IVGCVSW-4515 Add ConvertBf16ToFp32Layer and Ref workload support

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: Ida6d7e1d2c9abe0618f8b711bab9d62c011090d6

33 files changed, 370 insertions, 0 deletions

diff --git a/Android.mk b/Android.mk
index c63022c3aa..436ae2207b 100644
--- a/Android.mk
+++ b/Android.mk
@@ -132,6 +132,7 @@ LOCAL_SRC_FILES := \
         src/armnn/layers/ConcatLayer.cpp \
         src/armnn/layers/ConstantLayer.cpp \
         src/armnn/layers/Convolution2dLayer.cpp \
+        src/armnn/layers/ConvertBf16ToFp32Layer.cpp \
         src/armnn/layers/ConvertFp16ToFp32Layer.cpp \
         src/armnn/layers/ConvertFp32ToFp16Layer.cpp \
         src/armnn/layers/DebugLayer.cpp \
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5002eb4e0b..d5da0d3aad 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -272,6 +272,8 @@ list(APPEND armnn_sources
     src/armnn/layers/ConstantLayer.cpp
     src/armnn/layers/Convolution2dLayer.hpp
     src/armnn/layers/Convolution2dLayer.cpp
+    src/armnn/layers/ConvertBf16ToFp32Layer.cpp
+    src/armnn/layers/ConvertBf16ToFp32Layer.hpp
     src/armnn/layers/ConvertFp16ToFp32Layer.hpp
     src/armnn/layers/ConvertFp16ToFp32Layer.cpp
     src/armnn/layers/ConvertFp32ToFp16Layer.hpp
diff --git a/include/armnn/ILayerSupport.hpp b/include/armnn/ILayerSupport.hpp
index af91e87376..17ed4bef5a 100644
--- a/include/armnn/ILayerSupport.hpp
+++ b/include/armnn/ILayerSupport.hpp
@@ -75,6 +75,10 @@ public:
     virtual bool IsConstantSupported(const TensorInfo& output,
                                      Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0;
 
