MLCE-510 Add CpuRef Shape Operator to ArmNN

 * Add front end
 * Add reference workload
 * Serialization/Deserialization
 * Add unit tests
 * Update ArmNN Versioning

Signed-off-by: Keith Davis <keith.davis@arm.com>
Change-Id: I6fcb1fa341d6f08dea4003b13544e6e9f53fefd3

20 files changed, 540 insertions, 4 deletions

diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index 2e171f98ae..8a24e1161b 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -557,6 +557,13 @@ bool LayerSupportBase::IsRsqrtSupported(const TensorInfo&, // input
     return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
 }
 
+bool LayerSupportBase::IsShapeSupported(const TensorInfo&, // input
+                                        const TensorInfo&, // output
+                                        Optional<std::string&> reasonIfUnsupported) const
+{
+    return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
 bool LayerSupportBase::IsSliceSupported(const TensorInfo&, // input
                                         const TensorInfo&, // output
                                         const SliceDescriptor&, // descriptor
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index a4f972f961..0277a782a1 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -344,6 +344,10 @@ public:
                           const TensorInfo& output,
                           Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsShapeSupported(const TensorInfo& input,
+                          const TensorInfo& output,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsSliceSupported(const TensorInfo& input,
                           const TensorInfo& output,
                           const SliceDescriptor& descriptor,
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index 44a6a17b37..8c78136185 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -2805,6 +2805,33 @@ void MergeQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
     ValidateTensorDataTypesMatch(inputTensorInfo0, outputTensorInfo, descriptorName, "input_0", "output");
 }
 
+void ShapeQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+    const std::string& descriptorName{"ShapeQueueDescriptor"};
+
+    ValidateNumInputs(workloadInfo,  descriptorName, 1);
+    ValidateNumOutputs(workloadInfo, descriptorName, 1);
+
+    const TensorInfo& inputTensorInfo = workloadInfo.m_InputTensorInfos[0];
+    const TensorInfo& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0];
+
+    std::vector<DataType> supportedTypes =
+    {
+        DataType::BFloat16,
+        DataType::Float16,
+        DataType::Float32,
+        DataType::QAsymmS8,
+        DataType::QAsymmU8,
+        DataType::QAsymmS8,
+        DataType::QSymmS8,
+        DataType::QSymmS16,
+        DataType::Signed32
+    };
+
+    ValidateDataTypes(inputTensorInfo, supportedTypes, descriptorName);
+    ValidateDataTypes(outputTensorInfo, {DataType::Signed32}, descriptorName);
+}
+
 void SwitchQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
 {
     const std::string& descriptorName{"SwitchQueueDescriptor"};
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 11ce2cb44f..36653bdc0d 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -690,4 +690,9 @@ struct ReduceQueueDescriptor : QueueDescriptorWithParameters<ReduceDescriptor>
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct ShapeQueueDescriptor : QueueDescriptor
+{
+    void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
 } // namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index c5fc9d0fe2..61ad20995b 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -1003,6 +1003,16 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId,
                                                           reason);
             break;
         }
+        case LayerType::Shape:
+        {
+            const TensorInfo& input  = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+
+            result = layerSupportObject.IsShapeSupported(OverrideDataType(input, dataType),
+                                                         OverrideDataType(output, dataType),
+                                                         reason);
+            break;
+        }
         case LayerType::Slice:
         {
             auto cLayer = PolymorphicDowncast<const SliceLayer*>(&layer);
@@ -1673,6 +1683,12 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateRsqrt(const RsqrtQueueDescrip
     return std::unique_ptr<IWorkload>();
 }
 
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateShape(const ShapeQueueDescriptor& /*descriptor*/,
+                                                         const WorkloadInfo& /*info*/) const
+{
+    return std::unique_ptr<IWorkload>();
+}
+
 std::unique_ptr<IWorkload> IWorkloadFactory::CreateSlice(const SliceQueueDescriptor& /*descriptor*/,
                                                          const WorkloadInfo& /*info*/) const
 {
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index 42360d37ff..1987b9b664 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -252,6 +252,9 @@ public:
     virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const;
 
