COMPMID-1995: Prepare Graph to support different input/output quantization info

- Added support for different input/output qinfo in ActivationLayer and DepthwiseConv
- Added support for different input/output qinfo in ConcatenateLayer introducing ConcatDescriptor
- Added reshape validate
- Allow OutputLayer to return a specific connection index from the input
- Not run Inplace and Depth mutator when input/output quantization info are different

Change-Id: I03f5e416fc43ddd284e1501887202a3145f76d8a
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-on: https://review.mlplatform.org/c/852
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

10 files changed, 64 insertions, 27 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 30f1fc6894..74f60d5354 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -111,9 +111,10 @@ NodeID GraphBuilder::add_output_node(Graph &g, NodeParams params, NodeIdxPair in
     return nid;
 }
 
-NodeID GraphBuilder::add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info)
+NodeID GraphBuilder::add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info,
+                                         const QuantizationInfo out_quant_info)
 {
-    return create_simple_single_input_output_node<ActivationLayerNode>(g, params, input, act_info);
+    return create_simple_single_input_output_node<ActivationLayerNode>(g, params, input, act_info, out_quant_info);
 }
 
 NodeID GraphBuilder::add_batch_normalization_node(Graph &g, NodeParams params, NodeIdxPair input, float epsilon,
@@ -293,11 +294,11 @@ NodeID GraphBuilder::add_deconvolution_node(Graph &g, NodeParams params, NodeIdx
     return deconv_nid;
 }
 
-NodeID GraphBuilder::add_concatenate_node(Graph &g, NodeParams params, std::vector<NodeIdxPair> inputs, DataLayoutDimension axis)
+NodeID GraphBuilder::add_concatenate_node(Graph &g, NodeParams params, std::vector<NodeIdxPair> inputs, descriptors::ConcatLayerDescriptor concat_descriptor)
 {
     ARM_COMPUTE_ERROR_ON(inputs.size() == 0);
 
-    NodeID nid = g.add_node<ConcatenateLayerNode>(inputs.size(), axis);
+    NodeID nid = g.add_node<ConcatenateLayerNode>(inputs.size(), concat_descriptor);
 
     unsigned int i = 0;
     for(const auto &input : inputs)
@@ -312,7 +313,7 @@ NodeID GraphBuilder::add_concatenate_node(Graph &g, NodeParams params, std::vect
 
 NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params, NodeIdxPair input, Size2D kernel_spatial_extend,
                                                     PadStrideInfo conv_info, int depth_multiplier, DepthwiseConvolutionMethod method,
-                                                    ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const QuantizationInfo quant_info)
+                                                    ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const QuantizationInfo quant_info, const QuantizationInfo out_quant_info)
 {
     CHECK_NODEIDX_PAIR(input, g);
     ARM_COMPUTE_ERROR_ON((kernel_spatial_extend.width == 0) || (kernel_spatial_extend.height == 0));
@@ -351,7 +352,7 @@ NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params,
     }
 
