IVGCVSW-3705 Add Channel Shuffle Front end and Ref Implementation

* Add front end
* Add reference workload
* Add unit tests
* Add Serializer and Deserializer
* Update ArmNN Versioning

Signed-off-by: Simon Obute <simon.obute@arm.com>
Change-Id: I9ac1f953af3974382eac8e8d62d794d2344e8f47

12 files changed, 361 insertions, 0 deletions

diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index 138d45367e..2753c927d5 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -95,6 +95,14 @@ bool LayerSupportBase::IsCastSupported(const TensorInfo&, //input
     return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
 }
 
+bool LayerSupportBase::IsChannelShuffleSupported(const TensorInfo&, //input
+                                                 const TensorInfo&, //output
+                                                 const ChannelShuffleDescriptor&, //descriptor
+                                                 Optional<std::string &> reasonIfUnsupported) const
+{
+    return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
 bool LayerSupportBase::IsComparisonSupported(const TensorInfo&, // input0
                                              const TensorInfo&, // input1
                                              const TensorInfo&, // output
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index 533a2c6bdd..cc68a220e2 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -51,6 +51,11 @@ public:
                          const TensorInfo& output,
                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsChannelShuffleSupported(const TensorInfo& input,
+                                   const TensorInfo& output,
+                                   const ChannelShuffleDescriptor& descriptor,
+                                   Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsComparisonSupported(const TensorInfo& input0,
                                const TensorInfo& input1,
                                const TensorInfo& output,
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index d87f858601..a6def847fa 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -2971,6 +2971,19 @@ void TransposeQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
     ValidateTensorDataTypesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output");
 }
 
+void ChannelShuffleQueueDescriptor::Validate(const WorkloadInfo &workloadInfo) const
+{
+    const std::string descriptorName{"TransposeQueueDescriptor"};
+
+    ValidateNumInputs(workloadInfo, descriptorName, 1);
+    ValidateNumOutputs(workloadInfo, descriptorName, 1);
+
+    const TensorInfo& inputTensorInfo = workloadInfo.m_InputTensorInfos[0];
+    const TensorInfo& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0];
+
+    ValidateTensorDataTypesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output");
+}
+
 void QLstmQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
 {
     const std::string descriptorName{"QLstmQueueDescriptor"};
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 78da00be5d..b90c29c1b4 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -747,4 +747,9 @@ struct UnidirectionalSequenceLstmQueueDescriptor : QueueDescriptorWithParameters
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct ChannelShuffleQueueDescriptor : QueueDescriptorWithParameters<ChannelShuffleDescriptor>
+{
+    void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
 } // namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 3f5972dab6..00263eca04 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -152,6 +152,21 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId,
                                                         reason);
             break;
         }
+        case LayerType::ChannelShuffle:
+        {
+            auto cLayer = PolymorphicDowncast<const ChannelShuffleLayer*>(&layer);
+
+            const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            const TensorInfo& output = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+
+            const ChannelShuffleDescriptor descriptor = cLayer->GetParameters();
+
+            result = layerSupportObject.IsChannelShuffleSupported(OverrideDataType(input, dataType),
+                                                                  OverrideDataType(output, dataType),
+                                                                  descriptor,
+                                                                  reason);
+            break;
+        }
         case LayerType::Comparison:
         {
             auto cLayer = PolymorphicDowncast<const ComparisonLayer*>(&layer);
@@ -1501,6 +1516,12 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateCast(const CastQueueDescripto
     return std::unique_ptr<IWorkload>();
 }
 
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateChannelShuffle(const ChannelShuffleQueueDescriptor& /*descriptor*/,
+                                                                  const WorkloadInfo& /*info*/) const
+{
+    return std::unique_ptr<IWorkload>();
+}
+
 std::unique_ptr<IWorkload> IWorkloadFactory::CreateComparison(const ComparisonQueueDescriptor& /*descriptor*/,
                                                               const WorkloadInfo& /*info*/) const
 {
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index efb8d99fa0..e84657ea13 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -90,6 +90,9 @@ public:
     virtual std::unique_ptr<IWorkload> CreateCast(const CastQueueDescriptor& descriptor,
                                                   const WorkloadInfo& Info) const;
 
+    virtual std::unique_ptr<IWorkload> CreateChannelShuffle(const ChannelShuffleQueueDescriptor& descriptor,
+                                                            const WorkloadInfo& info) const;
+
     virtual std::unique_ptr<IWorkload> CreateComparison(const ComparisonQueueDescriptor& descriptor,
                                                         const WorkloadInfo& Info) const;
 
diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 7ebc9975c3..7d3558c804 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -45,6 +45,7 @@ COMMON_TEST_SOURCES := \
     test/layerTests/ArgMinMaxTestImpl.cpp \
     test/layerTests/BatchNormalizationTestImpl.cpp \
     test/layerTests/CastTestImpl.cpp \
+    test/layerTests/ChannelShuffleTestImpl.cpp \
     test/layerTests/ComparisonTestImpl.cpp \
     test/layerTests/ConcatTestImpl.cpp \
     test/layerTests/ConstantTestImpl.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index c9bc5e74b8..292ec0efd2 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -64,6 +64,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     layerTests/BatchToSpaceNdTestImpl.hpp
     layerTests/CastTestImpl.cpp
     layerTests/CastTestImpl.hpp
+    layerTests/ChannelShuffleTestImpl.cpp
+    layerTests/ChannelShuffleTestImpl.hpp
     layerTests/ComparisonTestImpl.cpp
     layerTests/ComparisonTestImpl.hpp
     layerTests/ConcatTestImpl.cpp
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index 21b33d297b..c2d21842f2 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -620,6 +620,8 @@ DECLARE_LAYER_POLICY_2_PARAM(BatchToSpaceNd)
 
 DECLARE_LAYER_POLICY_1_PARAM(Cast)
 
+DECLARE_LAYER_POLICY_2_PARAM(ChannelShuffle)
+
 DECLARE_LAYER_POLICY_2_PARAM(Comparison)
 