+    virtual std::unique_ptr<IWorkload> CreateShape(const ShapeQueueDescriptor& descriptor,
+                                                   const WorkloadInfo& info) const;
+
     virtual std::unique_ptr<IWorkload> CreateSlice(const SliceQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const;
 
diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 73a16d00e1..ff9375dec1 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -85,6 +85,7 @@ COMMON_TEST_SOURCES := \
     test/layerTests/SliceTestImpl.cpp \
     test/layerTests/QuantizeTestImpl.cpp \
     test/layerTests/SinTestImpl.cpp \
+    test/layerTests/ShapeTestImpl.cpp \
     test/layerTests/SoftmaxTestImpl.cpp \
     test/layerTests/SpaceToBatchNdTestImpl.cpp \
     test/layerTests/SpaceToDepthTestImpl.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index 82381a8b84..162368fd43 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -157,6 +157,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     layerTests/RsqrtTestImpl.hpp
     layerTests/SinTestImpl.cpp
     layerTests/SinTestImpl.hpp
+    layerTests/ShapeTestImpl.cpp
+    layerTests/ShapeTestImpl.cpp
     layerTests/SliceTestImpl.cpp
     layerTests/SliceTestImpl.hpp
     layerTests/SoftmaxTestImpl.cpp
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index 5a05ee1d85..adc7bc4c3c 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -663,6 +663,8 @@ DECLARE_LAYER_POLICY_2_PARAM(Resize)
 
 DECLARE_LAYER_POLICY_2_PARAM(Reshape)
 
+DECLARE_LAYER_POLICY_1_PARAM(Shape)
+
 DECLARE_LAYER_POLICY_2_PARAM(Slice)
 