+    virtual bool IsConvertBf16ToFp32Supported(const TensorInfo& input,
+                                              const TensorInfo& output,
+                                              Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0;
+
     virtual bool IsConvertFp16ToFp32Supported(const TensorInfo& input,
                                               const TensorInfo& output,
                                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0;
diff --git a/src/armnn/InternalTypes.cpp b/src/armnn/InternalTypes.cpp
index c032e44cd3..3f3eed56e7 100644
--- a/src/armnn/InternalTypes.cpp
+++ b/src/armnn/InternalTypes.cpp
@@ -22,6 +22,7 @@ char const* GetLayerTypeAsCString(LayerType type)
         case LayerType::Comparison: return "Comparison";
         case LayerType::Concat: return "Concat";
         case LayerType::Constant: return "Constant";
+        case LayerType::ConvertBf16ToFp32: return "ConvertBf16ToFp32";
         case LayerType::ConvertFp16ToFp32: return "ConvertFp16ToFp32";
         case LayerType::ConvertFp32ToFp16: return "ConvertFp32ToFp16";
         case LayerType::Convolution2d: return "Convolution2d";
diff --git a/src/armnn/InternalTypes.hpp b/src/armnn/InternalTypes.hpp
index 351f12c510..9330122246 100644
--- a/src/armnn/InternalTypes.hpp
+++ b/src/armnn/InternalTypes.hpp
@@ -22,6 +22,7 @@ enum class LayerType
     Comparison,
     Concat,
     Constant,
+    ConvertBf16ToFp32,
     ConvertFp16ToFp32,
     ConvertFp32ToFp16,
     Convolution2d,
diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp
index f3ce7e61fa..3dde908fc3 100644
--- a/src/armnn/LayersFwd.hpp
+++ b/src/armnn/LayersFwd.hpp
@@ -14,6 +14,7 @@
 #include "layers/ComparisonLayer.hpp"
 #include "layers/ConcatLayer.hpp"
 #include "layers/ConstantLayer.hpp"
+#include "layers/ConvertBf16ToFp32Layer.hpp"
 #include "layers/ConvertFp16ToFp32Layer.hpp"
 #include "layers/ConvertFp32ToFp16Layer.hpp"
 #include "layers/Convolution2dLayer.hpp"
@@ -99,6 +100,7 @@ DECLARE_LAYER(BatchToSpaceNd)
 DECLARE_LAYER(Comparison)
 DECLARE_LAYER(Concat)
 DECLARE_LAYER(Constant)
+DECLARE_LAYER(ConvertBf16ToFp32)
 DECLARE_LAYER(ConvertFp16ToFp32)
 DECLARE_LAYER(ConvertFp32ToFp16)
 DECLARE_LAYER(Convolution2d)
diff --git a/src/armnn/layers/ConvertBf16ToFp32Layer.cpp b/src/armnn/layers/ConvertBf16ToFp32Layer.cpp
new file mode 100644
index 0000000000..147aa8f46a
--- /dev/null
+++ b/src/armnn/layers/ConvertBf16ToFp32Layer.cpp
@@ -0,0 +1,55 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ConvertBf16ToFp32Layer.hpp"
+#include "LayerCloneBase.hpp"
+
+#include <armnn/TypesUtils.hpp>
+
+#include <backendsCommon/WorkloadData.hpp>
+#include <backendsCommon/WorkloadFactory.hpp>
+
+namespace armnn
+{
+
+ConvertBf16ToFp32Layer::ConvertBf16ToFp32Layer(const char* name)
+    : Layer(1, 1, LayerType::ConvertBf16ToFp32, name)
+{
+}
+
+std::unique_ptr<IWorkload> ConvertBf16ToFp32Layer::CreateWorkload(const IWorkloadFactory& factory) const
+{
+    ConvertBf16ToFp32QueueDescriptor descriptor;
+    return factory.CreateConvertBf16ToFp32(descriptor, PrepInfoAndDesc(descriptor));
+}
+
+ConvertBf16ToFp32Layer* ConvertBf16ToFp32Layer::Clone(Graph& graph) const
+{
+    return CloneBase<ConvertBf16ToFp32Layer>(graph, GetName());
+}
+
+void ConvertBf16ToFp32Layer::ValidateTensorShapesFromInputs()
+{
+    VerifyLayerConnections(1, CHECK_LOCATION());
+
+    auto inferredShapes = InferOutputShapes({ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape() });
+
+    BOOST_ASSERT(inferredShapes.size() == 1);
+
+    ConditionalThrowIfNotEqual<LayerValidationException>(
+        "ConvertBf16ToFp32Layer: TensorShape set on OutputSlot[0] does not match the inferred shape.",
+        GetOutputSlot(0).GetTensorInfo().GetShape(),
+        inferredShapes[0]);
+}
+
+void ConvertBf16ToFp32Layer::Accept(ILayerVisitor& visitor) const
+{
+    // these conversion layers are only inserted by the
+    // optimizer and so will never be in an input graph.
+    IgnoreUnused(visitor);
+    throw armnn::Exception("ConvertBf16ToFp32Layer should never appear in an input graph");
+}
+
+} // namespace armnn
diff --git a/src/armnn/layers/ConvertBf16ToFp32Layer.hpp b/src/armnn/layers/ConvertBf16ToFp32Layer.hpp
new file mode 100644
index 0000000000..2a79a1cb65
--- /dev/null
+++ b/src/armnn/layers/ConvertBf16ToFp32Layer.hpp
@@ -0,0 +1,42 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <Layer.hpp>
+
+namespace armnn
+{
+
+/// This layer converts data type BFloat16 to Float32.
+class ConvertBf16ToFp32Layer : public Layer
+{
+public:
+    /// Makes a workload for the ConvertBf16ToFp32 type.
+    /// @param [in] graph The graph where this layer can be found.
+    /// @param [in] factory The workload factory which will create the workload.
+    /// @return A pointer to the created workload, or nullptr if not created.
+    virtual std::unique_ptr<IWorkload> CreateWorkload(const IWorkloadFactory& factory) const override;
+
+    /// Creates a dynamically-allocated copy of this layer.
+    /// @param [in] graph The graph into which this layer is being cloned.
+    ConvertBf16ToFp32Layer* Clone(Graph& graph) const override;
+
+    /// Check if the input tensor shape(s)
+    /// will lead to a valid configuration of @ref ConvertBf16ToFp32Layer.
+    void ValidateTensorShapesFromInputs() override;
+
+    void Accept(ILayerVisitor& visitor) const override;
+
+protected:
+    /// Constructor to create a ConvertBf16ToFp32Layer.
+    /// @param [in] name Optional name for the layer.
+    ConvertBf16ToFp32Layer(const char* name);
+
+    /// Default destructor
+    ~ConvertBf16ToFp32Layer() = default;
+};
+
+} // namespace
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index e8ef46ecd8..1ac08afd7a 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -111,6 +111,13 @@ bool LayerSupportBase::IsConstantSupported(const TensorInfo& /*output*/,
     return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
 }
 
