COMPMID-797 Integrate Mobilenet QASYMM8 with new graph.

Change-Id: I4df63ec2f4eb27a8a6eec2bea27741bf8dec6910
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/126966
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

9 files changed, 106 insertions, 35 deletions

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 0d1bdc3596..4ad34e789c 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -206,7 +206,9 @@ NodeID GraphBuilder::add_batch_normalization_node(Graph &g, NodeParams params, N
 NodeID GraphBuilder::add_convolution_node(Graph &g, NodeParams params, NodeIdxPair input,
                                           Size2D kernel_spatial_extend, unsigned int depth, PadStrideInfo conv_info,
                                           unsigned int num_groups, ConvolutionMethod method,
-                                          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(depth == 0);
@@ -220,7 +222,13 @@ NodeID GraphBuilder::add_convolution_node(Graph &g, NodeParams params, NodeIdxPa
     // Create weights node
     TensorDescriptor w_desc = input_tensor_desc;
     w_desc.shape            = TensorShape(kernel_spatial_extend.width, kernel_spatial_extend.height, w_desc.shape.z() / num_groups, depth);
-    NodeID w_nid            = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
+
+    if(!weights_quant_info.empty())
+    {
+        w_desc.quant_info = weights_quant_info;
+    }
+
+    NodeID w_nid = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
 
     // Create bias nodes
     NodeID b_nid = EmptyNodeID;
@@ -234,7 +242,7 @@ NodeID GraphBuilder::add_convolution_node(Graph &g, NodeParams params, NodeIdxPa
     if(num_groups == 1)
     {
         // Create convolution node and connect
-        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, method);
+        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, 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)
@@ -270,7 +278,7 @@ NodeID GraphBuilder::add_depth_concatenate_node(Graph &g, NodeParams params, std
 
 NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params, NodeIdxPair input, Size2D kernel_spatial_extend, PadStrideInfo conv_info,
                                                     DepthwiseConvolutionMethod method,
-                                                    ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor)
+                                                    ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor, const QuantizationInfo quant_info)
 {
     CHECK_NODEIDX_PAIR(input, g);
     ARM_COMPUTE_ERROR_ON((kernel_spatial_extend.width == 0) || (kernel_spatial_extend.height == 0));
@@ -283,7 +291,13 @@ NodeID GraphBuilder::add_depthwise_convolution_node(Graph &g, NodeParams params,
     // Create weights node
     TensorDescriptor w_desc = input_tensor_desc;
     w_desc.shape            = TensorShape(kernel_spatial_extend.width, kernel_spatial_extend.height, w_desc.shape.z());
-    NodeID w_nid            = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
+
+    if(!quant_info.empty())
+    {
+        w_desc.quant_info = quant_info;
+    }
+
+    NodeID w_nid = add_const_node_with_name(g, params, "Weights", w_desc, std::move(weights_accessor));
 
     // Create bias nodes
     NodeID b_nid = EmptyNodeID;
diff --git a/src/graph/backends/CL/CLDeviceBackend.cpp b/src/graph/backends/CL/CLDeviceBackend.cpp
index f10eb33a98..92cb6936c3 100644
--- a/src/graph/backends/CL/CLDeviceBackend.cpp
+++ b/src/graph/backends/CL/CLDeviceBackend.cpp
@@ -126,7 +126,7 @@ std::unique_ptr<ITensorHandle> CLDeviceBackend::create_tensor(const Tensor &tens
     ARM_COMPUTE_ERROR_ON(tensor_desc.target != Target::CL);
 
     // Create backend tensor handle
-    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type);
+    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type, tensor_desc.quant_info);
     auto       backend_tensor_handle = support::cpp14::make_unique<CLTensorHandle>(info);
 
     return std::move(backend_tensor_handle);
diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index 1b448fefd2..ad73a797e3 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -154,10 +154,16 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    ICLTensor              *input          = get_backing_tensor(node.input(0));
-    ICLTensor              *weights        = get_backing_tensor(node.input(1));
-    ICLTensor              *biases         = get_backing_tensor(node.input(2));
-    ICLTensor              *output         = get_backing_tensor(node.output(0));
+    ICLTensor *input   = get_backing_tensor(node.input(0));
+    ICLTensor *weights = get_backing_tensor(node.input(1));
+    ICLTensor *biases  = get_backing_tensor(node.input(2));
+    ICLTensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo     conv_info      = node.convolution_info();
     const ConvolutionMethod conv_algorithm = node.convolution_method();
 
