decoder<float>& scores,
+                          Decoder<float>& anchors,
                           float* detectionBoxes,
                           float* detectionClasses,
                           float* detectionScores,
@@ -153,17 +158,51 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
     // Transform center-size format which is (ycenter, xcenter, height, width) to box-corner format,
     // which represents the lower left corner and the upper right corner (ymin, xmin, ymax, xmax)
     std::vector<float> boxCorners(boxEncodingsInfo.GetNumElements());
-    unsigned int numBoxes = boxEncodingsInfo.GetShape()[1];
+
+    const unsigned int numBoxes  = boxEncodingsInfo.GetShape()[1];
+    const unsigned int numScores = scoresInfo.GetNumElements();
+
     for (unsigned int i = 0; i < numBoxes; ++i)
     {
+        // Y
+        float boxEncodingY = boxEncodings.Get();
+        float anchorY      = anchors.Get();
+
+        ++boxEncodings;
+        ++anchors;
+
+        // X
+        float boxEncodingX = boxEncodings.Get();
+        float anchorX      = anchors.Get();
+
+        ++boxEncodings;
+        ++anchors;
+
+        // H
+        float boxEncodingH = boxEncodings.Get();
+        float anchorH      = anchors.Get();
+
+        ++boxEncodings;
+        ++anchors;
+
+        // W
+        float boxEncodingW = boxEncodings.Get();
+        float anchorW      = anchors.Get();
+
+        ++boxEncodings;
+        ++anchors;
+
+        float yCentre = boxEncodingY / desc.m_ScaleY * anchorH + anchorY;
+        float xCentre = boxEncodingX / desc.m_ScaleX * anchorW + anchorX;
+
+        float halfH = 0.5f * expf(boxEncodingH / desc.m_ScaleH) * anchorH;
+        float halfW = 0.5f * expf(boxEncodingW / desc.m_ScaleW) * anchorW;
+
         unsigned int indexY = i * 4;
         unsigned int indexX = indexY + 1;
         unsigned int indexH = indexX + 1;
         unsigned int indexW = indexH + 1;
-        float yCentre = boxEncodings[indexY] / desc.m_ScaleY * anchors[indexH] + anchors[indexY];
-        float xCentre = boxEncodings[indexX] / desc.m_ScaleX * anchors[indexW] + anchors[indexX];
-        float halfH = 0.5f * expf(boxEncodings[indexH] / desc.m_ScaleH) * anchors[indexH];
-        float halfW = 0.5f * expf(boxEncodings[indexW] / desc.m_ScaleW) * anchors[indexW];
+
         // ymin
         boxCorners[indexY] = yCentre - halfH;
         // xmin
@@ -179,14 +218,29 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
 
     unsigned int numClassesWithBg = desc.m_NumClasses + 1;
 
+    // Decode scores
+    std::vector<float> decodedScores;
+    decodedScores.reserve(numScores);
+
+    for (unsigned int i = 0u; i < numScores; ++i)
+    {
+        decodedScores.emplace_back(scores.Get());
+        ++scores;
+    }
+
     // Perform Non Max Suppression.
     if (desc.m_UseRegularNms)
     {
         // Perform Regular NMS.
         // For each class, perform NMS and select max detection numbers of the highest score across all classes.
         std::vector<float> classScores(numBoxes);
-        std::vector<unsigned int>selectedBoxesAfterNms;
+
+        std::vector<unsigned int> selectedBoxesAfterNms;
+        selectedBoxesAfterNms.reserve(numBoxes);
+
         std::vector<float> selectedScoresAfterNms;
+        selectedBoxesAfterNms.reserve(numScores);
+
         std::vector<unsigned int> selectedClasses;
 
         for (unsigned int c = 0; c < desc.m_NumClasses; ++c)
@@ -194,9 +248,11 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
             // For each boxes, get scores of the boxes for the class c.
             for (unsigned int i = 0; i < numBoxes; ++i)
             {
-                classScores[i] = scores[i * numClassesWithBg + c + 1];
+                classScores[i] = decodedScores[i * numClassesWithBg + c + 1];
             }
-            std::vector<unsigned int> selectedIndices = NonMaxSuppression(numBoxes, boxCorners, classScores,
+            std::vector<unsigned int> selectedIndices = NonMaxSuppression(numBoxes,
+                                                                          boxCorners,
+                                                                          classScores,
                                                                           desc.m_NmsScoreThreshold,
                                                                           desc.m_DetectionsPerClass,
                                                                           desc.m_NmsIouThreshold);
@@ -237,11 +293,12 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
 