+bool LayerSupportBase::IsConvertBf16ToFp32Supported(const TensorInfo& /*input*/,
+                                                    const TensorInfo& /*output*/,
+                                                    Optional<std::string&> reasonIfUnsupported) const
+{
+    return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
 bool LayerSupportBase::IsConvertFp16ToFp32Supported(const TensorInfo& /*input*/,
                                                     const TensorInfo& /*output*/,
                                                     Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index 888bef5f89..59e8b969bd 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -61,6 +61,10 @@ public:
     bool IsConstantSupported(const TensorInfo& output,
                              Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsConvertBf16ToFp32Supported(const TensorInfo& input,
+                                      const TensorInfo& output,
+                                      Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsConvertFp16ToFp32Supported(const TensorInfo& input,
                                       const TensorInfo& output,
                                       Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/backendsCommon/Workload.hpp b/src/backends/backendsCommon/Workload.hpp
index 8ec09f98b6..d7434c0d01 100644
--- a/src/backends/backendsCommon/Workload.hpp
+++ b/src/backends/backendsCommon/Workload.hpp
@@ -177,6 +177,11 @@ using BaseUint8ComparisonWorkload = MultiTypedWorkload<QueueDescriptor,
                                                        armnn::DataType::Boolean>;
 
 template <typename QueueDescriptor>
+using BFloat16ToFloat32Workload = MultiTypedWorkload<QueueDescriptor,
+                                                     armnn::DataType::BFloat16,
+                                                     armnn::DataType::Float32>;
+
+template <typename QueueDescriptor>
 using Float16ToFloat32Workload = MultiTypedWorkload<QueueDescriptor,
                                                     armnn::DataType::Float16,
                                                     armnn::DataType::Float32>;
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index b501b3dbec..81aefa94e7 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -2016,6 +2016,29 @@ void LstmQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
     }
 }
 
