IVGCVSW-3223 Fix ref convolution performance regression

 * Do not use DataLayoutIndexed::GetIndex for weights and inputs,
   as it causes a large regression in performance
 * It turned out that the calculation of the indexes for the weights
   and inputs was the way it was because of an optimization done
   many months ago
 * Reverted the relevant hunks and added some comments so
   we won't make the same mistake again in the future
 * Made the GetIndex function inline to speed up other usages

Change-Id: I343b2ef0446993086f58b9dea1f0de0ba2d92216
Signed-off-by: Matteo Martincigh <matteo.martincigh@arm.com>

3 files changed, 70 insertions, 50 deletions

diff --git a/src/armnnUtils/DataLayoutIndexed.cpp b/src/armnnUtils/DataLayoutIndexed.cpp
index b02f07ec85..02f1e816ac 100644
--- a/src/armnnUtils/DataLayoutIndexed.cpp
+++ b/src/armnnUtils/DataLayoutIndexed.cpp
@@ -4,9 +4,6 @@
 //
 
 #include "DataLayoutIndexed.hpp"
-
-#include <boost/assert.hpp>
-
 using namespace armnn;
 
 namespace armnnUtils
@@ -33,40 +30,6 @@ DataLayoutIndexed::DataLayoutIndexed(armnn::DataLayout dataLayout)
     }
 }
 
-unsigned int DataLayoutIndexed::GetIndex(const TensorShape& shape,
-                                         unsigned int batchIndex, unsigned int channelIndex,
-                                         unsigned int heightIndex, unsigned int widthIndex) const
-{
-    BOOST_ASSERT( batchIndex < shape[0] || ( shape[0] == 0 && batchIndex == 0 ) );
-    BOOST_ASSERT( channelIndex < shape[m_ChannelsIndex] ||
-                ( shape[m_ChannelsIndex] == 0 && channelIndex == 0) );
-    BOOST_ASSERT( heightIndex < shape[m_HeightIndex] ||
-                ( shape[m_HeightIndex] == 0 && heightIndex == 0) );
-    BOOST_ASSERT( widthIndex < shape[m_WidthIndex] ||
-                ( shape[m_WidthIndex] == 0 && widthIndex == 0) );
-
-    // Offset the given indices appropriately depending on the data layout
-    switch (m_DataLayout)
-    {
-    case DataLayout::NHWC:
-        batchIndex  *= shape[1] * shape[2] * shape[3]; // batchIndex *= heightIndex * widthIndex * channelIndex
-        heightIndex *= shape[m_WidthIndex] * shape[m_ChannelsIndex];
-        widthIndex  *= shape[m_ChannelsIndex];
-        // channelIndex stays unchanged
-        break;
-    case DataLayout::NCHW:
-    default:
-        batchIndex   *= shape[1] * shape[2] * shape[3]; // batchIndex *= heightIndex * widthIndex * channelIndex
-        channelIndex *= shape[m_HeightIndex] * shape[m_WidthIndex];
-        heightIndex  *= shape[m_WidthIndex];
-        // widthIndex stays unchanged
-        break;
-    }
-
-    // Get the value using the correct offset
-    return batchIndex + channelIndex + heightIndex + widthIndex;
-}
-
 bool operator==(const DataLayout& dataLayout, const DataLayoutIndexed& indexed)
 {
     return dataLayout == indexed.GetDataLayout();
diff --git a/src/armnnUtils/DataLayoutIndexed.hpp b/src/armnnUtils/DataLayoutIndexed.hpp
index 5bb8e0d93f..8bd9701a5e 100644
--- a/src/armnnUtils/DataLayoutIndexed.hpp
+++ b/src/armnnUtils/DataLayoutIndexed.hpp
@@ -8,6 +8,8 @@
 #include <armnn/Types.hpp>
 #include <armnn/Tensor.hpp>
 
