IVGCVSW-4516 Add ConvertFp32ToBf16Layer and Ref workload support

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: I9099a4f840fb747336f77d20a0868b64e801a310

36 files changed, 390 insertions, 3 deletions

diff --git a/Android.mk b/Android.mk
index 436ae2207b..2370b3392f 100644
--- a/Android.mk
+++ b/Android.mk
@@ -134,6 +134,7 @@ LOCAL_SRC_FILES := \
         src/armnn/layers/Convolution2dLayer.cpp \
         src/armnn/layers/ConvertBf16ToFp32Layer.cpp \
         src/armnn/layers/ConvertFp16ToFp32Layer.cpp \
+        src/armnn/layers/ConvertFp32ToBf16Layer.cpp \
         src/armnn/layers/ConvertFp32ToFp16Layer.cpp \
         src/armnn/layers/DebugLayer.cpp \
         src/armnn/layers/DepthToSpaceLayer.cpp \
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3d0f518fe2..034f346d1d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -276,6 +276,8 @@ list(APPEND armnn_sources
     src/armnn/layers/ConvertBf16ToFp32Layer.hpp
     src/armnn/layers/ConvertFp16ToFp32Layer.hpp
     src/armnn/layers/ConvertFp16ToFp32Layer.cpp
+    src/armnn/layers/ConvertFp32ToBf16Layer.hpp
+    src/armnn/layers/ConvertFp32ToBf16Layer.cpp
     src/armnn/layers/ConvertFp32ToFp16Layer.hpp
     src/armnn/layers/ConvertFp32ToFp16Layer.cpp
     src/armnn/layers/DebugLayer.hpp
diff --git a/include/armnn/ILayerSupport.hpp b/include/armnn/ILayerSupport.hpp
index 17ed4bef5a..8274b05535 100644
--- a/include/armnn/ILayerSupport.hpp
+++ b/include/armnn/ILayerSupport.hpp
@@ -79,6 +79,10 @@ public:
                                               const TensorInfo& output,
                                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const = 0;
 