@@ -190,6 +196,8 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
@@ -251,10 +259,16 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    ICLTensor                       *input         = get_backing_tensor(node.input(0));
-    ICLTensor                       *weights       = get_backing_tensor(node.input(1));
-    ICLTensor                       *biases        = get_backing_tensor(node.input(2));
-    ICLTensor                       *output        = get_backing_tensor(node.output(0));
+    ICLTensor *input   = get_backing_tensor(node.input(0));
+    ICLTensor *weights = get_backing_tensor(node.input(1));
+    ICLTensor *biases  = get_backing_tensor(node.input(2));
+    ICLTensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo              conv_info     = node.convolution_info();
     const DepthwiseConvolutionMethod dwc_algorithm = node.depthwise_convolution_method();
 
@@ -275,6 +289,8 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
diff --git a/src/graph/backends/GLES/GCDeviceBackend.cpp b/src/graph/backends/GLES/GCDeviceBackend.cpp
index 8cd9994744..a55215f058 100644
--- a/src/graph/backends/GLES/GCDeviceBackend.cpp
+++ b/src/graph/backends/GLES/GCDeviceBackend.cpp
@@ -87,7 +87,7 @@ std::unique_ptr<ITensorHandle> GCDeviceBackend::create_tensor(const Tensor &tens
     ARM_COMPUTE_ERROR_ON(tensor_desc.target != Target::GC);
 
     // Create backend tensor handle
-    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type);
+    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type, tensor_desc.quant_info);
     auto       backend_tensor_handle = support::cpp14::make_unique<GCTensorHandle>(info);
 
     return std::move(backend_tensor_handle);
diff --git a/src/graph/backends/GLES/GCFunctionsFactory.cpp b/src/graph/backends/GLES/GCFunctionsFactory.cpp
index 12e7c042d4..d3c5737e68 100644
--- a/src/graph/backends/GLES/GCFunctionsFactory.cpp
+++ b/src/graph/backends/GLES/GCFunctionsFactory.cpp
@@ -154,10 +154,16 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    IGCTensor              *input          = get_backing_tensor(node.input(0));
-    IGCTensor              *weights        = get_backing_tensor(node.input(1));
-    IGCTensor              *biases         = get_backing_tensor(node.input(2));
-    IGCTensor              *output         = get_backing_tensor(node.output(0));
+    IGCTensor *input   = get_backing_tensor(node.input(0));
+    IGCTensor *weights = get_backing_tensor(node.input(1));
+    IGCTensor *biases  = get_backing_tensor(node.input(2));
+    IGCTensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo     conv_info      = node.convolution_info();
     const ConvolutionMethod conv_algorithm = node.convolution_method();
 
@@ -180,6 +186,8 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
@@ -241,10 +249,16 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    IGCTensor                       *input         = get_backing_tensor(node.input(0));
-    IGCTensor                       *weights       = get_backing_tensor(node.input(1));
-    IGCTensor                       *biases        = get_backing_tensor(node.input(2));
-    IGCTensor                       *output        = get_backing_tensor(node.output(0));
+    IGCTensor *input   = get_backing_tensor(node.input(0));
+    IGCTensor *weights = get_backing_tensor(node.input(1));
+    IGCTensor *biases  = get_backing_tensor(node.input(2));
+    IGCTensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo              conv_info     = node.convolution_info();
     const DepthwiseConvolutionMethod dwc_algorithm = node.depthwise_convolution_method();
 
@@ -264,6 +278,8 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
diff --git a/src/graph/backends/NEON/NEDeviceBackend.cpp b/src/graph/backends/NEON/NEDeviceBackend.cpp
index aaf05829bb..9123196540 100644
--- a/src/graph/backends/NEON/NEDeviceBackend.cpp
+++ b/src/graph/backends/NEON/NEDeviceBackend.cpp
@@ -93,7 +93,7 @@ std::unique_ptr<ITensorHandle> NEDeviceBackend::create_tensor(const Tensor &tens
     ARM_COMPUTE_ERROR_ON(tensor_desc.target != Target::NEON);
 