+void ConvertBf16ToFp32QueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+    const std::string descriptorName{"ConvertBf16ToFp32QueueDescriptor"};
+
+    ValidateNumInputs(workloadInfo,  descriptorName, 1);
+    ValidateNumOutputs(workloadInfo, descriptorName, 1);
+
+    const TensorInfo& inputTensorInfo  = workloadInfo.m_InputTensorInfos[0];
+    const TensorInfo& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0];
+
+    if (inputTensorInfo.GetDataType() != DataType::BFloat16)
+    {
+        throw InvalidArgumentException(descriptorName + ": Input tensor type must be BFloat16.");
+    }
+
+    if (outputTensorInfo.GetDataType() != DataType::Float32)
+    {
+        throw InvalidArgumentException(descriptorName + ": Output tensor type must be Float32.");
+    }
+
+    ValidateTensorShapesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output");
+}
+
 void ConvertFp32ToFp16QueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
 {
     const std::string descriptorName{"ConvertFp32ToFp16QueueDescriptor"};
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 06289fa039..9c392d3219 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -407,6 +407,11 @@ struct LstmQueueDescriptor : QueueDescriptorWithParameters<LstmDescriptor>
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct ConvertBf16ToFp32QueueDescriptor : QueueDescriptor
+{
+    void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
 struct ConvertFp16ToFp32QueueDescriptor : QueueDescriptor
 {
     void Validate(const WorkloadInfo& workloadInfo) const;
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 2e1ce0a674..d932eef49f 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -159,6 +159,13 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId,
             result = layerSupportObject->IsConstantSupported(OverrideDataType(output, dataType), reason);
             break;
         }
+        case LayerType::ConvertBf16ToFp32:
+        {
+            const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+            result = layerSupportObject->IsConvertBf16ToFp32Supported(input, output, reason);
+            break;
+        }
         case LayerType::ConvertFp16ToFp32:
         {
             const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
@@ -1144,6 +1151,12 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateConstant(const ConstantQueueD
     return std::unique_ptr<IWorkload>();
 }
 
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertBf16ToFp32(const ConvertBf16ToFp32QueueDescriptor& /*desc*/,
+                                                                     const WorkloadInfo& /*info*/) const
+{
+    return std::unique_ptr<IWorkload>();
+}
+
 std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& /*desc*/,
                                                                      const WorkloadInfo& /*info*/) const
 {
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index dae58b6d93..8c22452f2e 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -79,6 +79,9 @@ public:
     virtual std::unique_ptr<IWorkload> CreateConstant(const ConstantQueueDescriptor& descriptor,
                                                       const WorkloadInfo& info) const;
 
+    virtual std::unique_ptr<IWorkload> CreateConvertBf16ToFp32(const ConvertBf16ToFp32QueueDescriptor& descriptor,
+                                                               const WorkloadInfo& info) const;
+
     virtual std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor,
                                                                const WorkloadInfo& info) const;
 
diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 56a21b386c..22de3db77c 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -44,6 +44,7 @@ COMMON_TEST_SOURCES := \
     test/layerTests/ConcatTestImpl.cpp \
     test/layerTests/ConstantTestImpl.cpp \
     test/layerTests/Conv2dTestImpl.cpp \
+    test/layerTests/ConvertBf16ToFp32TestImpl.cpp \
     test/layerTests/ConvertFp16ToFp32TestImpl.cpp \
     test/layerTests/ConvertFp32ToFp16TestImpl.cpp \
     test/layerTests/DebugTestImpl.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index 0376e3e5e6..dc8031f6b4 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -68,6 +68,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     layerTests/ConstantTestImpl.hpp
     layerTests/Conv2dTestImpl.cpp
     layerTests/Conv2dTestImpl.hpp
+    layerTests/ConvertBf16ToFp32TestImpl.cpp
+    layerTests/ConvertBf16ToFp32TestImpl.hpp
     layerTests/ConvertFp16ToFp32TestImpl.cpp
     layerTests/ConvertFp16ToFp32TestImpl.hpp
     layerTests/ConvertFp32ToFp16TestImpl.cpp
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index 15608ccdd8..a070ac05e8 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -443,6 +443,8 @@ DECLARE_LAYER_POLICY_2_PARAM(Concat)
 
 DECLARE_LAYER_POLICY_1_PARAM(Constant)
 
+DECLARE_LAYER_POLICY_1_PARAM(ConvertBf16ToFp32)
+
 DECLARE_LAYER_POLICY_1_PARAM(ConvertFp16ToFp32)
 
