IVGCVSW-4904 Refactor CpuRef PAD Workload

* Refactored templated workload creation
* Added int8_t unit tests

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I018b7f9f20496b5f9d7445901fe0d3dd04199cd0

8 files changed, 152 insertions, 164 deletions

diff --git a/src/backends/backendsCommon/test/layerTests/PadTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/PadTestImpl.cpp
index 120572ce29..405f4743bc 100644
--- a/src/backends/backendsCommon/test/layerTests/PadTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/PadTestImpl.cpp
@@ -525,3 +525,31 @@ LayerTestResult<armnn::BFloat16, 4> PadBFloat164dTest(
 {
     return Pad4dTestCommon<armnn::DataType::BFloat16>(workloadFactory, memoryManager, 0.0f, 0);
 }
+
+LayerTestResult<int8_t, 2> PadInt82dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return Pad2dTestCommon<armnn::DataType::QSymmS8>(workloadFactory, memoryManager, 1.0f, 0);
+}
+
+LayerTestResult<int8_t, 2> PadInt82dCustomPaddingTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return Pad2dTestCommon<armnn::DataType::QSymmS8>(workloadFactory, memoryManager, 1.0f, 0, 1.0f);
+}
+
+LayerTestResult<int8_t, 3> PadInt83dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return Pad3dTestCommon<armnn::DataType::QSymmS8>(workloadFactory, memoryManager, 1.0f, 0);
+}
+
+LayerTestResult<int8_t, 4> PadInt84dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return Pad4dTestCommon<armnn::DataType::QSymmS8>(workloadFactory, memoryManager, 1.0f, 0);
+}
diff --git a/src/backends/backendsCommon/test/layerTests/PadTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/PadTestImpl.hpp
index 34aa6c66a3..7b36455063 100644
--- a/src/backends/backendsCommon/test/layerTests/PadTestImpl.hpp
+++ b/src/backends/backendsCommon/test/layerTests/PadTestImpl.hpp
@@ -83,3 +83,19 @@ LayerTestResult<armnn::BFloat16, 3> PadBFloat163dTest(
 LayerTestResult<armnn::BFloat16, 4> PadBFloat164dTest(
     armnn::IWorkloadFactory& workloadFactory,
     const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<int8_t, 2> PadInt82dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<int8_t, 2> PadInt82dCustomPaddingTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<int8_t, 3> PadInt83dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<int8_t, 4> PadInt84dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
\ No newline at end of file
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp
index 4cd0b9e0e3..ba84de49da 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -456,23 +456,7 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateOutput(const OutputQueueDes
 std::unique_ptr<IWorkload> RefWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor,
                                                          const WorkloadInfo& info) const
 {
-    if (IsQSymmS16(info))
-    {
-        return std::make_unique<RefPadQSymm16Workload>(descriptor, info);
-    }
-    else if (IsFloat16(info))
-    {
-        return std::make_unique<RefPadFloat16Workload>(descriptor, info);
-    }
-    else if (IsBFloat16(info))
-    {
-        return std::make_unique<RefPadBFloat16Workload>(descriptor, info);
-    }
-    else if (IsQAsymmS8(info))
-    {
-        return std::make_unique<RefPadQAsymmS8Workload>(descriptor, info);
-    }
-    return MakeWorkload<RefPadFloat32Workload, RefPadQAsymm8Workload>(descriptor, info);
+    return std::make_unique<RefPadWorkload>(descriptor, info);
 }
 
 std::unique_ptr<IWorkload> RefWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor,
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index fc71eea713..eb2d0c55d9 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -1146,6 +1146,11 @@ ARMNN_AUTO_TEST_CASE(Pad2dQSymm16CustomPadding, Pad2dTestCommon<DataType::QSymmS
 ARMNN_AUTO_TEST_CASE(Pad3dQSymm16, Pad3dTestCommon<DataType::QSymmS16>, 2.0f, 0)
 ARMNN_AUTO_TEST_CASE(Pad4dQSymm16, Pad4dTestCommon<DataType::QSymmS16>, 2.0f, 0)
 