+    virtual bool IsConvertFp32ToBf16Supported(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 3f3eed56e7..d688257142 100644
--- a/src/armnn/InternalTypes.cpp
+++ b/src/armnn/InternalTypes.cpp
@@ -24,6 +24,7 @@ char const* GetLayerTypeAsCString(LayerType type)
         case LayerType::Constant: return "Constant";
         case LayerType::ConvertBf16ToFp32: return "ConvertBf16ToFp32";
         case LayerType::ConvertFp16ToFp32: return "ConvertFp16ToFp32";
+        case LayerType::ConvertFp32ToBf16: return "ConvertFp32ToBf16";
         case LayerType::ConvertFp32ToFp16: return "ConvertFp32ToFp16";
         case LayerType::Convolution2d: return "Convolution2d";
         case LayerType::Debug: return "Debug";
diff --git a/src/armnn/InternalTypes.hpp b/src/armnn/InternalTypes.hpp
index 9330122246..8dd9a9eb1c 100644
--- a/src/armnn/InternalTypes.hpp
+++ b/src/armnn/InternalTypes.hpp
@@ -24,6 +24,7 @@ enum class LayerType
     Constant,
     ConvertBf16ToFp32,
     ConvertFp16ToFp32,
+    ConvertFp32ToBf16,
     ConvertFp32ToFp16,
     Convolution2d,
     Debug,
diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp
index 3dde908fc3..4159f488c1 100644
--- a/src/armnn/LayersFwd.hpp
+++ b/src/armnn/LayersFwd.hpp
@@ -16,6 +16,7 @@
 #include "layers/ConstantLayer.hpp"
 #include "layers/ConvertBf16ToFp32Layer.hpp"
 #include "layers/ConvertFp16ToFp32Layer.hpp"
+#include "layers/ConvertFp32ToBf16Layer.hpp"
 #include "layers/ConvertFp32ToFp16Layer.hpp"
 #include "layers/Convolution2dLayer.hpp"
 #include "layers/DebugLayer.hpp"
@@ -102,6 +103,7 @@ DECLARE_LAYER(Concat)
 DECLARE_LAYER(Constant)
 DECLARE_LAYER(ConvertBf16ToFp32)
 DECLARE_LAYER(ConvertFp16ToFp32)
+DECLARE_LAYER(ConvertFp32ToBf16)
 DECLARE_LAYER(ConvertFp32ToFp16)
 DECLARE_LAYER(Convolution2d)
 DECLARE_LAYER(Debug)
diff --git a/src/armnn/layers/ConvertBf16ToFp32Layer.hpp b/src/armnn/layers/ConvertBf16ToFp32Layer.hpp
index 2a79a1cb65..b419e5c2a3 100644
--- a/src/armnn/layers/ConvertBf16ToFp32Layer.hpp
+++ b/src/armnn/layers/ConvertBf16ToFp32Layer.hpp
@@ -15,7 +15,6 @@ 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;
diff --git a/src/armnn/layers/ConvertFp16ToFp32Layer.hpp b/src/armnn/layers/ConvertFp16ToFp32Layer.hpp
index 03d7dfa568..e3b798beec 100644
--- a/src/armnn/layers/ConvertFp16ToFp32Layer.hpp
+++ b/src/armnn/layers/ConvertFp16ToFp32Layer.hpp
@@ -15,7 +15,6 @@ class ConvertFp16ToFp32Layer : public Layer
 {
 public:
     /// Makes a workload for the ConvertFp16ToFp32 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;
diff --git a/src/armnn/layers/ConvertFp32ToBf16Layer.cpp b/src/armnn/layers/ConvertFp32ToBf16Layer.cpp
new file mode 100644
index 0000000000..936acf61ab
--- /dev/null
+++ b/src/armnn/layers/ConvertFp32ToBf16Layer.cpp
@@ -0,0 +1,55 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ConvertFp32ToBf16Layer.hpp"
+#include "LayerCloneBase.hpp"
+
+#include <armnn/TypesUtils.hpp>
+
+#include <backendsCommon/WorkloadData.hpp>
+#include <backendsCommon/WorkloadFactory.hpp>
+
+namespace armnn
+{
+
+ConvertFp32ToBf16Layer::ConvertFp32ToBf16Layer(const char* name)
+    : Layer(1, 1, LayerType::ConvertFp32ToBf16, name)
+{
+}
+
+std::unique_ptr<IWorkload> ConvertFp32ToBf16Layer::CreateWorkload(const IWorkloadFactory& factory) const
+{
+    ConvertFp32ToBf16QueueDescriptor descriptor;
+    return factory.CreateConvertFp32ToBf16(descriptor, PrepInfoAndDesc(descriptor));
+}
+
+ConvertFp32ToBf16Layer* ConvertFp32ToBf16Layer::Clone(Graph& graph) const
+{
+    return CloneBase<ConvertFp32ToBf16Layer>(graph, GetName());
+}
+
+void ConvertFp32ToBf16Layer::ValidateTensorShapesFromInputs()
+{
+    VerifyLayerConnections(1, CHECK_LOCATION());
+
+    auto inferredShapes = InferOutputShapes({ GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape() });
+
+    BOOST_ASSERT(inferredShapes.size() == 1);
+
+    ConditionalThrowIfNotEqual<LayerValidationException>(
+        "ConvertFp32ToBf16Layer: TensorShape set on OutputSlot[0] does not match the inferred shape.",
+        GetOutputSlot(0).GetTensorInfo().GetShape(),
+        inferredShapes[0]);
+}
+
+void ConvertFp32ToBf16Layer::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("ConvertFp32ToBf16Layer should never appear in an input graph");
+}
+
+} // namespace armnn
diff --git a/src/armnn/layers/ConvertFp32ToBf16Layer.hpp b/src/armnn/layers/ConvertFp32ToBf16Layer.hpp
new file mode 100644
index 0000000000..225b03314e
--- /dev/null
+++ b/src/armnn/layers/ConvertFp32ToBf16Layer.hpp
@@ -0,0 +1,41 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <Layer.hpp>
+
+namespace armnn
+{
+
+/// This layer converts data type Float32 to BFloat16.
+class ConvertFp32ToBf16Layer : public Layer
+{
+public:
+    /// Makes a workload for the ConvertFp32ToBf16Layer type.
+    /// @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.
+    ConvertFp32ToBf16Layer* Clone(Graph& graph) const override;
+
+    /// Check if the input tensor shape(s)
+    /// will lead to a valid configuration of @ref ConvertFp32ToBf16Layer.
+    void ValidateTensorShapesFromInputs() override;
+
+    void Accept(ILayerVisitor& visitor) const override;
+
+protected:
+    /// Constructor to create a ConvertFp32ToBf16Layer.
+    /// @param [in] name Optional name for the layer.
+    ConvertFp32ToBf16Layer(const char* name);
+
+    /// Default destructor
+    ~ConvertFp32ToBf16Layer() = default;
+};
+
+} // namespace
diff --git a/src/armnn/layers/ConvertFp32ToFp16Layer.hpp b/src/armnn/layers/ConvertFp32ToFp16Layer.hpp
index 907a55f084..8bb28f84ad 100644
--- a/src/armnn/layers/ConvertFp32ToFp16Layer.hpp
+++ b/src/armnn/layers/ConvertFp32ToFp16Layer.hpp
@@ -14,7 +14,6 @@ class ConvertFp32ToFp16Layer : public Layer
 {
 public:
     /// Makes a workload for the ConvertFp32ToFp16 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;
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index 1ac08afd7a..c3c8421267 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -125,6 +125,14 @@ bool LayerSupportBase::IsConvertFp16ToFp32Supported(const TensorInfo& /*input*/,
     return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
 }
 
