arm_compute v18.01

Change-Id: I9bfa178c2e38bfd5fc812e62aab6760d87748e05
diff --git a/src/graph/SubGraph.cpp b/src/graph/SubGraph.cpp
index e975421..8ba2af6 100644
--- a/src/graph/SubGraph.cpp
+++ b/src/graph/SubGraph.cpp
@@ -52,12 +52,12 @@
     }
 }
 
-std::unique_ptr<Graph> SubGraph::construct(TargetHint hint, std::unique_ptr<ITensorObject> input, std::unique_ptr<ITensorObject> output)
+std::unique_ptr<Graph> SubGraph::construct(const GraphContext &ctx, std::unique_ptr<ITensorObject> input, std::unique_ptr<ITensorObject> output)
 {
     auto graph = arm_compute::support::cpp14::make_unique<Graph>();
 
     // Set hint
-    graph->hints().set_target_hint(hint);
+    graph->hints() = ctx.hints();
 
     // Configure input
     if(_input == nullptr)
diff --git a/src/graph/SubTensor.cpp b/src/graph/SubTensor.cpp
index 2edeb3b..2e640dd 100644
--- a/src/graph/SubTensor.cpp
+++ b/src/graph/SubTensor.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,21 +37,21 @@
 namespace
 {
 template <typename SubTensorType, typename ParentTensorType>
-std::unique_ptr<arm_compute::ITensor> initialise_subtensor(arm_compute::ITensor *parent, TensorShape shape, Coordinates coords)
+std::unique_ptr<arm_compute::ITensor> initialise_subtensor(arm_compute::ITensor *parent, TensorShape shape, Coordinates coords, bool extend_parent)
 {
     auto ptensor   = dynamic_cast<ParentTensorType *>(parent);
-    auto subtensor = arm_compute::support::cpp14::make_unique<SubTensorType>(ptensor, shape, coords);
+    auto subtensor = arm_compute::support::cpp14::make_unique<SubTensorType>(ptensor, shape, coords, extend_parent);
     return std::move(subtensor);
 }
 } // namespace
 
 SubTensor::SubTensor()
-    : _target(TargetHint::DONT_CARE), _tensor_shape(), _coords(), _parent(nullptr), _subtensor(nullptr)
+    : _target(TargetHint::DONT_CARE), _tensor_shape(), _coords(), _parent(nullptr), _subtensor(nullptr), _extend_parent(false)
 {
 }
 
-SubTensor::SubTensor(Tensor &parent, TensorShape tensor_shape, Coordinates coords)
-    : _target(TargetHint::DONT_CARE), _tensor_shape(tensor_shape), _coords(coords), _parent(nullptr), _subtensor(nullptr)
+SubTensor::SubTensor(Tensor &parent, TensorShape tensor_shape, Coordinates coords, bool extend_parent)
+    : _target(TargetHint::DONT_CARE), _tensor_shape(tensor_shape), _coords(coords), _parent(nullptr), _subtensor(nullptr), _extend_parent(extend_parent)
 {
     ARM_COMPUTE_ERROR_ON(parent.tensor() == nullptr);
     _parent = parent.tensor();
@@ -60,8 +60,8 @@
     instantiate_subtensor();
 }
 
-SubTensor::SubTensor(arm_compute::ITensor *parent, TensorShape tensor_shape, Coordinates coords, TargetHint target)
-    : _target(target), _tensor_shape(tensor_shape), _coords(coords), _parent(parent), _subtensor(nullptr)
+SubTensor::SubTensor(arm_compute::ITensor *parent, TensorShape tensor_shape, Coordinates coords, TargetHint target, bool extend_parent)
+    : _target(target), _tensor_shape(tensor_shape), _coords(coords), _parent(parent), _subtensor(nullptr), _extend_parent(extend_parent)
 {
     ARM_COMPUTE_ERROR_ON(parent == nullptr);
     instantiate_subtensor();
@@ -108,10 +108,10 @@
     switch(_target)
     {
         case TargetHint::OPENCL:
-            _subtensor = initialise_subtensor<arm_compute::CLSubTensor, arm_compute::ICLTensor>(_parent, _tensor_shape, _coords);
+            _subtensor = initialise_subtensor<arm_compute::CLSubTensor, arm_compute::ICLTensor>(_parent, _tensor_shape, _coords, _extend_parent);
             break;
         case TargetHint::NEON:
-            _subtensor = initialise_subtensor<arm_compute::SubTensor, arm_compute::ITensor>(_parent, _tensor_shape, _coords);
+            _subtensor = initialise_subtensor<arm_compute::SubTensor, arm_compute::ITensor>(_parent, _tensor_shape, _coords, _extend_parent);
             break;
         default:
             ARM_COMPUTE_ERROR("Invalid TargetHint");
diff --git a/src/graph/nodes/BranchLayer.cpp b/src/graph/nodes/BranchLayer.cpp
index d062e4b..7a20a56 100644
--- a/src/graph/nodes/BranchLayer.cpp
+++ b/src/graph/nodes/BranchLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -37,46 +37,6 @@
 
 using namespace arm_compute::graph;
 