+ARMNN_AUTO_TEST_CASE(PadInt82d, PadInt82dTest)
+ARMNN_AUTO_TEST_CASE(PadInt82dCustomPadding, PadInt82dCustomPaddingTest)
+ARMNN_AUTO_TEST_CASE(PadInt83d, PadInt83dTest)
+ARMNN_AUTO_TEST_CASE(PadInt84d, PadInt84dTest)
+
 // Constant
 ARMNN_AUTO_TEST_CASE(Constant, ConstantTest)
 ARMNN_AUTO_TEST_CASE(ConstantUint8, ConstantUint8CustomQuantizationScaleAndOffsetTest)
diff --git a/src/backends/reference/workloads/Pad.cpp b/src/backends/reference/workloads/Pad.cpp
index 1b634145fc..1f8b674c3a 100644
--- a/src/backends/reference/workloads/Pad.cpp
+++ b/src/backends/reference/workloads/Pad.cpp
@@ -4,97 +4,125 @@
 //
 
 #include "Pad.hpp"
-#include <backendsCommon/WorkloadData.hpp>
-#include "TensorBufferArrayView.hpp"
+
+#include "BaseIterator.hpp"
+#include "Decoders.hpp"
 #include "Encoders.hpp"
 
-#include <boost/numeric/conversion/cast.hpp>
+#include <armnnUtils/TensorUtils.hpp>
+
 #include <cmath>
 #include <cstddef>
 #include <functional>
 #include <limits>
 #include <cassert>
 
+namespace
+{
+
+void FillOutputWithPadValue(armnn::Encoder<float>& outputData,
+                            const float padValue,
+                            const unsigned int numOutputElements)
+{
+    for (unsigned int i = 0; i < numOutputElements; ++i)
+    {
+        outputData[i];
+        outputData.Set(padValue);
+    }
+}
+
+} // anonymous namespace
+
 namespace armnn
 {
 
-template <typename T>
 void Pad(const TensorInfo& inputInfo,
          const TensorInfo& outputInfo,
-         std::vector<std::pair<unsigned int, unsigned int>> m_padList,
-         const T* inputData,
-         T* outData,
-         const float padValue)
+         const PadQueueDescriptor& data)
 {
+    auto padList  = data.m_Parameters.m_PadList;
+    auto padValue = data.m_Parameters.m_PadValue;
+
     unsigned int numOutputElements = outputInfo.GetNumElements();
 
     TensorShape outputShape = outputInfo.GetShape();
-    TensorShape inputShape = inputInfo.GetShape();
+    TensorShape inputShape  = inputInfo.GetShape();
 
     unsigned int numInputDimensions = inputShape.GetNumDimensions();
 
-    #ifndef NDEBUG
+#ifndef NDEBUG
 
     unsigned int numOutputDimensions = outputShape.GetNumDimensions();
     assert(numInputDimensions == numOutputDimensions);
 
-    #endif
+#endif
 
-    unsigned int inputBatches = 0;
+    unsigned int inputBatches  = 0;
     unsigned int inputChannels = 0;
-    unsigned int inputHeight = 0;
-    unsigned int inputWidth = 0;
+    unsigned int inputHeight   = 0;
+    unsigned int inputWidth    = 0;
 
     unsigned int outputChannels = 0;
-    unsigned int outputHeight = 0;
-    unsigned int outputWidth = 0;
+    unsigned int outputHeight   = 0;
+    unsigned int outputWidth    = 0;
 
-    T convertedPadValue = static_cast<T>(padValue);
+    auto inputData = MakeDecoder<float>(inputInfo, data.m_Inputs[0]->Map());
+    auto outData   = MakeEncoder<float>(outputInfo, data.m_Outputs[0]->Map());
 