             // Get the max scores of the box.
             std::vector<unsigned int> maxScoreIndices = GenerateRangeK(desc.m_NumClasses);
-            TopKSort(numClassesPerBox, maxScoreIndices.data(), scores + scoreIndex, desc.m_NumClasses);
+            TopKSort(numClassesPerBox, maxScoreIndices.data(),
+                decodedScores.data() + scoreIndex, desc.m_NumClasses);
 
             for (unsigned int i = 0; i < numClassesPerBox; ++i)
             {
-                maxScores.push_back(scores[scoreIndex + maxScoreIndices[i]]);
+                maxScores.push_back(decodedScores[scoreIndex + maxScoreIndices[i]]);
                 maxScoreClasses.push_back(maxScoreIndices[i]);
                 boxIndices.push_back(box);
             }
diff --git a/src/backends/reference/workloads/DetectionPostProcess.hpp b/src/backends/reference/workloads/DetectionPostProcess.hpp
index 06e9e15781..8700a53317 100644
--- a/src/backends/reference/workloads/DetectionPostProcess.hpp
+++ b/src/backends/reference/workloads/DetectionPostProcess.hpp
@@ -7,6 +7,10 @@
 #include "armnn/Tensor.hpp"
 #include "armnn/Descriptors.hpp"
 
+#include "Decoders.hpp"
+
+#include <vector>
+
 namespace armnn
 {
 
@@ -18,12 +22,26 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
                           const TensorInfo& detectionScoresInfo,
                           const TensorInfo& numDetectionsInfo,
                           const DetectionPostProcessDescriptor& desc,
-                          const float* boxEncodings,
-                          const float* scores,
-                          const float* anchors,
+                          Decoder<float>& boxEncodings,
+                          Decoder<float>& scores,
+                          Decoder<float>& anchors,
                           float* detectionBoxes,
                           float* detectionClasses,
                           float* detectionScores,
                           float* numDetections);
 