 DECLARE_LAYER_POLICY_1_PARAM(ConvertFp32ToFp16)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 62a66df166..1c6277a333 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -13,6 +13,7 @@
 #include <backendsCommon/test/layerTests/BatchToSpaceNdTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ComparisonTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConcatTestImpl.hpp>
+#include <backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertFp16ToFp32TestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertFp32ToFp16TestImpl.hpp>
 #include <backendsCommon/test/layerTests/Conv2dTestImpl.hpp>
diff --git a/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.cpp b/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.cpp
new file mode 100644
index 0000000000..0dc3048ee3
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.cpp
@@ -0,0 +1,56 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ConvertBf16ToFp32TestImpl.hpp"
+
+#include <backendsCommon/test/TensorCopyUtils.hpp>
+#include <backendsCommon/test/WorkloadTestUtils.hpp>
+
+#include <test/TensorHelpers.hpp>
+
+LayerTestResult<float, 4> ConvertBf16ToFp32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    IgnoreUnused(memoryManager);
+
+    const armnn::TensorInfo inputTensorInfo({1, 3, 2, 3}, armnn::DataType::BFloat16);
+    const armnn::TensorInfo outputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float32);
+
+    std::vector<armnn::BFloat16> inputValues = armnnUtils::QuantizedVector<armnn::BFloat16>(
+        {
+            -37.5f, -15.2f, -8.76f, -2.0f, -1.5f, -1.3f, -0.5f, -0.4f, 0.0f,
+          1.0f, 0.4f, 0.5f, 1.3f, 1.5f, 2.0f, 8.76f, 15.2f, 37.5f
+        },
+        1.0f, 0);
+
+    auto input = MakeTensor<armnn::BFloat16, 4>(inputTensorInfo, std::vector<armnn::BFloat16>(inputValues));
+
+    LayerTestResult<float, 4> ret(outputTensorInfo);
+    ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo,
+        { -37.5f, -15.2f, -8.76f, -2.0f, -1.5f, -1.3f, -0.5f, -0.4f, 0.0f,
+          1.0f, 0.4f, 0.5f, 1.3f, 1.5f, 2.0f, 8.76f, 15.2f, 37.5f });
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::ConvertBf16ToFp32QueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvertBf16ToFp32(data, info);
+
+    inputHandle->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]);
+
+    workload->Execute();
+
+    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
+
+    return ret;
+}
diff --git a/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp b/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp
new file mode 100644
index 0000000000..717ec6a121
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp
@@ -0,0 +1,17 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "LayerTestResult.hpp"
+
+#include <BFloat16.hpp>
+
+#include <armnn/backends/IBackendInternal.hpp>
+#include <backendsCommon/WorkloadFactory.hpp>
+
+LayerTestResult<float, 4> ConvertBf16ToFp32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index 94128fe7cd..a4f4efd92a 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -376,6 +376,21 @@ bool RefLayerSupport::IsConstantSupported(const TensorInfo& output,
                                   "Reference constant: output is not a supported type.");
 }
 
+bool RefLayerSupport::IsConvertBf16ToFp32Supported(const TensorInfo& input,
+                                                   const TensorInfo& output,
+                                                   Optional<std::string&> reasonIfUnsupported) const
+{
+    bool supported = true;
+
+    supported &= CheckSupportRule(TypeIs(input, DataType::BFloat16), reasonIfUnsupported,
+                                  "Reference for ConvertBf16ToFp32 layer: input type not supported");
+
+    supported &= CheckSupportRule(TypeIs(output, DataType::Float32), reasonIfUnsupported,
+                                  "Reference for ConvertBf16ToFp32 layer: output type not supported");
+
+    return supported;
+}
+
 bool RefLayerSupport::IsConvertFp16ToFp32Supported(const TensorInfo& input,
                                                    const TensorInfo& output,
                                                    Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/reference/RefLayerSupport.hpp b/src/backends/reference/RefLayerSupport.hpp
index 27f3f81489..ff3478142f 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -60,6 +60,10 @@ public:
     bool IsConstantSupported(const TensorInfo& output,
                              Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsConvertBf16ToFp32Supported(const TensorInfo& input,
+                                      const TensorInfo& output,
+                                      Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsConvertFp16ToFp32Supported(const TensorInfo& input,
                                       const TensorInfo& output,
                                       Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp
index aebf19bf28..c9fc626bc1 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -180,6 +180,13 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConstant(const ConstantQueu
     return std::make_unique<RefConstantWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConvertBf16ToFp32(
+    const ConvertBf16ToFp32QueueDescriptor& descriptor,
+    const WorkloadInfo& info) const
+{
+    return std::make_unique<RefConvertBf16ToFp32Workload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConvertFp16ToFp32(
     const ConvertFp16ToFp32QueueDescriptor& descriptor,
     const WorkloadInfo& info) const
diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp
index b64479e207..119605b998 100644
--- a/src/backends/reference/RefWorkloadFactory.hpp
+++ b/src/backends/reference/RefWorkloadFactory.hpp
@@ -86,6 +86,9 @@ public:
     std::unique_ptr<IWorkload> CreateConstant(const ConstantQueueDescriptor& descriptor,
                                               const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateConvertBf16ToFp32(const ConvertBf16ToFp32QueueDescriptor& descriptor,
+                                                       const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor,
                                                        const WorkloadInfo& info) const override;
 
diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk
index 010d54871a..aeac2e28b9 100644
--- a/src/backends/reference/backend.mk
+++ b/src/backends/reference/backend.mk
@@ -49,6 +49,7 @@ BACKEND_SOURCES := \
         workloads/RefComparisonWorkload.cpp \
         workloads/RefConcatWorkload.cpp \
         workloads/RefConstantWorkload.cpp \
+        workloads/RefConvertBf16ToFp32Workload.cpp \
         workloads/RefConvertFp16ToFp32Workload.cpp \
         workloads/RefConvertFp32ToFp16Workload.cpp \
         workloads/RefConvolution2dWorkload.cpp \
diff --git a/src/backends/reference/test/RefLayerSupportTests.cpp b/src/backends/reference/test/RefLayerSupportTests.cpp
index ab749c1a5c..0b549dbb77 100644
--- a/src/backends/reference/test/RefLayerSupportTests.cpp
+++ b/src/backends/reference/test/RefLayerSupportTests.cpp
@@ -116,6 +116,38 @@ BOOST_AUTO_TEST_CASE(IsConvertFp16ToFp32SupportedFp16OutputReference)
     BOOST_CHECK_EQUAL(reasonIfUnsupported, "Layer is not supported with float16 data type output");
 }
 
+BOOST_AUTO_TEST_CASE(IsConvertBf16ToFp32SupportedReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertBf16ToFp32Layer,
+      armnn::DataType::BFloat16, armnn::DataType::Float32>(reasonIfUnsupported);
+
+    BOOST_CHECK(result);
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertBf16ToFp32SupportedFp32InputReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertBf16ToFp32Layer,
+      armnn::DataType::Float32, armnn::DataType::Float32>(reasonIfUnsupported);
+
+    BOOST_CHECK(!result);
+    BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertBf16ToFp32 layer: input type not supported\n");
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertBf16ToFp32SupportedBf16OutputReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertBf16ToFp32Layer,
+      armnn::DataType::BFloat16, armnn::DataType::BFloat16>(reasonIfUnsupported);
+
+    BOOST_CHECK(!result);
+    BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertBf16ToFp32 layer: output type not supported\n");
+}
+
 BOOST_AUTO_TEST_CASE(IsConvertFp32ToFp16SupportedReference)
 {
     std::string reasonIfUnsupported;
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index 73b2a05e09..08cfa2843f 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -1142,6 +1142,9 @@ ARMNN_AUTO_TEST_CASE(LstmLayerInt16NoCifgWithPeepholeWithProjection,
 ARMNN_AUTO_TEST_CASE(LstmLayerInt16NoCifgNoPeepholeNoProjectionInt16Constant,
                      LstmLayerInt16NoCifgNoPeepholeNoProjectionInt16ConstantTest)
 