+bool LayerSupportBase::IsConvertFp32ToBf16Supported(const TensorInfo& /*input*/,
+                                                    const TensorInfo& /*output*/,
+                                                    Optional<std::string&> reasonIfUnsupported) const
+{
+    return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
+
 bool LayerSupportBase::IsConvertFp32ToFp16Supported(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 59e8b969bd..063983357e 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -69,6 +69,10 @@ public:
                                       const TensorInfo& output,
                                       Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsConvertFp32ToBf16Supported(const TensorInfo& input,
+                                      const TensorInfo& output,
+                                      Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsConvertFp32ToFp16Supported(
             const TensorInfo& input,
             const TensorInfo& output,
diff --git a/src/backends/backendsCommon/Workload.hpp b/src/backends/backendsCommon/Workload.hpp
index 56736cebb3..984443b79b 100644
--- a/src/backends/backendsCommon/Workload.hpp
+++ b/src/backends/backendsCommon/Workload.hpp
@@ -182,6 +182,11 @@ using BFloat16ToFloat32Workload = MultiTypedWorkload<QueueDescriptor,
                                                      armnn::DataType::Float32>;
 
 template <typename QueueDescriptor>
+using Float32ToBFloat16Workload = MultiTypedWorkload<QueueDescriptor,
+                                                     armnn::DataType::Float32,
+                                                     armnn::DataType::BFloat16>;
+
+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 81aefa94e7..bf26056a97 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -2039,6 +2039,29 @@ void ConvertBf16ToFp32QueueDescriptor::Validate(const WorkloadInfo& workloadInfo
     ValidateTensorShapesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output");
 }
 
