IVGCVSW-7214 Disable BF16-Turbo-Mode and remove conversion layers

 - Remove Bf16ToFp32 Conversion Layer
 - Remove Fp32ToBf16 Conversion Layer
 - Remove B16 Conversion tests
 * Throw exception if m_ReduceFp32ToBf16 optimzer option is set to true
 * Provide comments to enable fast math in order to use bf16
 * Update docs to inform users to enable fast math for bf16

 Execute Network Changes
 * Require bf16_turbo_mode to also have fast_math_enabled set to true
 - Remove setting m_ReduceFp32ToBf16 optimizer option

Signed-off-by: Ryan OShea <ryan.oshea3@arm.com>
Change-Id: Ibaa6da9d29c96a1ce32ff5196b0847fde9f04a1c

6 files changed, 0 insertions, 231 deletions

diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt
index dd09ecf015..a3eb883079 100644
--- a/src/backends/neon/workloads/CMakeLists.txt
+++ b/src/backends/neon/workloads/CMakeLists.txt
@@ -28,12 +28,8 @@ list(APPEND armnnNeonBackendWorkloads_sources
     NeonConcatWorkload.hpp
     NeonConstantWorkload.cpp
     NeonConstantWorkload.hpp
-    NeonConvertBf16ToFp32Workload.cpp
-    NeonConvertBf16ToFp32Workload.hpp
     NeonConvertFp16ToFp32Workload.cpp
     NeonConvertFp16ToFp32Workload.hpp
-    NeonConvertFp32ToBf16Workload.cpp
-    NeonConvertFp32ToBf16Workload.hpp
     NeonConvertFp32ToFp16Workload.cpp
     NeonConvertFp32ToFp16Workload.hpp
     NeonConvolution2dWorkload.cpp
diff --git a/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp b/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp
deleted file mode 100644
index 7a2ff9ac1a..0000000000
--- a/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp
+++ /dev/null
@@ -1,81 +0,0 @@
-//
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "NeonConvertBf16ToFp32Workload.hpp"
-
-#include <armnnUtils/FloatingPointConverter.hpp>
-
-#include <BFloat16.hpp>
-
-#include <backendsCommon/WorkloadUtils.hpp>
-
-namespace armnn
-{
-
-NeonConvertBf16ToFp32Workload::NeonConvertBf16ToFp32Workload(const ConvertBf16ToFp32QueueDescriptor& descriptor,
-                                                             const WorkloadInfo& info)
-     : BFloat16ToFloat32Workload<ConvertBf16ToFp32QueueDescriptor>(descriptor, info)
-{
-    this->m_Data.ValidateInputsOutputs("NeonConvertBf16ToFp32Workload", 1, 1);
-    GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
-}
-
-void NeonConvertBf16ToFp32Workload::Execute() const
-{
-    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvertBf16ToFp32Workload_Execute", this->GetGuid());
-
-    auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
-        {
-            auto input = reinterpret_cast<const BFloat16*>(src);
-            auto output = reinterpret_cast<float*>(dst);
-            size_t numElements = size/2; // 2 bytes per Bf16
-            armnnUtils::FloatingPointConverter::ConvertBFloat16ToFloat32(input, numElements, output);
-        };
-
-    for (const auto& pair : m_TensorHandlePairs)
-    {
-        CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
-    }
-}
-
-void NeonConvertBf16ToFp32Workload::ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
-{
-    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
-    this->m_Data.m_Inputs[slot] = tensorHandle;
-    try
-    {
-        Reconfigure();
-    }
-    catch(armnn::UnimplementedException& e)
-    {
-        // Cannot reconfigure, revert the slot back and throw the exception.
-        this->m_Data.m_Inputs[slot] = backupHandle;
-        throw e;
-    }
-}
-
-// Replace output tensor handle with the given TensorHandle
-void NeonConvertBf16ToFp32Workload::ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
-{
-    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
-    this->m_Data.m_Inputs[slot] = tensorHandle;
-    try
-    {
-        Reconfigure();
-    }
-    catch(armnn::UnimplementedException& e)
-    {
-        // Cannot reconfigure, revert the slot back and throw the exception.
-        this->m_Data.m_Inputs[slot] = backupHandle;
-        throw e;
-    }
-}
-
-void NeonConvertBf16ToFp32Workload::Reconfigure()
-{
-    throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
-}
-
-} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.hpp b/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.hpp
deleted file mode 100644
index 9d44ad2cac..0000000000
--- a/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.hpp
+++ /dev/null
@@ -1,31 +0,0 @@
-//
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <armnn/backends/Workload.hpp>
-#include <armnn/backends/WorkloadData.hpp>
-#include <neon/workloads/NeonWorkloadUtils.hpp>
-
-namespace armnn
-{
-
-class NeonConvertBf16ToFp32Workload : public BFloat16ToFloat32Workload<ConvertBf16ToFp32QueueDescriptor>
-{
-public:
-    NeonConvertBf16ToFp32Workload(const ConvertBf16ToFp32QueueDescriptor& descriptor, const WorkloadInfo& info);
-    virtual void Execute() const override;
-    // Replace input tensor handle with the given TensorHandle
-    void ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot) override;
-
