IVGCVSW-3228 Fix bias quantization to be INT32 not QAsymm8

 * Add function to calculate bias tensor quantization scale
   from input and weights scales.
 * Change visitor method of Conv2d, DepthwiseConv and FullyConnected to use
   the new function.
 * Fix Unit tests to expect correctly calculated quantization parameters.

Change-Id: Ic36f47ceea81243c813d74ccf791e984c819cc71
Signed-off-by: Francis Murtagh <francis.murtagh@arm.com>

1 files changed, 6 insertions, 0 deletions

diff --git a/src/armnn/QuantizerVisitor.hpp b/src/armnn/QuantizerVisitor.hpp
index 6e5609df02..6463350f2b 100644
--- a/src/armnn/QuantizerVisitor.hpp
+++ b/src/armnn/QuantizerVisitor.hpp
@@ -139,6 +139,12 @@ private:
     /// Record the guids so we can easily find the layers later
     void RecordLayer(const IConnectableLayer* srcLayer, IConnectableLayer* qLayer);
 
+    /// Sets the bias quantization scale based on input and weight scales
+    ConstTensor CreateQuantizedBias(const IConnectableLayer* srcLayer,
+                                    const ConstTensor& weights,
+                                    const Optional<ConstTensor>& biases,
+                                    std::vector<int32_t>& weightsBacking);
+
     /// Reference to the static range visitor used to retrieve the quantization ranges
     const RangeTracker& m_Ranges;