+void ConvertFp32ToBf16QueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+    const std::string descriptorName{"ConvertFp32ToBf16QueueDescriptor"};
+
+    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::Float32)
+    {
+        throw InvalidArgumentException(descriptorName + ": Input tensor type must be Float32.");
+    }
+
+    if (outputTensorInfo.GetDataType() != DataType::BFloat16)
+    {
+        throw InvalidArgumentException(descriptorName + ": Output tensor type must be BFloat16.");
+    }
+
+    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 9c392d3219..85bda5469a 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -412,6 +412,11 @@ struct ConvertBf16ToFp32QueueDescriptor : QueueDescriptor
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct ConvertFp32ToBf16QueueDescriptor : 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 d932eef49f..5854bece3c 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -173,6 +173,13 @@ bool IWorkloadFactory::IsLayerSupported(const BackendId& backendId,
             result = layerSupportObject->IsConvertFp16ToFp32Supported(input, output, reason);
             break;
         }
+        case LayerType::ConvertFp32ToBf16:
+        {
+            const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+            result = layerSupportObject->IsConvertFp32ToBf16Supported(input, output, reason);
+            break;
+        }
         case LayerType::ConvertFp32ToFp16:
         {
             const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
@@ -1163,6 +1170,12 @@ std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp16ToFp32(const Conve
     return std::unique_ptr<IWorkload>();
 }
 
+std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp32ToBf16(const ConvertFp32ToBf16QueueDescriptor& /*desc*/,
+                                                                     const WorkloadInfo& /*info*/) const
+{
+    return std::unique_ptr<IWorkload>();
+}
+
 std::unique_ptr<IWorkload> IWorkloadFactory::CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& /*desc*/,
                                                                      const WorkloadInfo& /*info*/) const
 {
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index 8c22452f2e..0fc7ab9bce 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -85,6 +85,9 @@ public:
     virtual std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor,
                                                                const WorkloadInfo& info) const;
 
+    virtual std::unique_ptr<IWorkload> CreateConvertFp32ToBf16(const ConvertFp32ToBf16QueueDescriptor& descriptor,
+                                                               const WorkloadInfo& info) const;
+
     virtual std::unique_ptr<IWorkload> CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor,
                                                                const WorkloadInfo& info) const;
 
diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 22de3db77c..ba5518010d 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -46,6 +46,7 @@ COMMON_TEST_SOURCES := \
     test/layerTests/Conv2dTestImpl.cpp \
     test/layerTests/ConvertBf16ToFp32TestImpl.cpp \
     test/layerTests/ConvertFp16ToFp32TestImpl.cpp \
+    test/layerTests/ConvertFp32ToBf16TestImpl.cpp \
     test/layerTests/ConvertFp32ToFp16TestImpl.cpp \
     test/layerTests/DebugTestImpl.cpp \
     test/layerTests/DepthToSpaceTestImpl.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index ea214de771..bfaca6c44b 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -73,6 +73,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     layerTests/ConvertBf16ToFp32TestImpl.hpp
     layerTests/ConvertFp16ToFp32TestImpl.cpp
     layerTests/ConvertFp16ToFp32TestImpl.hpp
+    layerTests/ConvertFp32ToBf16TestImpl.cpp
+    layerTests/ConvertFp32ToBf16TestImpl.hpp
     layerTests/ConvertFp32ToFp16TestImpl.cpp
     layerTests/ConvertFp32ToFp16TestImpl.hpp
     layerTests/DebugTestImpl.cpp
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index a070ac05e8..d646847cbc 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -447,6 +447,8 @@ DECLARE_LAYER_POLICY_1_PARAM(ConvertBf16ToFp32)
 
 DECLARE_LAYER_POLICY_1_PARAM(ConvertFp16ToFp32)
 
+DECLARE_LAYER_POLICY_1_PARAM(ConvertFp32ToBf16)
+
 DECLARE_LAYER_POLICY_1_PARAM(ConvertFp32ToFp16)
 