     // Create backend tensor handle
-    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type);
+    TensorInfo info(tensor_desc.shape, 1, tensor_desc.data_type, tensor_desc.quant_info);
     auto       backend_tensor_handle = support::cpp14::make_unique<NETensorHandle>(info);
 
     return std::move(backend_tensor_handle);
diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp
index 228af9ca6f..906378c565 100644
--- a/src/graph/backends/NEON/NEFunctionFactory.cpp
+++ b/src/graph/backends/NEON/NEFunctionFactory.cpp
@@ -140,10 +140,16 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    ITensor                *input          = get_backing_tensor(node.input(0));
-    ITensor                *weights        = get_backing_tensor(node.input(1));
-    ITensor                *biases         = get_backing_tensor(node.input(2));
-    ITensor                *output         = get_backing_tensor(node.output(0));
+    ITensor *input   = get_backing_tensor(node.input(0));
+    ITensor *weights = get_backing_tensor(node.input(1));
+    ITensor *biases  = get_backing_tensor(node.input(2));
+    ITensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo     conv_info      = node.convolution_info();
     const ConvolutionMethod conv_algorithm = node.convolution_method();
 
@@ -175,6 +181,8 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node,
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
@@ -234,10 +242,16 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    ITensor                         *input         = get_backing_tensor(node.input(0));
-    ITensor                         *weights       = get_backing_tensor(node.input(1));
-    ITensor                         *biases        = get_backing_tensor(node.input(2));
-    ITensor                         *output        = get_backing_tensor(node.output(0));
+    ITensor *input   = get_backing_tensor(node.input(0));
+    ITensor *weights = get_backing_tensor(node.input(1));
+    ITensor *biases  = get_backing_tensor(node.input(2));
+    ITensor *output  = get_backing_tensor(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->info()->data_type()))
+    {
+        biases->info()->set_data_type(DataType::S32);
+    }
+
     const PadStrideInfo              conv_info     = node.convolution_info();
     const DepthwiseConvolutionMethod dwc_algorithm = node.depthwise_convolution_method();
 
@@ -258,6 +272,8 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << func_name
                                << " Data Type: " << input->info()->data_type()
+                               << " Input QuantInfo: " << input->info()->quantization_info()
+                               << " Weights QuantInfo: " << weights->info()->quantization_info()
                                << " Input shape: " << input->info()->tensor_shape()
                                << " Weights shape: " << weights->info()->tensor_shape()
                                << " Output shape: " << output->info()->tensor_shape()
diff --git a/src/graph/nodes/ConvolutionLayerNode.cpp b/src/graph/nodes/ConvolutionLayerNode.cpp
index 461728487f..eb0c6a1c1a 100644
--- a/src/graph/nodes/ConvolutionLayerNode.cpp
+++ b/src/graph/nodes/ConvolutionLayerNode.cpp
@@ -31,8 +31,8 @@ namespace arm_compute
 {
 namespace graph
 {
-ConvolutionLayerNode::ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method)
-    : _info(std::move(info)), _method(method)
+ConvolutionLayerNode::ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method, QuantizationInfo out_quant_info)
+    : _info(std::move(info)), _method(method), _out_quant_info(out_quant_info)
 {
     _input_edges.resize(3, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -90,6 +90,12 @@ TensorDescriptor ConvolutionLayerNode::configure_output(size_t idx) const
     TensorDescriptor output_info  = src->desc();
     TensorShape      output_shape = compute_output_shape(src->desc().shape, weights->desc().shape, _info);
     output_info.shape             = output_shape;
+
+    if(!_out_quant_info.empty())
+    {
+        output_info.quant_info = _out_quant_info;
+    }
+
     return output_info;
 }
 
diff --git a/src/graph/nodes/SoftmaxLayerNode.cpp b/src/graph/nodes/SoftmaxLayerNode.cpp
index 4c21ac6ad0..b6241e6654 100644
--- a/src/graph/nodes/SoftmaxLayerNode.cpp
+++ b/src/graph/nodes/SoftmaxLayerNode.cpp
@@ -63,7 +63,10 @@ TensorDescriptor SoftmaxLayerNode::configure_output(size_t idx) const
     const Tensor *src = input(0);
     ARM_COMPUTE_ERROR_ON(src == nullptr);
 
-    return src->desc();
+    TensorDescriptor out_desc = src->desc();
+    out_desc.quant_info       = QuantizationInfo(1.f / 256.f, 0);
+
+    return out_desc;
 }
 
 Status SoftmaxLayerNode::validate()