COMPMID-1188: Add quantization info support in graph FC layer.

Change-Id: Ie9a6a896da142198243139fb9f8be0f83b87ccce
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142130
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Vidhya Sudhan Loganathan <vidhyasudhan.loganathan@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

2 files changed, 24 insertions, 8 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index b3721719d9..7f567fd559 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -423,7 +423,8 @@ NodeID GraphBuilder::add_flatten_node(Graph &g, NodeParams params, NodeIdxPair i
 }
 
 NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs,
-                                               ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor)
+                                               ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor,
+                                               const QuantizationInfo weights_quant_info, const QuantizationInfo out_quant_info)
 {
     CHECK_NODEIDX_PAIR(input, g);
     ARM_COMPUTE_ERROR_ON(num_outputs == 0);
@@ -434,7 +435,7 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node
     const TensorDescriptor input_tensor_desc = get_tensor_descriptor(g, g.node(input.node_id)->outputs()[0]);
 
     // Create weights node
-    TensorDescriptor w_desc = FullyConnectedLayerNode::compute_weights_descriptor(input_tensor_desc, num_outputs);
+    TensorDescriptor w_desc = FullyConnectedLayerNode::compute_weights_descriptor(input_tensor_desc, num_outputs, weights_quant_info);
     NodeID           w_nid  = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
 
     // Create bias nodes
@@ -456,7 +457,7 @@ NodeID GraphBuilder::add_fully_connected_layer(Graph &g, NodeParams params, Node
     fc_info.weights_trained_layout = DataLayout::NCHW;
 
     // Create fully connected node and connect
-    NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, fc_info);
+    NodeID fc_nid = g.add_node<FullyConnectedLayerNode>(num_outputs, out_quant_info, fc_info);
     g.add_connection(input.node_id, input.index, fc_nid, 0);
     g.add_connection(w_nid, 0, fc_nid, 1);
     if(has_bias)
diff --git a/src/graph/nodes/FullyConnectedLayer.cpp b/src/graph/nodes/FullyConnectedLayer.cpp
index 6ba4eefa88..f41e1a3c22 100644
--- a/src/graph/nodes/FullyConnectedLayer.cpp
+++ b/src/graph/nodes/FullyConnectedLayer.cpp
@@ -31,15 +31,16 @@ namespace arm_compute
 {
 namespace graph
 {
-FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, FullyConnectedLayerInfo fc_info)
-    : _num_outputs(num_outputs), _info(fc_info)
+FullyConnectedLayerNode::FullyConnectedLayerNode(unsigned int num_outputs, QuantizationInfo out_quant_info, FullyConnectedLayerInfo fc_info)
+    : _num_outputs(num_outputs), _out_quant_info(out_quant_info), _info(fc_info)
 {
     _input_edges.resize(3, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
 }
 
 TensorDescriptor FullyConnectedLayerNode::compute_weights_descriptor(const TensorDescriptor &input_descriptor,
-                                                                     unsigned int            num_outputs)
+                                                                     unsigned int            num_outputs,
+                                                                     QuantizationInfo        weights_quant_info)
 {
     unsigned int num_weights    = 1;
     unsigned int num_dimensions = input_descriptor.shape.num_dimensions();
@@ -56,11 +57,18 @@ TensorDescriptor FullyConnectedLayerNode::compute_weights_descriptor(const Tenso
     TensorDescriptor weights_descriptor = input_descriptor;
     weights_descriptor.shape            = TensorShape(num_weights, num_outputs);
 
+    // Set quantization info if present
+    if(!weights_quant_info.empty())
+    {
+        weights_descriptor.quant_info = weights_quant_info;
+    }
+
     return weights_descriptor;
 }
 
 TensorDescriptor FullyConnectedLayerNode::compute_output_descriptor(const TensorDescriptor &input_descriptor,
-                                                                    unsigned int            num_outputs)
+                                                                    unsigned int            num_outputs,
+                                                                    QuantizationInfo        out_quant_info)
 {
     // Note: Only 1D batch space is supported at the moment
     unsigned int batches = input_descriptor.shape[1];
@@ -69,9 +77,16 @@ TensorDescriptor FullyConnectedLayerNode::compute_output_descriptor(const Tensor
         batches = input_descriptor.shape[3];
     }
 
+    // Set descriptor shape
     TensorDescriptor output_descriptor = input_descriptor;
     output_descriptor.shape            = TensorShape(num_outputs, batches);
 
+    // Set quantization info if present
+    if(!out_quant_info.empty())
+    {
+        output_descriptor.quant_info = out_quant_info;
+    }
+
     return output_descriptor;
 }
 
@@ -98,7 +113,7 @@ TensorDescriptor FullyConnectedLayerNode::configure_output(size_t idx) const
     const Tensor *src = input(0);
     ARM_COMPUTE_ERROR_ON(src == nullptr);
 
-    return compute_output_descriptor(src->desc(), _num_outputs);
+    return compute_output_descriptor(src->desc(), _num_outputs, _out_quant_info);
 }
 
 NodeType FullyConnectedLayerNode::type() const