IVGCVSW-4633 Add conversion of BF16 support to Neon

 * Add NeonConvertBf16ToFp32Workload
 * Add NeonConvertFp32ToBf16Workload
 * Add BFloat16 type support to NeonConstantWorkload and NeonTensorHandle
 * Add ConvertBf16ToFp32Weight when ConvertBf16ToFp32Layer is added
 * Unit tests

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: Id5b44a203add5e0c98c1ca4e2162115741b56644

1 files changed, 43 insertions, 0 deletions

diff --git a/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp b/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp
new file mode 100644
index 0000000000..79d1f22313
--- /dev/null
+++ b/src/backends/neon/workloads/NeonConvertBf16ToFp32Workload.cpp
@@ -0,0 +1,43 @@
+//
+// Copyright © 2020 Arm Ltd. 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("NeonConvertBf16ToFp32Workload_Execute");
+
+    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);
+    }
+}
+
+} //namespace armnn