+void TopKSort(unsigned int k,
+              unsigned int* indices,
+              const float* values,
+              unsigned int numElement);
+
+float IntersectionOverUnion(const float* boxI, const float* boxJ);
+
+std::vector<unsigned int> NonMaxSuppression(unsigned int numBoxes,
+                                            const std::vector<float>& boxCorners,
+                                            const std::vector<float>& scores,
+                                            float nmsScoreThreshold,
+                                            unsigned int maxDetection,
+                                            float nmsIouThreshold);
+
 } // namespace armnn
diff --git a/src/backends/reference/workloads/RefDetectionPostProcessFloat32Workload.cpp b/src/backends/reference/workloads/RefDetectionPostProcessFloat32Workload.cpp
deleted file mode 100644
index ddab046f9c..0000000000
--- a/src/backends/reference/workloads/RefDetectionPostProcessFloat32Workload.cpp
+++ /dev/null
@@ -1,48 +0,0 @@
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "RefDetectionPostProcessFloat32Workload.hpp"
-
-#include "DetectionPostProcess.hpp"
-#include "Profiling.hpp"
-#include "RefWorkloadUtils.hpp"
-
-namespace armnn
-{
-
-RefDetectionPostProcessFloat32Workload::RefDetectionPostProcessFloat32Workload(
-        const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info)
-        : Float32Workload<DetectionPostProcessQueueDescriptor>(descriptor, info),
-          m_Anchors(std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Anchors))) {}
-
-void RefDetectionPostProcessFloat32Workload::Execute() const
-{
-    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefDetectionPostProcessUint8Workload_Execute");
-
-    const TensorInfo& boxEncodingsInfo = GetTensorInfo(m_Data.m_Inputs[0]);
-    const TensorInfo& scoresInfo = GetTensorInfo(m_Data.m_Inputs[1]);
-    const TensorInfo& anchorsInfo = GetTensorInfo(m_Anchors.get());
-    const TensorInfo& detectionBoxesInfo = GetTensorInfo(m_Data.m_Outputs[0]);
-    const TensorInfo& detectionClassesInfo = GetTensorInfo(m_Data.m_Outputs[1]);
-    const TensorInfo& detectionScoresInfo = GetTensorInfo(m_Data.m_Outputs[2]);
-    const TensorInfo& numDetectionsInfo = GetTensorInfo(m_Data.m_Outputs[3]);
-
-    const float* boxEncodings = GetInputTensorDataFloat(0, m_Data);
-    const float* scores = GetInputTensorDataFloat(1, m_Data);
-    const float* anchors = m_Anchors->GetConstTensor<float>();
-
-    float* detectionBoxes = GetOutputTensorData<float>(0, m_Data);
-    float* detectionClasses = GetOutputTensorData<float>(1, m_Data);
-    float* detectionScores = GetOutputTensorData<float>(2, m_Data);
-    float* numDetections = GetOutputTensorData<float>(3, m_Data);
-
-    DetectionPostProcess(boxEncodingsInfo, scoresInfo, anchorsInfo,
-                         detectionBoxesInfo, detectionClassesInfo,
-                         detectionScoresInfo, numDetectionsInfo, m_Data.m_Parameters,
-                         boxEncodings, scores, anchors, detectionBoxes,
-                         detectionClasses, detectionScores, numDetections);
-}
-
-} //namespace armnn
diff --git a/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.cpp b/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.cpp
deleted file mode 100644
index ccdaf87c9a..0000000000
--- a/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.cpp
+++ /dev/null
@@ -1,52 +0,0 @@
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "RefDetectionPostProcessUint8Workload.hpp"
-
-#include "DetectionPostProcess.hpp"
-#include "Profiling.hpp"
-#include "RefWorkloadUtils.hpp"
-
-namespace armnn
-{
-
-RefDetectionPostProcessUint8Workload::RefDetectionPostProcessUint8Workload(
-        const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info)
-        : Uint8ToFloat32Workload<DetectionPostProcessQueueDescriptor>(descriptor, info),
-          m_Anchors(std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Anchors))) {}
-
-void RefDetectionPostProcessUint8Workload::Execute() const
-{
-    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefDetectionPostProcessUint8Workload_Execute");
-
-    const TensorInfo& boxEncodingsInfo = GetTensorInfo(m_Data.m_Inputs[0]);
-    const TensorInfo& scoresInfo = GetTensorInfo(m_Data.m_Inputs[1]);
-    const TensorInfo& anchorsInfo = GetTensorInfo(m_Anchors.get());
-    const TensorInfo& detectionBoxesInfo = GetTensorInfo(m_Data.m_Outputs[0]);
-    const TensorInfo& detectionClassesInfo = GetTensorInfo(m_Data.m_Outputs[1]);
-    const TensorInfo& detectionScoresInfo = GetTensorInfo(m_Data.m_Outputs[2]);
-    const TensorInfo& numDetectionsInfo = GetTensorInfo(m_Data.m_Outputs[3]);
-
-    const uint8_t* boxEncodingsData = GetInputTensorDataU8(0, m_Data);
-    const uint8_t* scoresData = GetInputTensorDataU8(1, m_Data);
-    const uint8_t* anchorsData = m_Anchors->GetConstTensor<uint8_t>();
-
-    auto boxEncodings = Dequantize(boxEncodingsData, boxEncodingsInfo);
-    auto scores = Dequantize(scoresData, scoresInfo);
-    auto anchors = Dequantize(anchorsData, anchorsInfo);
-
-    float* detectionBoxes = GetOutputTensorData<float>(0, m_Data);
-    float* detectionClasses = GetOutputTensorData<float>(1, m_Data);
-    float* detectionScores = GetOutputTensorData<float>(2, m_Data);
-    float* numDetections = GetOutputTensorData<float>(3, m_Data);
-
