IVGCVSW-4515 Add ConvertBf16ToFp32Layer and Ref workload support

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

12 files changed, 123 insertions, 0 deletions

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"