 DECLARE_LAYER_POLICY_2_PARAM(Convolution2d)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 1c6277a333..600a261fe3 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -15,6 +15,7 @@
 #include <backendsCommon/test/layerTests/ConcatTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertBf16ToFp32TestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertFp16ToFp32TestImpl.hpp>
+#include <backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConvertFp32ToFp16TestImpl.hpp>
 #include <backendsCommon/test/layerTests/Conv2dTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ConstantTestImpl.hpp>
diff --git a/src/backends/backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.cpp b/src/backends/backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.cpp
new file mode 100644
index 0000000000..66eb4ee28c
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.cpp
@@ -0,0 +1,77 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ConvertFp32ToBf16TestImpl.hpp"
+
+#include <backendsCommon/test/TensorCopyUtils.hpp>
+#include <backendsCommon/test/WorkloadTestUtils.hpp>
+
+#include <test/TensorHelpers.hpp>
+
+LayerTestResult<armnn::BFloat16, 4> ConvertFp32ToBf16Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    IgnoreUnused(memoryManager);
+
+    const armnn::TensorInfo inputTensorInfo({1, 2, 4, 3}, armnn::DataType::Float32);
+    const armnn::TensorInfo outputTensorInfo({1, 2, 4, 3}, armnn::DataType::BFloat16);
+
+    auto input = MakeTensor<float, 4>(inputTensorInfo,
+        { -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,
+          3.8f, // 0x40733333 Round down
+          3.1055E+29f, // 0x707ADC3C Round up
+          9.149516E-10f, // 0x307B7FFF Round down
+          -3.8f, // 0xC0733333 Round down
+          -3.1055E+29f, // 0xF07ADC3C Round up
+          -9.149516E-10f // 0xB07B7FFF Round down
+        });
+
+    std::vector<armnn::BFloat16> outputValues = 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,
+          3.796875f, // 0x4073
+          3.1072295E29f, // 0x707B
+          9.131327E-10f, // 0x307B
+          -3.796875f, // 0xC073
+          -3.1072295E29f, // 0xF07B
+          -9.131327E-10f // 0xB07B
+        },
+        1.0f, 0);
+
+    LayerTestResult<armnn::BFloat16, 4> ret(outputTensorInfo);
+    ret.outputExpected = MakeTensor<armnn::BFloat16, 4>(outputTensorInfo, outputValues);
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::ConvertFp32ToBf16QueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConvertFp32ToBf16(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/ConvertFp32ToBf16TestImpl.hpp b/src/backends/backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.hpp
new file mode 100644
index 0000000000..4c6125f585
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/ConvertFp32ToBf16TestImpl.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<armnn::BFloat16, 4> ConvertFp32ToBf16Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index a4f4efd92a..9f22b9ef0e 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -411,6 +411,21 @@ bool RefLayerSupport::IsConvertFp16ToFp32Supported(const TensorInfo& input,
                                           &FalseFuncU8<>));
 }
 
+bool RefLayerSupport::IsConvertFp32ToBf16Supported(const TensorInfo& input,
+                                                   const TensorInfo& output,
+                                                   Optional<std::string&> reasonIfUnsupported) const
+{
+    bool supported = true;
+
+    supported &= CheckSupportRule(TypeIs(input, DataType::Float32), reasonIfUnsupported,
+                                  "Reference for ConvertFp32ToBf16 layer: input type not supported");
+
+    supported &= CheckSupportRule(TypeIs(output, DataType::BFloat16), reasonIfUnsupported,
+                                  "Reference for ConvertFp32ToBf16 layer: output type not supported");
+
+    return supported;
+}
+
 bool RefLayerSupport::IsConvertFp32ToFp16Supported(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 ff3478142f..30f45c37f2 100644
--- a/src/backends/reference/RefLayerSupport.hpp
+++ b/src/backends/reference/RefLayerSupport.hpp
@@ -68,6 +68,10 @@ public:
                                       const TensorInfo& output,
                                       Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsConvertFp32ToBf16Supported(const TensorInfo& input,
+                                      const TensorInfo& output,
+                                      Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsConvertFp32ToFp16Supported(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 c9fc626bc1..5d3775a59d 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -194,6 +194,13 @@ std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConvertFp16ToFp32(
     return std::make_unique<RefConvertFp16ToFp32Workload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConvertFp32ToBf16(
+    const ConvertFp32ToBf16QueueDescriptor& descriptor,
+    const WorkloadInfo& info) const
+{
+    return std::make_unique<RefConvertFp32ToBf16Workload>(descriptor, info);
+}
+
 std::unique_ptr<IWorkload> RefWorkloadFactory::CreateConvertFp32ToFp16(
     const ConvertFp32ToFp16QueueDescriptor& descriptor,
     const WorkloadInfo& info) const
diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp
index 119605b998..9a53ae2e5a 100644
--- a/src/backends/reference/RefWorkloadFactory.hpp
+++ b/src/backends/reference/RefWorkloadFactory.hpp
@@ -92,6 +92,9 @@ public:
     std::unique_ptr<IWorkload> CreateConvertFp16ToFp32(const ConvertFp16ToFp32QueueDescriptor& descriptor,
                                                        const WorkloadInfo& info) const override;
 
