IVGCVSW-4211 Add Signed 8 bit Quantisation support into the Reference backend

!android-nn-driver:2435

Signed-off-by: Finn Williams <Finn.Williams@arm.com>
Change-Id: I10ecd4a8937725953396805f33a3562a5384c4d4

4 files changed, 62 insertions, 0 deletions

diff --git a/src/backends/reference/workloads/BaseIterator.hpp b/src/backends/reference/workloads/BaseIterator.hpp
index ca5110c2fd..ca6d3cbc60 100644
--- a/src/backends/reference/workloads/BaseIterator.hpp
+++ b/src/backends/reference/workloads/BaseIterator.hpp
@@ -137,6 +137,25 @@ private:
     const int32_t m_Offset;
 };
 
+class QSymmS8Decoder : public TypedIterator<const int8_t, Decoder<float>>
+{
+public:
+    QSymmS8Decoder(const int8_t* data, const float scale, const int32_t offset)
+            : TypedIterator(data), m_Scale(scale), m_Offset(offset) {}
+
+    QSymmS8Decoder(const float scale, const int32_t offset)
+            : QSymmS8Decoder(nullptr, scale, offset) {}
+
+    float Get() const override
+    {
+        return armnn::Dequantize(*m_Iterator, m_Scale, m_Offset);
+    }
+
+private:
+    const float m_Scale;
+    const int32_t m_Offset;
+};
+
 class QSymm16Decoder : public TypedIterator<const int16_t, Decoder<float>>
 {
 public:
@@ -245,6 +264,30 @@ private:
     const int32_t m_Offset;
 };
 
+class QSymmS8Encoder : public TypedIterator<int8_t, Encoder<float>>
+{
+public:
+    QSymmS8Encoder(int8_t* data, const float scale, const int32_t offset)
+            : TypedIterator(data), m_Scale(scale), m_Offset(offset) {}
+
+    QSymmS8Encoder(const float scale, const int32_t offset)
+            : QSymmS8Encoder(nullptr, scale, offset) {}
+
+    void Set(float right) override
+    {
+        *m_Iterator = armnn::Quantize<int8_t>(right, m_Scale, m_Offset);
+    }
+
+    float Get() const override
+    {
+        return armnn::Dequantize(*m_Iterator, m_Scale, m_Offset);
+    }
+
+private:
+    const float m_Scale;
+    const int32_t m_Offset;
+};
+
 class QSymm16Encoder : public TypedIterator<int16_t, Encoder<float>>
 {
 public:
diff --git a/src/backends/reference/workloads/Decoders.hpp b/src/backends/reference/workloads/Decoders.hpp
index b9cd7f9573..9d41c9e9e7 100644
--- a/src/backends/reference/workloads/Decoders.hpp
+++ b/src/backends/reference/workloads/Decoders.hpp
@@ -105,6 +105,13 @@ inline std::unique_ptr<Decoder<float>> MakeDecoder(const TensorInfo& info, const
         {
             return MakeSigned32Decoder(info, data);
         }
+        case DataType::QSymmS8:
+        {
+            return std::make_unique<QSymmS8Decoder>(
+            static_cast<const int8_t*>(data),
+            info.GetQuantizationScale(),
+            info.GetQuantizationOffset());
+        }
         default:
         {
             BOOST_ASSERT_MSG(false, "Unsupported Data Type!");
diff --git a/src/backends/reference/workloads/Encoders.hpp b/src/backends/reference/workloads/Encoders.hpp
index 0d578d68de..92493ed641 100644
--- a/src/backends/reference/workloads/Encoders.hpp
+++ b/src/backends/reference/workloads/Encoders.hpp
@@ -37,6 +37,13 @@ inline std::unique_ptr<Encoder<float>> MakeEncoder(const TensorInfo& info, void*
                 info.GetQuantizationScale(),
                 info.GetQuantizationOffset());
         }
+        case DataType::QSymmS8:
+        {
+            return std::make_unique<QSymmS8Encoder>(
+                    static_cast<int8_t*>(data),
+                    info.GetQuantizationScale(),
+                    info.GetQuantizationOffset());
+        }
         case armnn::DataType::QuantisedSymm16:
         {
             return std::make_unique<QSymm16Encoder>(
diff --git a/src/backends/reference/workloads/RefQuantizeWorkload.cpp b/src/backends/reference/workloads/RefQuantizeWorkload.cpp
index b7ace32e14..a78804b709 100644
--- a/src/backends/reference/workloads/RefQuantizeWorkload.cpp
+++ b/src/backends/reference/workloads/RefQuantizeWorkload.cpp
@@ -48,6 +48,11 @@ void RefQuantizeWorkload::Execute() const
             QuantizeImpl<uint8_t>(input, output, m_NumElements, m_Scale, m_Offset);
             break;
         }
+        case DataType::QSymmS8:
+        {
+            QuantizeImpl<int8_t>(input, output, m_NumElements, m_Scale, m_Offset);
+            break;
+        }
         case DataType::QuantisedSymm16:
         {
             QuantizeImpl<int16_t>(input, output, m_NumElements, m_Scale, 0);