 DECLARE_LAYER_POLICY_2_PARAM(Softmax)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 4ae6553362..46eb6ee2a5 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -55,6 +55,7 @@
 #include <backendsCommon/test/layerTests/ReshapeTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ResizeTestImpl.hpp>
 #include <backendsCommon/test/layerTests/RsqrtTestImpl.hpp>
+#include <backendsCommon/test/layerTests/ShapeTestImpl.hpp>
 #include <backendsCommon/test/layerTests/SinTestImpl.hpp>
 #include <backendsCommon/test/layerTests/SliceTestImpl.hpp>
 #include <backendsCommon/test/layerTests/SoftmaxTestImpl.hpp>
diff --git a/src/backends/backendsCommon/test/layerTests/ShapeTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/ShapeTestImpl.cpp
new file mode 100644
index 0000000000..d6c03141ab
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ShapeTestImpl.cpp
@@ -0,0 +1,306 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ShapeTestImpl.hpp"
+
+#include <backendsCommon/test/DataTypeUtils.hpp>
+#include <backendsCommon/test/TensorCopyUtils.hpp>
+#include <backendsCommon/test/WorkloadTestUtils.hpp>
+
+#include <test/TensorHelpers.hpp>
+
+template<typename T, std::size_t n>
+LayerTestResult<int32_t, 1> ShapeTest(
+    armnn::TensorInfo inputTensorInfo,
+    std::vector<T> input,
+    armnn::TensorInfo outputTensorInfo,
+    std::vector<int32_t> expectedOutputData,
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    IgnoreUnused(memoryManager);
+
+    std::vector<int32_t> actualOutput(outputTensorInfo.GetNumElements());
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::ShapeQueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateShape(data, info);
+
+    inputHandle->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle.get(), input.data());
+
+    workload->Execute();
+
+    CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get());
+
+    return LayerTestResult<int32_t, 1>(actualOutput,
+                                       expectedOutputData,
+                                       outputHandle->GetShape(),
+                                       outputTensorInfo.GetShape());
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<int32_t, 1> ShapeDimSize1Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo({ 1 }, ArmnnType, 1.0f, 0);
+    armnn::TensorInfo outputTensorInfo({ 1 }, armnn::DataType::Signed32);
+
+    auto input = ConvertToDataType<ArmnnType>({ 1.0f }, inputTensorInfo);
+
+    return ShapeTest<T, 1>(inputTensorInfo, input, outputTensorInfo, { 1 }, workloadFactory, memoryManager,
+                           tensorHandleFactory);
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<int32_t, 1> ShapeDimSize2Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo({ 1, 2 }, ArmnnType, 1.0f, 0);
+    armnn::TensorInfo outputTensorInfo({ 2 }, armnn::DataType::Signed32);
+
+    auto input = ConvertToDataType<ArmnnType>({ 1.0f, 1.0f }, inputTensorInfo);
+
+    return ShapeTest<T, 2>(inputTensorInfo, input, outputTensorInfo, { 1, 2 }, workloadFactory, memoryManager,
+                           tensorHandleFactory);
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<int32_t, 1> ShapeDimSize3Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo({ 1, 2, 3 }, ArmnnType, 1.0f, 0);
+    armnn::TensorInfo outputTensorInfo({ 3 }, armnn::DataType::Signed32);
+
+    auto input = ConvertToDataType<ArmnnType>({ 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f }, inputTensorInfo);
+
+    return ShapeTest<T, 3>(inputTensorInfo, input, outputTensorInfo, { 1, 2, 3 }, workloadFactory, memoryManager,
+                           tensorHandleFactory);
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<int32_t, 1> ShapeDimSize4Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    armnn::TensorInfo inputTensorInfo({ 2, 2, 3, 2 }, ArmnnType, 1.0f, 0);
+    armnn::TensorInfo outputTensorInfo({ 4 }, armnn::DataType::Signed32);
+
+    auto input = ConvertToDataType<ArmnnType>({ 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+                                                1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f,
+                                                1.0f },
+                                              inputTensorInfo);
+
+    return ShapeTest<T, 4>(inputTensorInfo, input, outputTensorInfo, { 2, 2, 3, 2 }, workloadFactory, memoryManager,
+                           tensorHandleFactory);
+}
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::Signed32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::QSymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize4Test<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::Signed32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::QSymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize3Test<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::Signed32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::QSymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize2Test<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::Signed32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::QSymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<int32_t, 1>
+ShapeDimSize1Test<armnn::DataType::BFloat16>(
+    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/ShapeTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/ShapeTestImpl.hpp
new file mode 100644
index 0000000000..85f7c0a453
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ShapeTestImpl.hpp
@@ -0,0 +1,45 @@
+//
+// 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<typename T , std::size_t n>
+LayerTestResult<int32_t, 1> ShapeTest(
+        armnn::TensorInfo inputTensorInfo,
+        std::vector<T> input,
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<int32_t, 1> ShapeDimSize1Test(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<int32_t, 1> ShapeDimSize2Test(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<int32_t, 1> ShapeDimSize3Test(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<int32_t, 1> ShapeDimSize4Test(
+        armnn::IWorkloadFactory& workloadFactory,
+        const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+        const armnn::ITensorHandleFactory& tensorHandleFactory);
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index 14a40f9d5d..1b05c4e0f4 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -1859,6 +1859,24 @@ bool RefLayerSupport::IsRsqrtSupported(const TensorInfo& input,
                                        reasonIfUnsupported);
 }
 