+    std::unique_ptr<IWorkload> CreateConvertFp32ToBf16(const ConvertFp32ToBf16QueueDescriptor& descriptor,
+                                                       const WorkloadInfo& info) const override;
+
     std::unique_ptr<IWorkload> CreateConvertFp32ToFp16(const ConvertFp32ToFp16QueueDescriptor& descriptor,
                                                        const WorkloadInfo& info) const override;
 
diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk
index aeac2e28b9..239863f2c7 100644
--- a/src/backends/reference/backend.mk
+++ b/src/backends/reference/backend.mk
@@ -51,6 +51,7 @@ BACKEND_SOURCES := \
         workloads/RefConstantWorkload.cpp \
         workloads/RefConvertBf16ToFp32Workload.cpp \
         workloads/RefConvertFp16ToFp32Workload.cpp \
+        workloads/RefConvertFp32ToBf16Workload.cpp \
         workloads/RefConvertFp32ToFp16Workload.cpp \
         workloads/RefConvolution2dWorkload.cpp \
         workloads/RefDebugWorkload.cpp \
diff --git a/src/backends/reference/test/RefLayerSupportTests.cpp b/src/backends/reference/test/RefLayerSupportTests.cpp
index 0b549dbb77..f0c69f92cc 100644
--- a/src/backends/reference/test/RefLayerSupportTests.cpp
+++ b/src/backends/reference/test/RefLayerSupportTests.cpp
@@ -148,6 +148,38 @@ BOOST_AUTO_TEST_CASE(IsConvertBf16ToFp32SupportedBf16OutputReference)
     BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertBf16ToFp32 layer: output type not supported\n");
 }
 
+BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertFp32ToBf16Layer,
+      armnn::DataType::Float32, armnn::DataType::BFloat16>(reasonIfUnsupported);
+
+    BOOST_CHECK(result);
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedBf16InputReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertFp32ToBf16Layer,
+      armnn::DataType::BFloat16, armnn::DataType::BFloat16>(reasonIfUnsupported);
+
+    BOOST_CHECK(!result);
+    BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertFp32ToBf16 layer: input type not supported\n");
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertFp32ToBf16SupportedFp32OutputReference)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::RefWorkloadFactory, armnn::ConvertFp32ToBf16Layer,
+      armnn::DataType::Float32, armnn::DataType::Float32>(reasonIfUnsupported);
+
+    BOOST_CHECK(!result);
+    BOOST_CHECK_EQUAL(reasonIfUnsupported, "Reference for ConvertFp32ToBf16 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 08cfa2843f..f5c1c65c88 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -1145,6 +1145,9 @@ ARMNN_AUTO_TEST_CASE(LstmLayerInt16NoCifgNoPeepholeNoProjectionInt16Constant,
 // Convert from BFloat16 to Float32
 ARMNN_AUTO_TEST_CASE(ConvertBf16ToFp32, ConvertBf16ToFp32Test)
 