-    for (unsigned int i = 0; i < numOutputElements; ++i)
+    // Fill the output tensor with Pad value first
+    if (outputInfo.IsQuantized())
     {
-       outData[i] = convertedPadValue;
+        // For Quantized types Pad Value should not be quantized with scale and offset of the tensor info
+        auto temporaryInfo = TensorInfo(outputInfo.GetShape(), outputInfo.GetDataType(), 1.0f, 0);
+        auto outputData = MakeEncoder<float>(temporaryInfo, data.m_Outputs[0]->Map());
+        FillOutputWithPadValue(*outputData, padValue, numOutputElements);
     }
+    else
+    {
+        FillOutputWithPadValue(*outData, padValue, numOutputElements);
+    }
+
+    Decoder<float>& input  = *inputData;
+    Encoder<float>& output = *outData;
 
     switch(numInputDimensions) {
 
         case 1:
-
             inputWidth = inputShape[0];
-
             for (unsigned int w = 0; w < inputWidth ; w++)
             {
-                outData[w+std::get<0>(m_padList[0])] = inputData[w];
+                input[w];
+                auto inputValue = input.Get();
+                auto outputIndex = w + std::get<0>(padList[0]);
+                output[outputIndex];
+                output.Set(inputValue);
             }
 
             break;
-
         case 2  :
-
             inputHeight = inputShape[0];
-            inputWidth = inputShape[1];
-            outputHeight = outputShape[0];
+            inputWidth  = inputShape[1];
             outputWidth = outputShape[1];
 
             for (unsigned int h = 0; h < inputHeight; h++)
             {
                 for (unsigned int w = 0; w < inputWidth ; w++)
                 {
-                    outData[(h+std::get<0>(m_padList[0]))*outputWidth
-                    + (w+std::get<0>(m_padList[1]))] = inputData[h * inputWidth + w];
+                    input[h * inputWidth + w];
+                    auto inputValue  = input.Get();
+                    auto outputIndex = (h + std::get<0>(padList[0])) * outputWidth + (w + std::get<0>(padList[1]));
+                    output[outputIndex];
+                    output.Set(inputValue);
                 }
             }
 
             break;
-
         case 3  :
-
             inputChannels = inputShape[0];
-            inputHeight = inputShape[1];
-            inputWidth = inputShape[2];
-            outputChannels = outputShape[0];
-            outputHeight = outputShape[1];
-            outputWidth = outputShape[2];
+            inputHeight   = inputShape[1];
+            inputWidth    = inputShape[2];
+            outputHeight  = outputShape[1];
+            outputWidth   = outputShape[2];
 
             for (unsigned int c = 0; c < inputChannels; c++)
             {
@@ -102,26 +130,26 @@ void Pad(const TensorInfo& inputInfo,
                 {
                     for (unsigned int w = 0; w < inputWidth ; w++)
                     {
-                        outData[(c+std::get<0>(m_padList[0]))*outputHeight*outputWidth
-                        + (h+std::get<0>(m_padList[1]))*outputWidth
-                        + (w+std::get<0>(m_padList[2]))] = inputData[c * inputHeight * inputWidth
-                                                                      + h * inputWidth
-                                                                      + w];
+                        input[c * inputHeight * inputWidth + h * inputWidth + w];
+                        auto inputValue  = input.Get();
+                        auto outputIndex = (c + std::get<0>(padList[0])) * outputHeight * outputWidth
+                                           + (h + std::get<0>(padList[1])) * outputWidth
+                                           + (w + std::get<0>(padList[2]));
+                        output[outputIndex];
+                        output.Set(inputValue);
                     }
                 }
             }
 
             break;
-
         case 4  :
-
-            inputBatches = inputShape[0];
-            inputChannels = inputShape[1];
-            inputHeight = inputShape[2];
-            inputWidth = inputShape[3];
+            inputBatches   = inputShape[0];
+            inputChannels  = inputShape[1];
+            inputHeight    = inputShape[2];
+            inputWidth     = inputShape[3];
             outputChannels = outputShape[1];
-            outputHeight = outputShape[2];
-            outputWidth = outputShape[3];
+            outputHeight   = outputShape[2];
+            outputWidth    = outputShape[3];
 