+#include <boost/assert.hpp>
+
 namespace armnnUtils
 {
 
@@ -21,9 +23,40 @@ public:
     unsigned int      GetChannelsIndex() const { return m_ChannelsIndex; }
     unsigned int      GetHeightIndex()   const { return m_HeightIndex; }
     unsigned int      GetWidthIndex()    const { return m_WidthIndex; }
-    unsigned int      GetIndex(const armnn::TensorShape& shape,
-                               unsigned int batchIndex, unsigned int channelIndex,
-                               unsigned int heightIndex, unsigned int widthIndex) const;
+
+    inline unsigned int GetIndex(const armnn::TensorShape& shape,
+                                 unsigned int batchIndex, unsigned int channelIndex,
+                                 unsigned int heightIndex, unsigned int widthIndex) const
+    {
+        BOOST_ASSERT( batchIndex < shape[0] || ( shape[0] == 0 && batchIndex == 0 ) );
+        BOOST_ASSERT( channelIndex < shape[m_ChannelsIndex] ||
+                    ( shape[m_ChannelsIndex] == 0 && channelIndex == 0) );
+        BOOST_ASSERT( heightIndex < shape[m_HeightIndex] ||
+                    ( shape[m_HeightIndex] == 0 && heightIndex == 0) );
+        BOOST_ASSERT( widthIndex < shape[m_WidthIndex] ||
+                    ( shape[m_WidthIndex] == 0 && widthIndex == 0) );
+
+        // Offset the given indices appropriately depending on the data layout
+        switch (m_DataLayout)
+        {
+        case armnn::DataLayout::NHWC:
+            batchIndex  *= shape[1] * shape[2] * shape[3]; // batchIndex *= heightIndex * widthIndex * channelIndex
+            heightIndex *= shape[m_WidthIndex] * shape[m_ChannelsIndex];
+            widthIndex  *= shape[m_ChannelsIndex];
+            // channelIndex stays unchanged
+            break;
+        case armnn::DataLayout::NCHW:
+        default:
+            batchIndex   *= shape[1] * shape[2] * shape[3]; // batchIndex *= heightIndex * widthIndex * channelIndex
+            channelIndex *= shape[m_HeightIndex] * shape[m_WidthIndex];
+            heightIndex  *= shape[m_WidthIndex];
+            // widthIndex stays unchanged
+            break;
+        }
+
+        // Get the value using the correct offset
+        return batchIndex + channelIndex + heightIndex + widthIndex;
+    }
 
 private:
     armnn::DataLayout m_DataLayout;
diff --git a/src/backends/reference/workloads/ConvImpl.cpp b/src/backends/reference/workloads/ConvImpl.cpp
index 801a29af1a..92e3b2d7dd 100644
--- a/src/backends/reference/workloads/ConvImpl.cpp
+++ b/src/backends/reference/workloads/ConvImpl.cpp
@@ -147,11 +147,22 @@ void Convolve(const TensorShape& rInputShape,
                                 }
                                 else
                                 {
-                                    filterIndex = dataLayoutIndexed.GetIndex(rFilterShape,
-                                                                             cOutput,
-                                                                             cInput,
-                                                                             yFilter,
-                                                                             xFilter);
+                                    // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great
+                                    // performance regression.
+                                    if (dataLayout == DataLayout::NHWC)
+                                    {
+                                        filterIndex = cOutput * filterHeight * filterWidth * inputChannels +
+                                                      yFilter * filterWidth * inputChannels +
+                                                      xFilter * inputChannels +
+                                                      cInput;
+                                    }
+                                    else
+                                    {
+                                        filterIndex = cOutput * filterWidth * filterHeight * inputChannels +
+                                                      cInput  * filterWidth * filterHeight +
+                                                      yFilter * filterWidth +
+                                                      xFilter;
+                                    }
                                 }
 
                                 rFilterDecoder[filterIndex];
@@ -170,11 +181,24 @@ void Convolve(const TensorShape& rInputShape,
                                 }
                                 else
                                 {
-                                    unsigned int inputIndex = dataLayoutIndexed.GetIndex(rInputShape,
-                                                                                         batchIdx,
-                                                                                         cInput,
-                                                                                         yInput - paddingTop,
-                                                                                         xInput - paddingLeft);
+                                    unsigned int inputIndex = 0;
+
+                                    // Keep this implementation, as using DataLayoutIndexed::GetIndex causes great
+                                    // performance regression.
+                                    if (dataLayout == DataLayout::NHWC)
+                                    {
+                                        inputIndex = batchIdx * inputHeight * inputWidth  * inputChannels +
+                                                     (yInput - paddingTop) * inputWidth * inputChannels +
+                                                     (xInput - paddingLeft) * inputChannels +
+                                                     cInput;
+                                    }
+                                    else
+                                    {
+                                        inputIndex = batchIdx * inputWidth * inputHeight * inputChannels +
+                                                     inputWidth * inputHeight * cInput +
+                                                     inputWidth * (yInput - paddingTop) +
+                                                     xInput - paddingLeft;
+                                    }
 
                                     rInputDecoder[inputIndex];
                                     inputValue = rInputDecoder.Get();