+// Convert from Float32 to BFloat16
+ARMNN_AUTO_TEST_CASE(ConvertFp32ToBf16, ConvertFp32ToBf16Test)
+
 // 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 86764d8a01..9f3880e077 100644
--- a/src/backends/reference/workloads/CMakeLists.txt
+++ b/src/backends/reference/workloads/CMakeLists.txt
@@ -73,6 +73,8 @@ list(APPEND armnnRefBackendWorkloads_sources
     RefConvertBf16ToFp32Workload.hpp
     RefConvertFp16ToFp32Workload.cpp
     RefConvertFp16ToFp32Workload.hpp
+    RefConvertFp32ToBf16Workload.cpp
+    RefConvertFp32ToBf16Workload.hpp
     RefConvertFp32ToFp16Workload.cpp
     RefConvertFp32ToFp16Workload.hpp
     RefConvolution2dWorkload.cpp
diff --git a/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp
new file mode 100644
index 0000000000..181b236e83
--- /dev/null
+++ b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.cpp
@@ -0,0 +1,27 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "RefConvertFp32ToBf16Workload.hpp"
+#include "RefWorkloadUtils.hpp"
+
+#include <armnnUtils/FloatingPointConverter.hpp>
+
+#include <BFloat16.hpp>
+
+namespace armnn
+{
+
+void RefConvertFp32ToBf16Workload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefConvertFp32ToBf16Workload_Execute");
+
+    const float* const input = GetInputTensorDataFloat(0, m_Data);
+    BFloat16* const output = GetOutputTensorDataBFloat16(0, m_Data);
+
+    unsigned int numElements = GetTensorInfo(m_Data.m_Inputs[0]).GetNumElements();
+    armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(input, numElements, output);
+}
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.hpp
new file mode 100644
index 0000000000..409603bb6c
--- /dev/null
+++ b/src/backends/reference/workloads/RefConvertFp32ToBf16Workload.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 RefConvertFp32ToBf16Workload : public Float32ToBFloat16Workload<ConvertFp32ToBf16QueueDescriptor>
+{
+public:
+    using Float32ToBFloat16Workload<ConvertFp32ToBf16QueueDescriptor>::Float32ToBFloat16Workload;
+    virtual void Execute() const override;
+};
+
+} //namespace armnn
diff --git a/src/backends/reference/workloads/RefWorkloadUtils.hpp b/src/backends/reference/workloads/RefWorkloadUtils.hpp
index 69713145c0..f1b31571db 100644
--- a/src/backends/reference/workloads/RefWorkloadUtils.hpp
+++ b/src/backends/reference/workloads/RefWorkloadUtils.hpp
@@ -75,6 +75,12 @@ const BFloat16* GetInputTensorDataBFloat16(unsigned int idx, const PayloadType&
     return GetInputTensorData<BFloat16>(idx, data);
 }
 
+template <typename PayloadType>
+BFloat16* GetOutputTensorDataBFloat16(unsigned int idx, const PayloadType& data)
+{
+    return GetOutputTensorData<BFloat16>(idx, data);
+}
+
 ////////////////////////////////////////////
 /// u8 helpers
 ////////////////////////////////////////////
diff --git a/src/backends/reference/workloads/RefWorkloads.hpp b/src/backends/reference/workloads/RefWorkloads.hpp
index 37d79f0fc0..cbfade3c02 100644
--- a/src/backends/reference/workloads/RefWorkloads.hpp
+++ b/src/backends/reference/workloads/RefWorkloads.hpp
@@ -24,6 +24,7 @@
 #include "RefConcatWorkload.hpp"
 #include "RefConvertBf16ToFp32Workload.hpp"
 #include "RefConvertFp16ToFp32Workload.hpp"
+#include "RefConvertFp32ToBf16Workload.hpp"
 #include "RefConvertFp32ToFp16Workload.hpp"
 #include "RefDebugWorkload.hpp"
 #include "RefDepthToSpaceWorkload.hpp"