-    // Replace output tensor handle with the given TensorHandle
-    void ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot) override;
-private:
-    using TensorHandlePair = std::pair<const ITensorHandle*, ITensorHandle*>;
-    std::vector<TensorHandlePair> m_TensorHandlePairs;
-    virtual void Reconfigure();
-};
-
-} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.cpp b/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.cpp
deleted file mode 100644
index acd1a1ea8f..0000000000
--- a/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.cpp
+++ /dev/null
@@ -1,82 +0,0 @@
-//
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#include "NeonConvertFp32ToBf16Workload.hpp"
-
-#include <BFloat16.hpp>
-#include <Profiling.hpp>
-
-#include <armnnUtils/FloatingPointConverter.hpp>
-
-#include <backendsCommon/WorkloadUtils.hpp>
-
-namespace armnn
-{
-
-NeonConvertFp32ToBf16Workload::NeonConvertFp32ToBf16Workload(const ConvertFp32ToBf16QueueDescriptor& descriptor,
-                                                             const WorkloadInfo& info)
-    : Float32ToBFloat16Workload<ConvertFp32ToBf16QueueDescriptor>(descriptor, info)
-{
-    this->m_Data.ValidateInputsOutputs("NeonConvertFp32ToBf16Workload", 1, 1);
-    GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
-}
-
-void NeonConvertFp32ToBf16Workload::Execute() const
-{
-    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvertFp32ToBf16Workload_Execute", this->GetGuid());
-
-    auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
-        {
-            auto input = reinterpret_cast<const float*>(src);
-            auto output = reinterpret_cast<BFloat16*>(dst);
-            size_t numElements = size/2; // 2 bytes per bf16
-            armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(input, numElements, output);
-        };
-
-    for (const auto& pair : m_TensorHandlePairs)
-    {
-        CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
-    }
-}
-
-void NeonConvertFp32ToBf16Workload::ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
-{
-    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
-    this->m_Data.m_Inputs[slot] = tensorHandle;
-    try
-    {
-        Reconfigure();
-    }
-    catch(armnn::UnimplementedException& e)
-    {
-        // Cannot reconfigure, revert the slot back and throw the exception.
-        this->m_Data.m_Inputs[slot] = backupHandle;
-        throw e;
-    }
-}
-
-// Replace output tensor handle with the given TensorHandle
-void NeonConvertFp32ToBf16Workload::ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
-{
-    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
-    this->m_Data.m_Inputs[slot] = tensorHandle;
-    try
-    {
-        Reconfigure();
-    }
-    catch(armnn::UnimplementedException& e)
-    {
-        // Cannot reconfigure, revert the slot back and throw the exception.
-        this->m_Data.m_Inputs[slot] = backupHandle;
-        throw e;
-    }
-}
-
-void NeonConvertFp32ToBf16Workload::Reconfigure()
-{
-    throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
-}
-
-} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.hpp b/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.hpp
deleted file mode 100644
index 2304f8a1d4..0000000000
--- a/src/backends/neon/workloads/NeonConvertFp32ToBf16Workload.hpp
+++ /dev/null
@@ -1,31 +0,0 @@
-//
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
-// SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include <armnn/backends/Workload.hpp>
-#include <armnn/backends/WorkloadData.hpp>
-#include <neon/workloads/NeonWorkloadUtils.hpp>
-
-namespace armnn
-{
-
-class NeonConvertFp32ToBf16Workload : public Float32ToBFloat16Workload<ConvertFp32ToBf16QueueDescriptor>
-{
-public:
-    NeonConvertFp32ToBf16Workload(const ConvertFp32ToBf16QueueDescriptor& descriptor, const WorkloadInfo& info);
-    virtual void Execute() const override;
-    // Replace input tensor handle with the given TensorHandle
-    void ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot) override;
-
-    // Replace output tensor handle with the given TensorHandle
-    void ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot) override;
-private:
-    using TensorHandlePair = std::pair<const ITensorHandle*, ITensorHandle*>;
-    std::vector<TensorHandlePair> m_TensorHandlePairs;
-    virtual void Reconfigure();
-};
-
-} //namespace armnn
diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp
index c9c5421804..01fd2f7dba 100644
--- a/src/backends/neon/workloads/NeonWorkloads.hpp
+++ b/src/backends/neon/workloads/NeonWorkloads.hpp
@@ -16,9 +16,7 @@
 #include "NeonComparisonWorkload.hpp"
 #include "NeonConcatWorkload.hpp"
 #include "NeonConstantWorkload.hpp"
-#include "NeonConvertBf16ToFp32Workload.hpp"
 #include "NeonConvertFp16ToFp32Workload.hpp"
-#include "NeonConvertFp32ToBf16Workload.hpp"
 #include "NeonConvertFp32ToFp16Workload.hpp"
 #include "NeonConvolution2dWorkload.hpp"
 #include "NeonConvolution3dWorkload.hpp"