+bool RefLayerSupport::IsShapeSupported(const TensorInfo& input,
+                                       const TensorInfo& output,
+                                       Optional<std::string&> reasonIfUnsupported) const
+{
+    IgnoreUnused(input);
+    bool supported = true;
+
+    std::array<DataType, 1> supportedTypes =
+    {
+        DataType::Signed32
+    };
+
+    supported &= CheckSupportRule(TypeAnyOf(output, supportedTypes), reasonIfUnsupported,
+                                  "Reference Shape: output type not supported");
+
+    return supported;
+}
+
 bool RefLayerSupport::IsSliceSupported(const TensorInfo& input,
                                        const TensorInfo& output,
                                        const SliceDescriptor& descriptor,
diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp
index 7a95bb06ac..c060f79b5a 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -297,12 +297,16 @@ public:
                            const TensorInfo& output,
                            const ResizeDescriptor& descriptor,
                            Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
-    
+
     ARMNN_DEPRECATED_MSG("Use IsElementwiseUnarySupported instead")
     bool IsRsqrtSupported(const TensorInfo& input,
                           const TensorInfo& output,
                           Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsShapeSupported(const TensorInfo& input,
+                          const TensorInfo& output,
+                          Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsSliceSupported(const TensorInfo& input,
                           const TensorInfo& output,
                           const SliceDescriptor& descriptor,
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp
index 8e3bbe468f..606f531630 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -7,7 +7,6 @@
 #include <backendsCommon/MemImportWorkload.hpp>
 #include <backendsCommon/MakeWorkloadHelper.hpp>
 #include <backendsCommon/TensorHandle.hpp>
-#include <reference/workloads/RefFillWorkload.hpp>
 #include "RefWorkloadFactory.hpp"
 #include "RefBackendId.hpp"
 #include "workloads/RefWorkloads.hpp"
@@ -626,6 +625,12 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateRsqrt(const RsqrtQueueDescr
     return CreateElementwiseUnary(elementwiseUnaryDescriptor, info);
 }
 
+std::unique_ptr<IWorkload> RefWorkloadFactory::CreateShape(const ShapeQueueDescriptor& descriptor,
+                                                           const WorkloadInfo& info) const
+{
+    return std::make_unique<RefShapeWorkload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> RefWorkloadFactory::CreateSlice(const SliceQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
 {
diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp
index 734c5e49c3..2beffa77f3 100644
--- a/src/backends/reference/RefWorkloadFactory.hpp
+++ b/src/backends/reference/RefWorkloadFactory.hpp
@@ -243,6 +243,9 @@ public:
     std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                            const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateShape(const ShapeQueueDescriptor& descriptor,
+                                           const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateSlice(const SliceQueueDescriptor& descriptor,
                                            const WorkloadInfo& info) const override;
 
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index 1cc6fa8d36..45e3717268 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -1733,6 +1733,43 @@ ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_2, DepthToSpaceTest2<DataType::QSymmS
 ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_3, DepthToSpaceTest3<DataType::QSymmS16>, DataLayout::NHWC);
 ARMNN_AUTO_TEST_CASE(DepthToSpaceNhwcInt16_4, DepthToSpaceTest4<DataType::QSymmS16>, DataLayout::NHWC);
 