-namespace
-{
-void depth_concatenate_output_info(ITensorInfo *info, ITensorInfo *sub_tensor_info)
-{
-    ARM_COMPUTE_ERROR_ON(info == nullptr);
-    ARM_COMPUTE_ERROR_ON(sub_tensor_info == nullptr);
-
-    TensorShape        info_shape            = info->tensor_shape();
-    const TensorShape &sub_tensor_info_shape = sub_tensor_info->tensor_shape();
-
-    // Update parent info and valid region
-    if(info_shape.total_size() == 0)
-    {
-        arm_compute::auto_init_if_empty(*info,
-                                        sub_tensor_info->tensor_shape(),
-                                        sub_tensor_info->num_channels(),
-                                        sub_tensor_info->data_type(), sub_tensor_info->fixed_point_position());
-        info->set_valid_region(sub_tensor_info->valid_region());
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR_ON(info->num_channels() != sub_tensor_info->num_channels());
-        ARM_COMPUTE_ERROR_ON(info->data_type() != sub_tensor_info->data_type());
-        ARM_COMPUTE_ERROR_ON(info->fixed_point_position() != sub_tensor_info->fixed_point_position());
-
-        // Concatenate depth
-        ARM_COMPUTE_ERROR_ON(info_shape.x() != sub_tensor_info_shape.x());
-        ARM_COMPUTE_ERROR_ON(info_shape.y() != sub_tensor_info_shape.y());
-        info_shape.set(2, info_shape.z() + sub_tensor_info_shape.z());
-        info->set_tensor_shape(info_shape);
-
-        // Update valid region
-        arm_compute::ValidRegion info_valid_region = info->valid_region();
-        info_valid_region.shape.set(2, info_shape.z());
-        arm_compute::ValidRegion updated_region = arm_compute::intersect_valid_regions(info_valid_region, sub_tensor_info->valid_region());
-        info->set_valid_region(updated_region);
-    }
-}
-} // namespace
-
 /** Branch function */
 class BranchFunction final : public arm_compute::IFunction
 {
@@ -117,9 +77,8 @@
     // Create branch function
     auto func = arm_compute::support::cpp14::make_unique<BranchFunction>();
 
-    // Track output SubTensorInfo and depth
-    TensorInfo out_info;
-    int        depth = 0;
+    // Track output depth
+    int depth = 0;
 
     // Constuct all sub-graphs given the input/output
     for(auto &sg : _sub_graphs)
@@ -143,15 +102,18 @@
         // Create output sub-tensor
         if(!sg->has_output())
         {
-            ARM_COMPUTE_ERROR_ON(dynamic_cast<Tensor *>(output) == nullptr);
-            out = arm_compute::support::cpp14::make_unique<SubTensor>(*dynamic_cast<Tensor *>(output),
-                                                                      output->tensor()->info()->tensor_shape(),
-                                                                      Coordinates(0, 0, depth));
+            ARM_COMPUTE_ERROR_ON((dynamic_cast<Tensor *>(output) == nullptr) && (dynamic_cast<SubTensor *>(output) == nullptr));
+
+            out = arm_compute::support::cpp14::make_unique<SubTensor>(output->tensor(),
+                                                                      TensorShape(),
+                                                                      Coordinates(0, 0, depth),
+                                                                      output->target(),
+                                                                      true);
             out_sub_tensor = dynamic_cast<SubTensor *>(out.get());
         }
 
         // Construct sub_graph
-        auto g = sg->construct(ctx.hints().target_hint(), std::move(in), std::move(out));
+        auto g = sg->construct(ctx, std::move(in), std::move(out));
 
