IVGCVSW-4517 Implement BFloat16 Encoder and Decoder

 * Add ConvertFloat32ToBFloat16
 * Add ConvertBFloat16ToFloat32
 * Add BFloat16Encoder
 * Add BFloat16Decoder
 * Unit tests

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: I198888384c923aba28cfbed09a02edc6f8194b3e

2 files changed, 32 insertions, 1 deletions

diff --git a/src/armnnUtils/BFloat16.hpp b/src/armnnUtils/BFloat16.hpp
index bb56b7d37c..965fc31c17 100644
--- a/src/armnnUtils/BFloat16.hpp
+++ b/src/armnnUtils/BFloat16.hpp
@@ -6,7 +6,7 @@
 #pragma once
 
 #include <ostream>
-#include <math.h>
+#include <cmath>
 #include <stdint.h>
 
 namespace armnn
diff --git a/src/armnnUtils/FloatingPointConverter.cpp b/src/armnnUtils/FloatingPointConverter.cpp
index 3bdde11eb8..e9b338ac7c 100644
--- a/src/armnnUtils/FloatingPointConverter.cpp
+++ b/src/armnnUtils/FloatingPointConverter.cpp
@@ -5,6 +5,7 @@
 
 #include <armnnUtils/FloatingPointConverter.hpp>
 
+#include "BFloat16.hpp"
 #include "Half.hpp"
 
 #include <boost/assert.hpp>
@@ -42,4 +43,34 @@ void FloatingPointConverter::ConvertFloat16To32(const void* srcFloat16Buffer,
     }
 }
 
+void FloatingPointConverter::ConvertFloat32ToBFloat16(const float* srcFloat32Buffer,
+                                                      size_t numElements,
+                                                      void* dstBFloat16Buffer)
+{
+    BOOST_ASSERT(srcFloat32Buffer != nullptr);
+    BOOST_ASSERT(dstBFloat16Buffer != nullptr);
+
+    armnn::BFloat16* bf16 = reinterpret_cast<armnn::BFloat16*>(dstBFloat16Buffer);
+
+    for (size_t i = 0; i < numElements; i++)
+    {
+        bf16[i] = armnn::BFloat16(srcFloat32Buffer[i]);
+    }
+}
+
+void FloatingPointConverter::ConvertBFloat16ToFloat32(const void* srcBFloat16Buffer,
+                                                      size_t numElements,
+                                                      float* dstFloat32Buffer)
+{
+    BOOST_ASSERT(srcBFloat16Buffer != nullptr);
+    BOOST_ASSERT(dstFloat32Buffer != nullptr);
+
+    const armnn::BFloat16* bf16 = reinterpret_cast<const armnn::BFloat16*>(srcBFloat16Buffer);
+
+    for (size_t i = 0; i < numElements; i++)
+    {
+        dstFloat32Buffer[i] = bf16[i].toFloat32();
+    }
+}
+
 } //namespace armnnUtils