+// Convert from BFloat16 to Float32
+ARMNN_AUTO_TEST_CASE(ConvertBf16ToFp32, ConvertBf16ToFp32Test)
+
 // Convert from Float16 to Float32
 ARMNN_AUTO_TEST_CASE(SimpleConvertFp16ToFp32, SimpleConvertFp16ToFp32Test)
 // Convert from Float32 to Float16
diff --git a/src/backends/reference/workloads/CMakeLists.txt b/src/backends/reference/workloads/CMakeLists.txt
index b2d8938745..86764d8a01 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -69,6 +69,8 @@ list(APPEND armnnRefBackendWorkloads_sources
     RefConcatWorkload.hpp
     RefConstantWorkload.cpp
     RefConstantWorkload.hpp
+    RefConvertBf16ToFp32Workload.cpp
+    RefConvertBf16ToFp32Workload.hpp
     RefConvertFp16ToFp32Workload.cpp
     RefConvertFp16ToFp32Workload.hpp
     RefConvertFp32ToFp16Workload.cpp
diff --git a/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp b/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp
new file mode 100644
index 0000000000..c4b5416836
--- /dev/null
+++ b/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.cpp
@@ -0,0 +1,27 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefConvertBf16ToFp32Workload.hpp"
+#include "RefWorkloadUtils.hpp"
+
+#include <armnnUtils/FloatingPointConverter.hpp>
+
+#include <BFloat16.hpp>
+
+namespace armnn
+{
+
+void RefConvertBf16ToFp32Workload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefConvertBf16ToFp32Workload_Execute");
+
+    const BFloat16* const input = GetInputTensorDataBFloat16(0, m_Data);
+    float* const output = GetOutputTensorDataFloat(0, m_Data);
+
+    unsigned int numElements = GetTensorInfo(m_Data.m_Inputs[0]).GetNumElements();
+    armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(input, numElements, output);
+}
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.hpp b/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.hpp
new file mode 100644
index 0000000000..87cdc3e1e3
--- /dev/null
+++ b/src/backends/reference/workloads/RefConvertBf16ToFp32Workload.hpp
@@ -0,0 +1,21 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <backendsCommon/Workload.hpp>
+#include <backendsCommon/WorkloadData.hpp>
+
+namespace armnn
+{
+
+class RefConvertBf16ToFp32Workload : public BFloat16ToFloat32Workload<ConvertBf16ToFp32QueueDescriptor>
+{
+public:
+    using BFloat16ToFloat32Workload<ConvertBf16ToFp32QueueDescriptor>::BFloat16ToFloat32Workload;
+    virtual void Execute() const override;
+};
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefWorkloadUtils.hpp b/src/backends/reference/workloads/RefWorkloadUtils.hpp
index c3260c8142..69713145c0 100644
--- a/src/backends/reference/workloads/RefWorkloadUtils.hpp
+++ b/src/backends/reference/workloads/RefWorkloadUtils.hpp
@@ -12,6 +12,7 @@
 
 #include <reference/RefTensorHandle.hpp>
 
+#include <BFloat16.hpp>
 #include <Half.hpp>
 #include <boost/polymorphic_cast.hpp>
 
@@ -68,6 +69,12 @@ Half* GetOutputTensorDataHalf(unsigned int idx, const PayloadType& data)
     return GetOutputTensorData<Half>(idx, data);
 }
 
+template <typename PayloadType>
+const BFloat16* GetInputTensorDataBFloat16(unsigned int idx, const PayloadType& data)
+{
+    return GetInputTensorData<BFloat16>(idx, data);
+}
+
 ////////////////////////////////////////////
 /// u8 helpers
 ////////////////////////////////////////////
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index a0558ff06e..37d79f0fc0 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -22,6 +22,7 @@
 #include "RefConvolution2dWorkload.hpp"
 #include "RefConstantWorkload.hpp"
 #include "RefConcatWorkload.hpp"
+#include "RefConvertBf16ToFp32Workload.hpp"
 #include "RefConvertFp16ToFp32Workload.hpp"
 #include "RefConvertFp32ToFp16Workload.hpp"
 #include "RefDebugWorkload.hpp"