     // Create convolution node and connect
-    NodeID conv_nid = g.add_node<DepthwiseConvolutionLayerNode>(conv_info, depth_multiplier, method);
+    NodeID conv_nid = g.add_node<DepthwiseConvolutionLayerNode>(conv_info, depth_multiplier, method, out_quant_info);
     g.add_connection(input.node_id, input.index, conv_nid, 0);
     g.add_connection(w_nid, 0, conv_nid, 1);
     if(has_bias)
diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp
index 4b71837a49..cb8dc0a172 100644
--- a/src/graph/backends/CL/CLNodeValidator.cpp
+++ b/src/graph/backends/CL/CLNodeValidator.cpp
@@ -74,6 +74,8 @@ Status CLNodeValidator::validate(INode *node)
             return detail::validate_priorbox_layer<CLPriorBoxLayer>(*polymorphic_downcast<PriorBoxLayerNode *>(node));
         case NodeType::ReorgLayer:
             return detail::validate_reorg_layer<CLReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
+        case NodeType::ReshapeLayer:
+            return detail::validate_reshape_layer<CLReshapeLayer>(*polymorphic_downcast<ReshapeLayerNode *>(node));
         case NodeType::ROIAlignLayer:
             return detail::validate_roi_align_layer<CLROIAlignLayer>(*polymorphic_downcast<ROIAlignLayerNode *>(node));
         case NodeType::SliceLayer:
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index f23845c314..81c6e09f92 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -143,8 +143,10 @@ std::unique_ptr<IFunction> create_convolution_layer<NEConvolutionLayerFunctions,
             << " Weights QuantInfo: " << weights->info()->quantization_info()
             << " Output QuantInfo: " << output->info()->quantization_info();
     }
-    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
-                               << " Target " << NETargetInfo::TargetType
+    ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
+                               << node.name()
+                               << " Type: " << func_name
+                               << " Target: " << NETargetInfo::TargetType
                                << " Data Type: " << input->info()->data_type()
                                << qss.str()
                                << " Input shape: " << input->info()->tensor_shape()
diff --git a/src/graph/backends/NEON/NENodeValidator.cpp b/src/graph/backends/NEON/NENodeValidator.cpp
index b0feec563b..77f2e7f21d 100644
--- a/src/graph/backends/NEON/NENodeValidator.cpp
+++ b/src/graph/backends/NEON/NENodeValidator.cpp
@@ -74,6 +74,8 @@ Status NENodeValidator::validate(INode *node)
             return detail::validate_priorbox_layer<NEPriorBoxLayer>(*polymorphic_downcast<PriorBoxLayerNode *>(node));
         case NodeType::ReorgLayer:
             return detail::validate_reorg_layer<NEReorgLayer>(*polymorphic_downcast<ReorgLayerNode *>(node));
+        case NodeType::ReshapeLayer:
+            return detail::validate_reshape_layer<NEReshapeLayer>(*polymorphic_downcast<ReshapeLayerNode *>(node));
         case NodeType::ROIAlignLayer:
             return ARM_COMPUTE_CREATE_ERROR(arm_compute::ErrorCode::RUNTIME_ERROR, "Unsupported operation : ROIAlignLayer");
         case NodeType::SliceLayer:
diff --git a/src/graph/mutators/DepthConcatSubTensorMutator.cpp b/src/graph/mutators/DepthConcatSubTensorMutator.cpp
index a170c4d899..0e0a26b886 100644
--- a/src/graph/mutators/DepthConcatSubTensorMutator.cpp
+++ b/src/graph/mutators/DepthConcatSubTensorMutator.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -69,11 +69,12 @@ void DepthConcatSubTensorMutator::mutate(Graph &g)
                 continue;
             }
 
-            // Check that all tensor have the same target and valid inputs
+            // Check that all tensor have the same target, valid inputs and same quantization info
             bool is_valid = std::all_of(node->input_edges().cbegin(), node->input_edges().cend(),
                                         [&](const EdgeID & eid)
             {
-                return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target);
+                return (g.edge(eid) != nullptr) && (g.edge(eid)->tensor() != nullptr) && (g.edge(eid)->tensor()->desc().target == output_tensor->desc().target)
+                       && (g.edge(eid)->tensor()->desc().quant_info == output_tensor->desc().quant_info);
             });
 
             // Create subtensors
diff --git a/src/graph/mutators/InPlaceOperationMutator.cpp b/src/graph/mutators/InPlaceOperationMutator.cpp
index 31921b328e..1c2985dce6 100644
--- a/src/graph/mutators/InPlaceOperationMutator.cpp
+++ b/src/graph/mutators/InPlaceOperationMutator.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,8 +56,8 @@ void InPlaceOperationMutator::mutate(Graph &g)
 
                 ARM_COMPUTE_ERROR_ON(current_output_tensor == nullptr || new_output_tensor == nullptr);
 
-                // Prevent in-place operation if there is an accessor bound to the in-place tensor
-                if(new_output_tensor->accessor() == nullptr)
+                // Prevent in-place operation if there is an accessor bound to the in-place tensor or quantization info are different
+                if(new_output_tensor->accessor() == nullptr || current_output_tensor->desc().quant_info == new_output_tensor->desc().quant_info)
                 {
                     ARM_COMPUTE_LOG_GRAPH_VERBOSE("Switching to in-place computation for the node with ID : "
                                                   << node->id() << " and name : " << node->name() << std::endl);
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp
index 5927a597bb..724307e7b7 100644
--- a/src/graph/mutators/NodeFusionMutator.cpp
+++ b/src/graph/mutators/NodeFusionMutator.cpp
@@ -211,10 +211,17 @@ void NodeFusionMutator::mutate(Graph &g)
     {
         return true;
     };
-    auto qs8_prec = [](INode & n)
+    auto qs8_prec = [&g](INode & n)
     {
         ARM_COMPUTE_ERROR_ON(n.output(0) == nullptr);
-        return n.output(0)->desc().data_type == DataType::QASYMM8;
+
+        const auto output_edge_id = *n.output_edges().begin();
+        const auto output_edge    = g.edge(output_edge_id);
+        // To perform fusion the two nodes must have same output quantization information
+        const bool same_qinfo     = n.output(0)->desc().quant_info == output_edge->producer()->output(0)->desc().quant_info;
+        const bool output_qasymm8 = n.output(0)->desc().data_type == DataType::QASYMM8;
+
+        return output_qasymm8 && same_qinfo;
     };
 