+// Shape
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1Float16,  ShapeDimSize1Test<DataType::Float16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1Float32,  ShapeDimSize1Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1QAsymmU8, ShapeDimSize1Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1Signed32, ShapeDimSize1Test<DataType::Signed32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1QSymmS16, ShapeDimSize1Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1QSymmS8,  ShapeDimSize1Test<DataType::QSymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1QAsymmS8, ShapeDimSize1Test<DataType::QAsymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize1BFloat16, ShapeDimSize1Test<DataType::BFloat16>)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2Float16,  ShapeDimSize2Test<DataType::Float16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2Float32,  ShapeDimSize2Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2QAsymmU8, ShapeDimSize2Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2Signed32, ShapeDimSize2Test<DataType::Signed32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2QSymmS16, ShapeDimSize2Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2QSymmS8,  ShapeDimSize2Test<DataType::QSymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2QAsymmS8, ShapeDimSize2Test<DataType::QAsymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize2BFloat16, ShapeDimSize2Test<DataType::BFloat16>)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3Float16,  ShapeDimSize3Test<DataType::Float16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3Float32,  ShapeDimSize3Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3QAsymmU8, ShapeDimSize3Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3Signed32, ShapeDimSize3Test<DataType::Signed32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3QSymmS16, ShapeDimSize3Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3QSymmS8,  ShapeDimSize3Test<DataType::QSymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3QAsymmS8, ShapeDimSize3Test<DataType::QAsymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize3BFloat16, ShapeDimSize3Test<DataType::BFloat16>)
+
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4Float16,  ShapeDimSize4Test<DataType::Float16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4Float32,  ShapeDimSize4Test<DataType::Float32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4QAsymmU8, ShapeDimSize4Test<DataType::QAsymmU8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4Signed32, ShapeDimSize4Test<DataType::Signed32>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4QSymmS16, ShapeDimSize4Test<DataType::QSymmS16>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4QSymmS8,  ShapeDimSize4Test<DataType::QSymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4QAsymmS8, ShapeDimSize4Test<DataType::QAsymmS8>)
+ARMNN_AUTO_TEST_CASE_WITH_THF(ShapeDimSize4BFloat16, ShapeDimSize4Test<DataType::BFloat16>)
+
 // SpaceToDepth
 ARMNN_AUTO_TEST_CASE_WITH_THF(SpaceToDepthNchwAsymmQ8, SpaceToDepthNchwAsymmQ8Test)
 ARMNN_AUTO_TEST_CASE_WITH_THF(SpaceToDepthNhwcAsymmQ8, SpaceToDepthNhwcAsymmQ8Test)
diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index 09e02e67bd..7a769e5246 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -143,6 +143,7 @@ list(APPEND armnnRefBackendWorkloads_sources
     RefResizeBilinearWorkload.hpp
     RefResizeWorkload.cpp
     RefResizeWorkload.hpp
+    RefShapeWorkload.hpp
     RefSliceWorkload.cpp
     RefSliceWorkload.hpp
     RefSoftmaxWorkload.cpp
diff --git a/src/backends/reference/workloads/RefShapeWorkload.hpp b/src/backends/reference/workloads/RefShapeWorkload.hpp
new file mode 100644
index 0000000000..8e2a410b0c
--- /dev/null
+++ b/src/backends/reference/workloads/RefShapeWorkload.hpp
@@ -0,0 +1,48 @@
+//
+// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
+
+#include "RefWorkloadUtils.hpp"
+
+namespace armnn
+{
+
+struct RefShapeWorkload : public BaseWorkload<ShapeQueueDescriptor>
+{
+public:
+    using BaseWorkload<ShapeQueueDescriptor>::BaseWorkload;
+    virtual void Execute() const override
+    {
+        Execute(m_Data.m_Inputs, m_Data.m_Outputs);
+    }
+    void ExecuteAsync(WorkingMemDescriptor& workingMemDescriptor)  override
+    {
+        Execute(workingMemDescriptor.m_Inputs, workingMemDescriptor.m_Outputs);
+    }
+
+private:
+    void Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const
+    {
+        const TensorShape Shape = GetTensorInfo(inputs[0]).GetShape();
+
+        const TensorInfo& outputInfo = GetTensorInfo(outputs[0]);
+
+        unsigned int numBytes =
+            GetTensorInfo(inputs[0]).GetNumDimensions() * GetDataTypeSize(outputInfo.GetDataType());
+
+        std::memcpy(outputs[0]->Map(), &Shape, numBytes);
+        outputs[0]->Unmap();
+    }
+};
+
+} //namespace armnn
+
+
+
+
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index d3995f2b82..afe63d13c0 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -35,10 +35,10 @@
 #include "RefDequantizeWorkload.hpp"
 #include "RefElementwiseWorkload.hpp"
 #include "RefElementwiseUnaryWorkload.hpp"
+#include "RefFakeQuantizationFloat32Workload.hpp"
 #include "RefFillWorkload.hpp"
-#include "RefFullyConnectedWorkload.hpp"
 #include "RefFloorWorkload.hpp"
-#include "RefFakeQuantizationFloat32Workload.hpp"
+#include "RefFullyConnectedWorkload.hpp"
 #include "RefGatherWorkload.hpp"
 #include "RefInstanceNormalizationWorkload.hpp"
 #include "RefL2NormalizationWorkload.hpp"
@@ -59,6 +59,7 @@
 #include "RefReshapeWorkload.hpp"
 #include "RefResizeBilinearWorkload.hpp"
 #include "RefResizeWorkload.hpp"
+#include "RefShapeWorkload.hpp"
 #include "RefSliceWorkload.hpp"
 #include "RefSplitterWorkload.hpp"
 #include "RefSoftmaxWorkload.hpp"