-    DetectionPostProcess(boxEncodingsInfo, scoresInfo, anchorsInfo,
-                         detectionBoxesInfo, detectionClassesInfo,
-                         detectionScoresInfo, numDetectionsInfo, m_Data.m_Parameters,
-                         boxEncodings.data(), scores.data(), anchors.data(),
-                         detectionBoxes, detectionClasses, detectionScores, numDetections);
-}
-
-} //namespace armnn
diff --git a/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.hpp b/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.hpp
deleted file mode 100644
index 91590f57bd..0000000000
--- a/src/backends/reference/workloads/RefDetectionPostProcessUint8Workload.hpp
+++ /dev/null
@@ -1,25 +0,0 @@
-//
-// Copyright © 2017 Arm Ltd. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <backendsCommon/Workload.hpp>
-#include <backendsCommon/WorkloadData.hpp>
-
-namespace armnn
-{
-
-class RefDetectionPostProcessUint8Workload : public Uint8ToFloat32Workload<DetectionPostProcessQueueDescriptor>
-{
-public:
-    explicit RefDetectionPostProcessUint8Workload(const DetectionPostProcessQueueDescriptor& descriptor,
-                                                  const WorkloadInfo& info);
-    virtual void Execute() const override;
-
-private:
-    std::unique_ptr<ScopedCpuTensorHandle> m_Anchors;
-};
-
-} //namespace armnn
diff --git a/src/backends/reference/workloads/RefDetectionPostProcessWorkload.cpp b/src/backends/reference/workloads/RefDetectionPostProcessWorkload.cpp
new file mode 100644
index 0000000000..db24cc53e4
--- /dev/null
+++ b/src/backends/reference/workloads/RefDetectionPostProcessWorkload.cpp
@@ -0,0 +1,50 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefDetectionPostProcessWorkload.hpp"
+
+#include "Decoders.hpp"
+#include "DetectionPostProcess.hpp"
+#include "Profiling.hpp"
+#include "RefWorkloadUtils.hpp"
+
+namespace armnn
+{
+
+RefDetectionPostProcessWorkload::RefDetectionPostProcessWorkload(
+        const DetectionPostProcessQueueDescriptor& descriptor, const WorkloadInfo& info)
+        : BaseWorkload<DetectionPostProcessQueueDescriptor>(descriptor, info),
+          m_Anchors(std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Anchors))) {}
+
+void RefDetectionPostProcessWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefDetectionPostProcessWorkload_Execute");
+
+    const TensorInfo& boxEncodingsInfo = GetTensorInfo(m_Data.m_Inputs[0]);
+    const TensorInfo& scoresInfo       = GetTensorInfo(m_Data.m_Inputs[1]);
+    const TensorInfo& anchorsInfo      = GetTensorInfo(m_Anchors.get());
+
+    const TensorInfo& detectionBoxesInfo   = GetTensorInfo(m_Data.m_Outputs[0]);
+    const TensorInfo& detectionClassesInfo = GetTensorInfo(m_Data.m_Outputs[1]);
+    const TensorInfo& detectionScoresInfo  = GetTensorInfo(m_Data.m_Outputs[2]);
+    const TensorInfo& numDetectionsInfo    = GetTensorInfo(m_Data.m_Outputs[3]);
+
+    auto boxEncodings = MakeDecoder<float>(boxEncodingsInfo, m_Data.m_Inputs[0]->Map());
+    auto scores       = MakeDecoder<float>(scoresInfo, m_Data.m_Inputs[1]->Map());
+    auto anchors      = MakeDecoder<float>(anchorsInfo, m_Anchors->Map(false));
+
+    float* detectionBoxes   = GetOutputTensorData<float>(0, m_Data);
+    float* detectionClasses = GetOutputTensorData<float>(1, m_Data);
+    float* detectionScores  = GetOutputTensorData<float>(2, m_Data);
+    float* numDetections    = GetOutputTensorData<float>(3, m_Data);
+
+    DetectionPostProcess(boxEncodingsInfo, scoresInfo, anchorsInfo,
+                         detectionBoxesInfo, detectionClassesInfo,
+                         detectionScoresInfo, numDetectionsInfo, m_Data.m_Parameters,
+                         *boxEncodings, *scores, *anchors, detectionBoxes,
+                         detectionClasses, detectionScores, numDetections);
+}
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefDetectionPostProcessFloat32Workload.hpp b/src/backends/reference/workloads/RefDetectionPostProcessWorkload.hpp
index 9f2a697ada..799d0c6219 100644
--- a/src/backends/reference/workloads/RefDetectionPostProcessFloat32Workload.hpp
+++ b/src/backends/reference/workloads/RefDetectionPostProcessWorkload.hpp
@@ -11,11 +11,11 @@
 namespace armnn
 {
 
-class RefDetectionPostProcessFloat32Workload : public Float32Workload<DetectionPostProcessQueueDescriptor>
+class RefDetectionPostProcessWorkload : public BaseWorkload<DetectionPostProcessQueueDescriptor>
 {
 public:
-    explicit RefDetectionPostProcessFloat32Workload(const DetectionPostProcessQueueDescriptor& descriptor,
-                                                    const WorkloadInfo& info);
+    explicit RefDetectionPostProcessWorkload(const DetectionPostProcessQueueDescriptor& descriptor,
+                                             const WorkloadInfo& info);
     virtual void Execute() const override;
 
 private:
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index 7ccd4efc54..8d99b69685 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -31,8 +31,7 @@
 #include "RefResizeBilinearFloat32Workload.hpp"
 #include "ResizeBilinear.hpp"
 #include "RefNormalizationFloat32Workload.hpp"
-#include "RefDetectionPostProcessFloat32Workload.hpp"
-#include "RefDetectionPostProcessUint8Workload.hpp"
+#include "RefDetectionPostProcessWorkload.hpp"
 #include "RefPooling2dUint8Workload.hpp"
 #include "BatchNormImpl.hpp"
 #include "Activation.hpp"