     // Fusion mutations
diff --git a/src/graph/nodes/ActivationLayerNode.cpp b/src/graph/nodes/ActivationLayerNode.cpp
index 85cb10bbdb..ada6cf981f 100644
--- a/src/graph/nodes/ActivationLayerNode.cpp
+++ b/src/graph/nodes/ActivationLayerNode.cpp
@@ -30,8 +30,8 @@ namespace arm_compute
 {
 namespace graph
 {
-ActivationLayerNode::ActivationLayerNode(ActivationLayerInfo info)
-    : _info(info)
+ActivationLayerNode::ActivationLayerNode(ActivationLayerInfo info, QuantizationInfo out_quant_info)
+    : _info(info), _out_quant_info(out_quant_info)
 {
     _input_edges.resize(1, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -62,7 +62,13 @@ TensorDescriptor ActivationLayerNode::configure_output(size_t idx) const
     const Tensor *src = input(0);
     ARM_COMPUTE_ERROR_ON(src == nullptr);
 
-    return src->desc();
+    TensorDescriptor output_info = src->desc();
+    if(!_out_quant_info.empty())
+    {
+        output_info.quant_info = _out_quant_info;
+    }
+
+    return output_info;
 }
 
 NodeType ActivationLayerNode::type() const
diff --git a/src/graph/nodes/ConcatenateLayerNode.cpp b/src/graph/nodes/ConcatenateLayerNode.cpp
index 3ce09d0073..bbdc4dc029 100644
--- a/src/graph/nodes/ConcatenateLayerNode.cpp
+++ b/src/graph/nodes/ConcatenateLayerNode.cpp
@@ -34,8 +34,8 @@ namespace arm_compute
 {
 namespace graph
 {
-ConcatenateLayerNode::ConcatenateLayerNode(unsigned int total_nodes, DataLayoutDimension axis)
-    : _total_nodes(total_nodes), _axis(axis), _is_enabled(true)
+ConcatenateLayerNode::ConcatenateLayerNode(unsigned int total_nodes, descriptors::ConcatLayerDescriptor concat_descriptor)
+    : _total_nodes(total_nodes), _concat_descriptor(std::move(concat_descriptor)), _is_enabled(true)
 {
     _input_edges.resize(_total_nodes, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -53,7 +53,12 @@ bool ConcatenateLayerNode::is_enabled() const
 
 DataLayoutDimension ConcatenateLayerNode::concatenation_axis() const
 {
-    return _axis;
+    return _concat_descriptor.axis;
+}
+
+QuantizationInfo ConcatenateLayerNode::output_quantization_info() const
+{
+    return _concat_descriptor.output_qinfo;
 }
 
 TensorDescriptor ConcatenateLayerNode::compute_output_descriptor(const std::vector<TensorDescriptor> &input_descriptors,
@@ -121,7 +126,11 @@ TensorDescriptor ConcatenateLayerNode::configure_output(size_t idx) const
             ARM_COMPUTE_ERROR_ON(t == nullptr);
             inputs_descriptors.push_back(t->desc());
         }
-        output_info = compute_output_descriptor(inputs_descriptors, _axis);
+        output_info = compute_output_descriptor(inputs_descriptors, _concat_descriptor.axis);
+        if(!_concat_descriptor.output_qinfo.empty())
+        {
+            output_info.quant_info = _concat_descriptor.output_qinfo;
+        }
     }
 
     return output_info;
diff --git a/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp b/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
index 75ca5f4e03..935902d3fd 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayerNode.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,8 +32,9 @@ namespace arm_compute
 {
 namespace graph
 {
-DepthwiseConvolutionLayerNode::DepthwiseConvolutionLayerNode(PadStrideInfo info, int depth_multiplier, DepthwiseConvolutionMethod method)
-    : _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _fused_activation()
+DepthwiseConvolutionLayerNode::DepthwiseConvolutionLayerNode(PadStrideInfo info, int depth_multiplier, DepthwiseConvolutionMethod method,
+                                                             QuantizationInfo out_quant_info)
+    : _info(std::move(info)), _depth_multiplier(depth_multiplier), _method(method), _out_quant_info(out_quant_info), _fused_activation()
 {
     _input_edges.resize(3, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -113,7 +114,13 @@ TensorDescriptor DepthwiseConvolutionLayerNode::configure_output(size_t idx) con
 
     ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr);
 
-    return compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+    TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info, _depth_multiplier);
+    if(!_out_quant_info.empty())
+    {
+        output_info.quant_info = _out_quant_info;
+    }
+
+    return output_info;
 }
 
 NodeType DepthwiseConvolutionLayerNode::type() const