         // Register graph to function
         func->register_graph(std::move(g));
@@ -161,16 +123,8 @@
         {
             ARM_COMPUTE_ERROR_ON(out_sub_tensor->tensor() == nullptr);
             depth += out_sub_tensor->tensor()->info()->tensor_shape()[2];
-            depth_concatenate_output_info(&out_info, out_sub_tensor->tensor()->info());
         }
     }
 
-    // Auto-init output
-    arm_compute::auto_init_if_empty(*output->tensor()->info(),
-                                    out_info.tensor_shape(),
-                                    out_info.num_channels(),
-                                    out_info.data_type(),
-                                    out_info.fixed_point_position());
-
     return std::move(func);
 }
\ No newline at end of file
diff --git a/src/graph/nodes/ConvolutionLayer.cpp b/src/graph/nodes/ConvolutionLayer.cpp
index ae4a8d7..f292b89 100644
--- a/src/graph/nodes/ConvolutionLayer.cpp
+++ b/src/graph/nodes/ConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -106,13 +106,16 @@
                                                                         const WeightsInfo    &weights_info,
                                                                         ConvolutionMethodHint conv_method)
 {
-    if(conv_method == ConvolutionMethodHint::GEMM)
+    if((conv_method == ConvolutionMethodHint::DIRECT)
+       && arm_compute::CLDirectConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info)) // NOLINT
     {
-        return instantiate_function<arm_compute::CLConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info, weights_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLDirectConvolutionLayer");
+        return instantiate_direct_function<arm_compute::CLDirectConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info);
     }
     else
     {
-        return instantiate_direct_function<arm_compute::CLDirectConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
+        return instantiate_function<arm_compute::CLConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info, weights_info);
     }
 }
 
@@ -122,13 +125,16 @@
                                                                       const WeightsInfo    &weights_info,
                                                                       ConvolutionMethodHint conv_method)
 {
-    if(conv_method == ConvolutionMethodHint::GEMM)
+    if((conv_method == ConvolutionMethodHint::DIRECT)
+       && arm_compute::NEDirectConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info)) // NOLINT
     {
-        return instantiate_function<arm_compute::NEConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info, weights_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDirectConvolutionLayer");
+        return instantiate_direct_function<arm_compute::NEDirectConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info);
     }
     else
     {
-        return instantiate_direct_function<arm_compute::NEDirectConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
+        return instantiate_function<arm_compute::NEConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info, weights_info);
     }
 }
 } // namespace
@@ -184,14 +190,17 @@
     // Set weights and biases info
     if(_weights.tensor() == nullptr)
     {
-        _weights.set_info(TensorInfo(TensorShape(_conv_width, _conv_height, in->info()->dimension(2) / _num_groups, _ofm),
+        TensorInfo info = TensorInfo(TensorShape(_conv_width, _conv_height, in->info()->dimension(2) / _num_groups, _ofm),
                                      in->info()->num_channels(),
                                      in->info()->data_type(),
-                                     in->info()->fixed_point_position()));
+                                     in->info()->fixed_point_position());
+        info.set_quantization_info(_weights_quant_info);
+        _weights.set_info(std::move(info));
     }
     if(_biases.has_accessor() && _biases.tensor() == nullptr)
     {
-        _biases.set_info(TensorInfo(TensorShape(_ofm), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        DataType dt = in->info()->data_type();
+        _biases.set_info(TensorInfo(TensorShape(_ofm), in->info()->num_channels(), is_data_type_quantized_asymmetric(dt) ? DataType::S32 : dt, in->info()->fixed_point_position()));
     }
 
     std::unique_ptr<arm_compute::IFunction> func;
@@ -213,7 +222,8 @@
     TensorShape output_shape = calculate_convolution_layer_output_shape(in->info()->tensor_shape(), _weights.info().tensor_shape(), _conv_info);
 
     // Output auto inizialitation if not yet initialized
-    arm_compute::auto_init_if_empty(*out->info(), output_shape, 1, in->info()->data_type(), in->info()->fixed_point_position());
+    arm_compute::auto_init_if_empty(*out->info(), output_shape, 1, in->info()->data_type(), in->info()->fixed_point_position(),
+                                    (_out_quant_info.empty()) ? in->info()->quantization_info() : _out_quant_info);
 
