IVGCVSW-3835 Create Encoder and Decoder for QSymm8PerAxis

 * Add QuantizedSymm8PerAxis to armnn DataType (types.hpp) and
 * Add Quantize and Dequantize template for int8 in TypeUtils to be able to compute QSymm8 of the weight
 * Create PerAxisIterator for per-axis quantization
 * Create QSymm8PerAxisDecoder
 * Create QSymm8PerAxisEncoder

Signed-off-by: Keith Davis <keith.davis@arm.com>
Change-Id: Ibcfe0288a197b7ee50b543bdbd77b7edb8a547c2

3 files changed, 133 insertions, 3 deletions

diff --git a/src/backends/reference/workloads/BaseIterator.hpp b/src/backends/reference/workloads/BaseIterator.hpp
index 18270faf46..9fe3f15f9b 100644
--- a/src/backends/reference/workloads/BaseIterator.hpp
+++ b/src/backends/reference/workloads/BaseIterator.hpp
@@ -339,4 +339,116 @@ public:
     }
 };
 
-} //namespace armnn
+// PerAxisIterator for per-axis quantization
+template<typename T, typename Base>
+class PerAxisIterator : public Base
+{
+public:
+    // axisFactor is used to calculate axisIndex
+    PerAxisIterator(T* data = nullptr, unsigned int axisFactor = 0)
+        : m_Iterator(data), m_Start(data), m_AxisIndex(0), m_AxisFactor(axisFactor)
+    {}
+
+    // This should be called to set index for per-axis Encoder/Decoder
+    PerAxisIterator& SetIndex(unsigned int index, unsigned int axisIndex)
+    {
+         BOOST_ASSERT(m_Iterator);
+         m_Iterator = m_Start + index;
+         m_AxisIndex = axisIndex;
+         return *this;
+    }
+
+    void Reset(void* data) override
+    {
+        m_Iterator = reinterpret_cast<T*>(data);
+        m_Start = m_Iterator;
+        m_AxisIndex = 0;
+    }
+
+    PerAxisIterator& operator++() override
+    {
+        BOOST_ASSERT(m_Iterator);
+        ++m_Iterator;
+        m_AxisIndex = static_cast<unsigned int>(*m_Iterator) % m_AxisFactor;
+        return *this;
+    }
+
+    PerAxisIterator& operator+=(const unsigned int increment) override
+    {
+        BOOST_ASSERT(m_Iterator);
+        m_Iterator += increment;
+        m_AxisIndex = static_cast<unsigned int>(*m_Iterator) % m_AxisFactor;
+        return *this;
+    }
+
+    PerAxisIterator& operator-=(const unsigned int decrement) override
+    {
+        BOOST_ASSERT(m_Iterator);
+        m_Iterator -= decrement;
+        m_AxisIndex = static_cast<unsigned int>(*m_Iterator) % m_AxisFactor;
+        return *this;
+    }
+
+    PerAxisIterator& operator[](const unsigned int index) override
+    {
+        BOOST_ASSERT(m_Iterator);
+        m_Iterator = m_Start + index;
+        m_AxisIndex = static_cast<unsigned int>(*m_Iterator) % m_AxisFactor;
+        return *this;
+    }
+
+    protected:
+        T* m_Iterator;
+        T* m_Start;
+        unsigned int m_AxisIndex;
+        unsigned int m_AxisFactor;
+};
+
+class QSymm8PerAxisDecoder : public PerAxisIterator<const int8_t, Decoder<float>>
+{
+public:
+    QSymm8PerAxisDecoder(const int8_t* data, const std::vector<float>& scale, unsigned int axisFactor)
+        : PerAxisIterator(data, axisFactor), m_Scale(scale) {}
+
+    float Get() const override
+    {
+        return armnn::Dequantize(*m_Iterator, m_Scale[m_AxisIndex], 0);
+    }
+
+    // Get scale of the current value
+    float GetScale() const
+    {
+        return m_Scale[m_AxisIndex];
+    }
+
+private:
+    std::vector<float> m_Scale;
+};
+
+class QSymm8PerAxisEncoder : public PerAxisIterator<int8_t, Encoder<float>>
+{
+public:
+    QSymm8PerAxisEncoder(int8_t* data, const std::vector<float>& scale, unsigned int axisFactor)
+        : PerAxisIterator(data, axisFactor), m_Scale(scale) {}
+
+    void Set(float right)
+    {
+        *m_Iterator = armnn::Quantize<int8_t>(right, m_Scale[m_AxisIndex], 0);
+    }
+
+    float Get() const
+    {
+        return armnn::Dequantize(*m_Iterator, m_Scale[m_AxisIndex], 0);
+    }
+
+    // Get scale of the current value
+    float GetScale() const
+    {
+        return m_Scale[m_AxisIndex];
+    }
+
+private:
+    std::vector<float> m_Scale;
+};
+
+} //namespace armnn
\ No newline at end of file
diff --git a/src/backends/reference/workloads/Decoders.hpp b/src/backends/reference/workloads/Decoders.hpp
index 328a5eb0f7..dd2b28a50f 100644
--- a/src/backends/reference/workloads/Decoders.hpp
+++ b/src/backends/reference/workloads/Decoders.hpp
@@ -7,6 +7,7 @@
 
