IVGCVSW-3837 Add support for per-axis quantization to reference Convolution2d workload

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: I0ac08ba4864d48e6f64c4ac645dad8ea850be112

4 files changed, 49 insertions, 18 deletions

diff --git a/src/backends/reference/RefLayerSupport.cpp b/src/backends/reference/RefLayerSupport.cpp
index 716e8d9492..4252fecf35 100644
--- a/src/backends/reference/RefLayerSupport.cpp
+++ b/src/backends/reference/RefLayerSupport.cpp
@@ -433,11 +433,12 @@ bool RefLayerSupport::IsConvolution2dSupported(const TensorInfo& input,
     bool supported = true;
 
     // Define supported types.
-    std::array<DataType,4> supportedTypes = {
-            DataType::Float32,
-            DataType::Float16,
-            DataType::QuantisedAsymm8,
-            DataType::QuantisedSymm16
+    std::array<DataType,4> supportedTypes =
+    {
+        DataType::Float32,
+        DataType::Float16,
+        DataType::QuantisedAsymm8,
+        DataType::QuantisedSymm16
     };
 
     supported &= CheckSupportRule(TypeAnyOf(input, supportedTypes), reasonIfUnsupported,
@@ -446,22 +447,39 @@ bool RefLayerSupport::IsConvolution2dSupported(const TensorInfo& input,
     supported &= CheckSupportRule(TypeAnyOf(output, supportedTypes), reasonIfUnsupported,
                                   "Reference convolution2d: output is not a supported type.");
 
-    supported &= CheckSupportRule(TypeAnyOf(weights, supportedTypes), reasonIfUnsupported,
-                                  "Reference convolution2d: weights is not a supported type.");
-
     supported &= CheckSupportRule(TypesAreEqual(input, output), reasonIfUnsupported,
                                   "Reference convolution2d: input and output types mismatched.");
 
-    supported &= CheckSupportRule(TypesAreEqual(input, weights), reasonIfUnsupported,
-                                  "Reference convolution2d: input and weights types mismatched.");
+    const DataType inputType = input.GetDataType();
+    if (inputType == DataType::QuantisedAsymm8)
+    {
+        std::array<DataType, 2> supportedWeightTypes =
+        {
+            DataType::QuantisedAsymm8,
+            DataType::QuantizedSymm8PerAxis
+        };
+
+        supported &= CheckSupportRule(TypeAnyOf(weights, supportedWeightTypes), reasonIfUnsupported,
+                                      "Reference convolution2d: weights type not supported for quantized input.");
+    }
+    else
+    {
+        supported &= CheckSupportRule(TypeAnyOf(weights, supportedTypes), reasonIfUnsupported,
+                                      "Reference convolution2d: weights is not a supported type.");
+
+        supported &= CheckSupportRule(TypesAreEqual(input, weights), reasonIfUnsupported,
+                                      "Reference convolution2d: input and weights types mismatched.");
+    }
 
     if (biases.has_value())
     {
-        std::array<DataType,3> biasesSupportedTypes = {
-                DataType::Float32,
-                DataType::Float16,
-                DataType::Signed32
+        std::array<DataType,3> biasesSupportedTypes =
+        {
+            DataType::Float32,
+            DataType::Float16,
+            DataType::Signed32
         };
+
         supported &= CheckSupportRule(TypeAnyOf(biases.value(), biasesSupportedTypes), reasonIfUnsupported,
                                       "Reference convolution2d: biases is not a supported type.");
     }
diff --git a/src/backends/reference/test/RefLayerTests.cpp b/src/backends/reference/test/RefLayerTests.cpp
index 2c38ed52cd..c407828f06 100644
--- a/src/backends/reference/test/RefLayerTests.cpp
+++ b/src/backends/reference/test/RefLayerTests.cpp
@@ -145,6 +145,8 @@ ARMNN_AUTO_TEST_CASE(Convolution2d2x2Dilation2x2Padding2x2Stride3x3NhwcInt16,
                      false,
                      DataLayout::NHWC)
 
+ARMNN_AUTO_TEST_CASE(Convolution2dPerAxisQuantTestNchw, Convolution2dPerAxisQuantTest, DataLayout::NCHW);
+ARMNN_AUTO_TEST_CASE(Convolution2dPerAxisQuantTestNhwc, Convolution2dPerAxisQuantTest, DataLayout::NHWC);
 
 // Depthwise Convolution
 ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2d, DepthwiseConvolution2dTest, true, DataLayout::NCHW)
diff --git a/src/backends/reference/workloads/BaseIterator.hpp b/src/backends/reference/workloads/BaseIterator.hpp
index 50475312a5..95a31fbdd6 100644
--- a/src/backends/reference/workloads/BaseIterator.hpp
+++ b/src/backends/reference/workloads/BaseIterator.hpp
@@ -11,6 +11,7 @@
 #include <ResolveType.hpp>
 
 #include <boost/assert.hpp>
+#include <boost/core/ignore_unused.hpp>
 
 namespace armnn
 {
@@ -22,6 +23,8 @@ public:
 
     virtual ~BaseIterator() {}
 
+    virtual BaseIterator& SetIndex(unsigned int index, unsigned int axisIndex = 0) = 0;
+
     virtual BaseIterator& operator++() = 0;
 
     virtual BaseIterator& operator+=(const unsigned int increment) = 0;
@@ -101,6 +104,14 @@ public:
         return *this;
     }
 
+    TypedIterator& SetIndex(unsigned int index, unsigned int axisIndex = 0) override
+    {
+        boost::ignore_unused(axisIndex);
+        BOOST_ASSERT(m_Iterator);
+        m_Iterator = m_Start + index;
+        return *this;
+    }
+
 protected:
     T* m_Iterator;
     T* m_Start;
@@ -350,7 +361,7 @@ public:
     {}
 
     // This should be called to set index for per-axis Encoder/Decoder
-    PerAxisIterator& SetIndex(unsigned int index, unsigned int axisIndex)
+    PerAxisIterator& SetIndex(unsigned int index, unsigned int axisIndex) override
     {
          BOOST_ASSERT(m_Iterator);
          m_Iterator = m_Start + index;
diff --git a/src/backends/reference/workloads/ConvImpl.cpp b/src/backends/reference/workloads/ConvImpl.cpp
index 92e3b2d7dd..0c13e3ba0d 100644
--- a/src/backends/reference/workloads/ConvImpl.cpp
+++ b/src/backends/reference/workloads/ConvImpl.cpp
@@ -165,7 +165,7 @@ void Convolve(const TensorShape& rInputShape,
                                     }
                                 }
 
-                                rFilterDecoder[filterIndex];
+                                rFilterDecoder.SetIndex(filterIndex, cOutput);
                                 float filterValue = rFilterDecoder.Get();
 
                                 unsigned int yInput = yOutput * yStride + yFilter * yDilation;
@@ -211,7 +211,7 @@ void Convolve(const TensorShape& rInputShape,
 
                     if (biasEnabled)
                     {
-                        (*pBiasDecoder)[cOutput];
+                        (*pBiasDecoder).SetIndex(cOutput, cOutput);
                         sum += pBiasDecoder->Get();
                     }
 
@@ -225,4 +225,4 @@ void Convolve(const TensorShape& rInputShape,
     }
 }
 
-} //namespace armnn
+} // namespace armnn