             for (unsigned int b = 0; b < inputBatches; b++)
             {
@@ -131,63 +159,27 @@ void Pad(const TensorInfo& inputInfo,
                     {
                         for (unsigned int w = 0; w < inputWidth ; w++)
                         {
-                            outData[(b+std::get<0>(m_padList[0])) * outputChannels * outputHeight * outputWidth
-                                   + (c+std::get<0>(m_padList[1])) * outputHeight * outputWidth
-                                   + (h+std::get<0>(m_padList[2])) * outputWidth
-                                   + (w+std::get<0>(m_padList[3]))] = inputData[b * inputChannels * inputHeight
-                                                                                * inputWidth
-                                                                             + c * inputHeight * inputWidth
-                                                                             + h * inputWidth
-                                                                             + w];
+                            input[b * inputChannels * inputHeight * inputWidth
+                                      + c * inputHeight * inputWidth
+                                      + h * inputWidth
+                                      + w];
+                            auto inputValue  = input.Get();
+                            auto outputIndex = (b + std::get<0>(padList[0]))
+                                               * outputChannels * outputHeight * outputWidth
+                                               + (c + std::get<0>(padList[1])) * outputHeight * outputWidth
+                                               + (h + std::get<0>(padList[2])) * outputWidth
+                                               + (w + std::get<0>(padList[3]));
+                            output[outputIndex];
+                            output.Set(inputValue);
                         }
                     }
                 }
             }
 
             break;
-
         default :
-
             break;
     }
 }
 
-template void Pad<BFloat16>(const TensorInfo& inputInfo,
-                            const TensorInfo& outputInfo,
-                            std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                            const BFloat16* inputData,
-                            BFloat16* outData,
-                            const float padValue);
-
-template void Pad<float>(const TensorInfo& inputInfo,
-                         const TensorInfo& outputInfo,
-                         std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                         const float* inputData,
-                         float* outData,
-                         const float padValue);
-template void Pad<Half>(const TensorInfo& inputInfo,
-                        const TensorInfo& outputInfo,
-                        std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                        const Half* inputData,
-                        Half* outData,
-                        const float padValue);
-template void Pad<uint8_t>(const TensorInfo& inputInfo,
-                           const TensorInfo& outputInfo,
-                           std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                           const uint8_t* inputData,
-                           uint8_t* outData,
-                           const float padValue);
-template void Pad<int8_t>(const TensorInfo& inputInfo,
-                          const TensorInfo& outputInfo,
-                          std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                          const int8_t* inputData,
-                          int8_t* outData,
-                          const float padValue);
-template void Pad<int16_t>(const TensorInfo& inputInfo,
-                           const TensorInfo& outputInfo,
-                           std::vector<std::pair<unsigned int, unsigned int>> m_PadList,
-                           const int16_t* inputData,
-                           int16_t* outData,
-                           const float padValue);
-
 } //namespace armnn
\ No newline at end of file
diff --git a/src/backends/reference/workloads/Pad.hpp b/src/backends/reference/workloads/Pad.hpp
index 429718596e..e7be44e88c 100644
--- a/src/backends/reference/workloads/Pad.hpp
+++ b/src/backends/reference/workloads/Pad.hpp
@@ -5,18 +5,16 @@
 
 #pragma once
 
-#include "armnn/DescriptorsFwd.hpp"
 #include "armnn/Tensor.hpp"
 