 #include "BaseIterator.hpp"
 #include "FloatingPointConverter.hpp"
+#include "TensorUtils.hpp"
 
 #include <boost/assert.hpp>
 
@@ -21,6 +22,14 @@ inline std::unique_ptr<Decoder<float>> MakeDecoder(const TensorInfo& info, const
 {
     switch(info.GetDataType())
     {
+        case armnn::DataType::QuantizedSymm8PerAxis:
+        {
+            std::pair<unsigned int, std::vector<float>> params = armnnUtils::GetPerAxisParams(info);
+            return std::make_unique<QSymm8PerAxisDecoder>(
+                static_cast<const int8_t*>(data),
+                params.second,
+                params.first);
+        }
         case DataType::QuantisedAsymm8:
         {
             return std::make_unique<QASymm8Decoder>(
@@ -55,7 +64,7 @@ inline std::unique_ptr<Decoder<float>> MakeDecoder(const TensorInfo& info, const
         }
         default:
         {
-            BOOST_ASSERT_MSG(false, "Not supported Data Type!");
+            BOOST_ASSERT_MSG(false, "Unsupported Data Type!");
             break;
         }
     }
diff --git a/src/backends/reference/workloads/Encoders.hpp b/src/backends/reference/workloads/Encoders.hpp
index 2b3a11af06..5c0cffa7ca 100644
--- a/src/backends/reference/workloads/Encoders.hpp
+++ b/src/backends/reference/workloads/Encoders.hpp
@@ -6,6 +6,7 @@
 #pragma once
 
 #include "BaseIterator.hpp"
+#include "TensorUtils.hpp"
 
 #include <boost/assert.hpp>
 
@@ -20,6 +21,14 @@ inline std::unique_ptr<Encoder<float>> MakeEncoder(const TensorInfo& info, void*
 {
     switch(info.GetDataType())
     {
+        case armnn::DataType::QuantizedSymm8PerAxis:
+        {
+            std::pair<unsigned int, std::vector<float>> params = armnnUtils::GetPerAxisParams(info);
+            return std::make_unique<QSymm8PerAxisEncoder>(
+                static_cast<int8_t*>(data),
+                params.second,
+                params.first);
+        }
         case armnn::DataType::QuantisedAsymm8:
         {
             return std::make_unique<QASymm8Encoder>(
@@ -48,7 +57,7 @@ inline std::unique_ptr<Encoder<float>> MakeEncoder(const TensorInfo& info, void*
         }
         default:
         {
-            BOOST_ASSERT_MSG(false, "Cannot encode from float. Not supported target Data Type!");
+            BOOST_ASSERT_MSG(false, "Unsupported target Data Type!");
             break;
         }
     }