     // Create appropriate convolution function
     if(_num_groups == 1)
@@ -254,12 +264,10 @@
     std::unique_ptr<arm_compute::IFunction> func;
     if(_target_hint == TargetHint::OPENCL)
     {
-        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
         func = instantiate<TargetHint::OPENCL>(input, _weights.tensor(), _biases.tensor(), output, _conv_info, _weights_info, conv_method_hint);
     }
     else
     {
-        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
         func = instantiate<TargetHint::NEON>(input, _weights.tensor(), _biases.tensor(), output, _conv_info, _weights_info, conv_method_hint);
     }
     return func;
@@ -321,12 +329,10 @@
         // Instantiate convolution function
         if(_target_hint == TargetHint::OPENCL)
         {
-            ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
             func = instantiate<TargetHint::OPENCL>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint);
         }
         else
         {
-            ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
             func = instantiate<TargetHint::NEON>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint);
         }
 
diff --git a/src/graph/nodes/DepthwiseConvolutionLayer.cpp b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
index b459853..1209d03 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayer.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,11 +43,14 @@
         TensorShape shape = in->info()->tensor_shape();
         shape.set(Window::DimX, _conv_width);
         shape.set(Window::DimY, _conv_height);
-        _weights.set_info(TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        TensorInfo info = TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position());
+        info.set_quantization_info(_quant_info);
+        _weights.set_info(std::move(info));
     }
     if(_biases.has_accessor() && _biases.tensor() == nullptr)
     {
-        _biases.set_info(TensorInfo(TensorShape(in->info()->dimension(2)), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        DataType dt = in->info()->data_type();
+        _biases.set_info(TensorInfo(TensorShape(in->info()->dimension(2)), in->info()->num_channels(), is_data_type_quantized_asymmetric(dt) ? DataType::S32 : dt, in->info()->fixed_point_position()));
     }
 
     bool weights_is_loaded = _weights.tensor() != nullptr;
diff --git a/src/graph/nodes/ReshapeLayer.cpp b/src/graph/nodes/ReshapeLayer.cpp
index bbe0739..b0c117e 100644
--- a/src/graph/nodes/ReshapeLayer.cpp
+++ b/src/graph/nodes/ReshapeLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,7 +44,7 @@
     arm_compute::ITensor *out = output->tensor();
 
     // Auto configure output
-    arm_compute::auto_init_if_empty(*out->info(), _shape, 1, in->info()->data_type(), in->info()->fixed_point_position());
+    arm_compute::auto_init_if_empty(*out->info(), _shape, 1, in->info()->data_type(), in->info()->fixed_point_position(), in->info()->quantization_info());
 
     // Create node context
     NodeContext node_ctx(OperationType::ReshapeLayer);
diff --git a/src/graph/operations/CLSimpleOperations.cpp b/src/graph/operations/CLSimpleOperations.cpp
index 8f2bf23..61315e7 100644
--- a/src/graph/operations/CLSimpleOperations.cpp
+++ b/src/graph/operations/CLSimpleOperations.cpp
@@ -156,13 +156,13 @@
     bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
     if(run_3x3_opt)
     {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
         depwthwise_conv->configure(in, weights, biases, out, conv_info);
         func = std::move(depwthwise_conv);
     }
     else
     {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
         depwthwise_conv->configure(in, weights, biases, out, conv_info);
         func = std::move(depwthwise_conv);
     }
diff --git a/src/graph/operations/NESimpleOperations.cpp b/src/graph/operations/NESimpleOperations.cpp
index bb99e8d..49adbe9 100644
--- a/src/graph/operations/NESimpleOperations.cpp
+++ b/src/graph/operations/NESimpleOperations.cpp
@@ -149,23 +149,12 @@
     auto      *biases    = ctx.num_inputs() == 3 ? dynamic_cast<arm_compute::ITensor *>(ctx.input(2)) : nullptr;
     auto      *out       = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
     const auto conv_info = ctx.parameter<PadStrideInfo>("ConvolutionInfo");
-    const auto opt3x3    = ctx.parameter<bool>("Optimized3x3");
 
     // Create and configure function
     std::unique_ptr<arm_compute::IFunction> func;
-    bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
-    if(run_3x3_opt)
-    {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
-        depwthwise_conv->configure(in, weights, biases, out, conv_info);
-        func = std::move(depwthwise_conv);
-    }
-    else
-    {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>();
-        depwthwise_conv->configure(in, weights, biases, out, conv_info);
-        func = std::move(depwthwise_conv);
-    }
+    auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
+    depwthwise_conv->configure(in, weights, biases, out, conv_info);
+    func = std::move(depwthwise_conv);
 
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDepthwiseConvolutionLayer"
@@ -460,4 +449,4 @@
                                << std::endl);
 
     return std::move(smx);
-}
\ No newline at end of file
+}