 DECLARE_LAYER_POLICY_2_PARAM(Concat)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 0690637500..9f1fa88b16 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -12,6 +12,7 @@
 #include <backendsCommon/test/layerTests/BatchNormalizationTestImpl.hpp>
 #include <backendsCommon/test/layerTests/BatchToSpaceNdTestImpl.hpp>
 #include <backendsCommon/test/layerTests/CastTestImpl.hpp>
+#include <backendsCommon/test/layerTests/ChannelShuffleTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ComparisonTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConcatTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp>
diff --git a/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.cpp
new file mode 100644
index 0000000000..46ee7d960e
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.cpp
@@ -0,0 +1,269 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ChannelShuffleTestImpl.hpp"
+
+#include <backendsCommon/test/DataTypeUtils.hpp>
+#include <backendsCommon/test/TensorCopyUtils.hpp>
+#include <backendsCommon/test/WorkloadTestUtils.hpp>
+
+namespace
+{
+
+template<typename T, size_t NumDims>
+LayerTestResult<T, NumDims> ChannelShuffleTestImpl(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory,
+        armnn::ChannelShuffleDescriptor descriptor,
+        armnn::TensorInfo inputTensorInfo,
+        armnn::TensorInfo outputTensorInfo,
+        const std::vector<T>& inputData,
+        const std::vector<T>& outputExpectedData)
+{
+    IgnoreUnused(memoryManager);
+    std::vector<T> actualOutput(outputTensorInfo.GetNumElements());
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::ChannelShuffleQueueDescriptor data;
+    data.m_Parameters = descriptor;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateChannelShuffle(data, info);
+
+    inputHandle->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle.get(), inputData.data());
+
+    workload->Execute();
+
+    CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
+
+    return LayerTestResult<T, NumDims>(actualOutput,
+                                       outputExpectedData,
+                                       outputHandle->GetShape(),
+                                       outputTensorInfo.GetShape());
+}
+} // anonymous namespace
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 4> SimpleChannelShuffleTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo;
+    armnn::TensorInfo outputTensorInfo;
+
+    unsigned int inputShape[] = { 1,9,1,1 };
+    unsigned int outputShape[] = { 1,9,1,1 };
+
+    armnn::ChannelShuffleDescriptor descriptor;
+    descriptor.m_Axis = 1;
+    descriptor.m_NumGroups = 3;
+
+    inputTensorInfo = armnn::TensorInfo(4, inputShape, ArmnnType);
+    inputTensorInfo.SetQuantizationScale(1.0f);
+    outputTensorInfo = armnn::TensorInfo(4, outputShape, ArmnnType);
+    outputTensorInfo.SetQuantizationScale(1.0f);
+
+    auto input = ConvertToDataType<ArmnnType>(
+            {
+                0.0f, 1.0f, 2.0f,   3.0f, 4.0f, 5.0f,   6.0f, 7.0f, 8.0f
+            },
+            inputTensorInfo);
+    auto outputExpected = ConvertToDataType<ArmnnType>(
+            {
+                0.0f, 3.0f, 6.0f, 1.0f, 4.0f, 7.0f, 2.0f, 5.0f, 8.0f
+            },
+            outputTensorInfo);
+
+    return ChannelShuffleTestImpl<T, 4>(
+                workloadFactory,
+                memoryManager,
+                tensorHandleFactory,
+                descriptor,
+                inputTensorInfo,
+                outputTensorInfo,
+                input,
+                outputExpected);
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 2> ChannelShuffle2DTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo;
+    armnn::TensorInfo outputTensorInfo;
+
+    unsigned int inputShape[] = { 3, 12 };
+    unsigned int outputShape[] = { 3, 12 };
+
+    armnn::ChannelShuffleDescriptor descriptor;
+    descriptor.m_Axis = 1;
+    descriptor.m_NumGroups = 3;
+
+    inputTensorInfo = armnn::TensorInfo(2, inputShape, ArmnnType);
+    inputTensorInfo.SetQuantizationScale(1.0f);
+    outputTensorInfo = armnn::TensorInfo(2, outputShape, ArmnnType);
+    outputTensorInfo.SetQuantizationScale(1.0f);
+
+    auto input = ConvertToDataType<ArmnnType>(
+            {
+                0, 1, 2, 3,       4, 5, 6, 7,       8, 9, 10, 11,
+               12, 13, 14, 15,   16, 17, 18, 19,   20, 21, 22, 23,
+               24, 25, 26, 27,   28, 29, 30, 31,   32, 33, 34, 35
+            },
+            inputTensorInfo);
+
+    auto outputExpected = ConvertToDataType<ArmnnType>(
+            {
+                0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11,
+                12, 16, 20, 13, 17, 21, 14, 18, 22, 15, 19, 23,
+                24, 28, 32, 25, 29, 33, 26, 30, 34, 27, 31, 35
+            },
+            outputTensorInfo);
+
+    return ChannelShuffleTestImpl<T, 2>(
+            workloadFactory,
+            memoryManager,
+            tensorHandleFactory,
+            descriptor,
+            inputTensorInfo,
+            outputTensorInfo,
+            input,
+            outputExpected);
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 4> ChannelShuffle4DTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo;
+    armnn::TensorInfo outputTensorInfo;
+
+    unsigned int inputShape[] = { 2, 9, 1, 2 };
+    unsigned int outputShape[] = { 2, 9, 1, 2 };
+
+    armnn::ChannelShuffleDescriptor descriptor;
+    descriptor.m_Axis = 1;
+    descriptor.m_NumGroups = 3;
+
+    inputTensorInfo = armnn::TensorInfo(4, inputShape, ArmnnType);
+    inputTensorInfo.SetQuantizationScale(1.0f);
+    outputTensorInfo = armnn::TensorInfo(4, outputShape, ArmnnType);
+    outputTensorInfo.SetQuantizationScale(1.0f);
+
+    auto input = ConvertToDataType<ArmnnType>(
+            {
+                    0.0f, 1.0f,
+                    2.0f, 3.0f,
+                    4.0f, 5.0f,
+
+                    6.0f, 7.0f,
+                    8.0f, 9.0f,
+                    10.0f, 11.0f,
+
+                    12.0f, 13.0f,
+                    14.0f, 15.0f,
+                    16.0f, 17.0f,
+
+                    18.0f, 19.0f,
+                    20.0f, 21.0f,
+                    22.0f, 23.0f,
+
+                    24.0f, 25.0f,
+                    26.0f, 27.0f,
+                    28.0f, 29.0f,
+
+                    30.0f, 31.0f,
+                    32.0f, 33.0f,
+                    34.0f, 35.0f
+            },
+            inputTensorInfo);
+
+    auto outputExpected = ConvertToDataType<ArmnnType>(
+            {
+                    0.0f, 1.0f,
+                    6.0f, 7.0f,
+                    12.0f, 13.0f,
+                    2.0f, 3.0f,
+                    8.0f, 9.0f,
+                    14.0f, 15.0f,
+                    4.0f, 5.0f,
+                    10.0f, 11.0f,
+                    16.0f, 17.0f,
+
+                    18.0f, 19.0f,
+                    24.0f, 25.0f,
+                    30.0f, 31.0f,
+                    20.0f, 21.0f,
+                    26.0f, 27.0f,
+                    32.0f, 33.0f,
+                    22.0f, 23.0f,
+                    28.0f, 29.0f,
+                    34.0f, 35.0f
+            },
+            outputTensorInfo);
+
+    return ChannelShuffleTestImpl<T, 4>(
+            workloadFactory,
+            memoryManager,
+            tensorHandleFactory,
+            descriptor,
+            inputTensorInfo,
+            outputTensorInfo,
+            input,
+            outputExpected);
+}
+
+//
+// Explicit template specializations
+//
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4>
+SimpleChannelShuffleTest<armnn::DataType::Float32>(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4>
+SimpleChannelShuffleTest<armnn::DataType::QAsymmU8>(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2>
+ChannelShuffle2DTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2>
+ChannelShuffle2DTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4>
+ChannelShuffle4DTest<armnn::DataType::Float32>(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4>
+ChannelShuffle4DTest<armnn::DataType::QAsymmU8>(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
\ No newline at end of file
diff --git a/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.hpp
new file mode 100644
index 0000000000..3500e72ae7
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ChannelShuffleTestImpl.hpp
@@ -0,0 +1,31 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "LayerTestResult.hpp"
+
+#include <ResolveType.hpp>
+
+#include <armnn/backends/IBackendInternal.hpp>
+#include <backendsCommon/WorkloadFactory.hpp>
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> SimpleChannelShuffleTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 2> ChannelShuffle2DTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> ChannelShuffle4DTest(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
\ No newline at end of file