-#include <vector>
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
 
 namespace armnn
 {
-template <typename T>
+
 void Pad(const TensorInfo& inputInfo,
          const TensorInfo& outputInfo,
-         std::vector<std::pair<unsigned int, unsigned int>> m_padList,
-         const T* inputData,
-         T* outData,
-         const float padValue);
+         const PadQueueDescriptor& data);
+
 } //namespace armnn
diff --git a/src/backends/reference/workloads/RefPadWorkload.cpp b/src/backends/reference/workloads/RefPadWorkload.cpp
index 6f82d5ffdb..af22c31001 100644
--- a/src/backends/reference/workloads/RefPadWorkload.cpp
+++ b/src/backends/reference/workloads/RefPadWorkload.cpp
@@ -6,38 +6,22 @@
 #include "RefPadWorkload.hpp"
 
 #include "Pad.hpp"
-#include "RefWorkloadUtils.hpp"
-
 #include "Profiling.hpp"
-
-#include <ResolveType.hpp>
-
-#include <vector>
+#include "RefWorkloadUtils.hpp"
 
 namespace armnn
 {
 
-template <armnn::DataType DataType>
-void RefPadWorkload<DataType>::Execute() const
+void RefPadWorkload::Execute() const
 {
-    using T = ResolveType<DataType>;
-
     ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefPadWorkload_Execute");
 
-    const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
+    const TensorInfo& inputInfo  = GetTensorInfo(m_Data.m_Inputs[0]);
     const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
 
-    const T* inputData = GetInputTensorData<T>(0, m_Data);
-    T* outputData = GetOutputTensorData<T>(0, m_Data);
-
-    Pad(inputInfo, outputInfo, m_Data.m_Parameters.m_PadList, inputData, outputData, m_Data.m_Parameters.m_PadValue);
+    armnn::Pad(inputInfo,
+               outputInfo,
+               m_Data);
 }
 
-template class RefPadWorkload<DataType::BFloat16>;
-template class RefPadWorkload<DataType::Float32>;
-template class RefPadWorkload<DataType::Float16>;
-template class RefPadWorkload<DataType::QAsymmS8>;
-template class RefPadWorkload<DataType::QAsymmU8>;
-template class RefPadWorkload<DataType::QSymmS16>;
-
 } //namespace armnn
\ No newline at end of file
diff --git a/src/backends/reference/workloads/RefPadWorkload.hpp b/src/backends/reference/workloads/RefPadWorkload.hpp
index 74dcab1967..0b8379a60f 100644
--- a/src/backends/reference/workloads/RefPadWorkload.hpp
+++ b/src/backends/reference/workloads/RefPadWorkload.hpp
@@ -8,33 +8,14 @@
 #include <backendsCommon/Workload.hpp>
 #include <backendsCommon/WorkloadData.hpp>
 
-#include <armnn/TypesUtils.hpp>
-
 namespace armnn
 {
 
-template <armnn::DataType DataType>
-class RefPadWorkload : public TypedWorkload<PadQueueDescriptor, DataType>
+class RefPadWorkload : public BaseWorkload<PadQueueDescriptor>
 {
 public:
-
-    static const std::string& GetName()
-    {
-        static const std::string name = std::string("RefPad") + GetDataTypeName(DataType) + "Workload";
-        return name;
-    }
-
-    using TypedWorkload<PadQueueDescriptor, DataType>::m_Data;
-    using TypedWorkload<PadQueueDescriptor, DataType>::TypedWorkload;
-
-    void Execute() const override;
+    using BaseWorkload<PadQueueDescriptor>::BaseWorkload;
+    virtual void Execute() const override;
 };
 
-using RefPadBFloat16Workload = RefPadWorkload<DataType::BFloat16>;
-using RefPadFloat32Workload =  RefPadWorkload<DataType::Float32>;
-using RefPadFloat16Workload =  RefPadWorkload<DataType::Float16>;
-using RefPadQAsymmS8Workload = RefPadWorkload<DataType::QAsymmS8>;
-using RefPadQAsymm8Workload =  RefPadWorkload<DataType::QAsymmU8>;
-using RefPadQSymm16Workload =  RefPadWorkload<DataType::QSymmS16>;
-
 } //namespace armnn