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>

6 files changed, 89 insertions, 54 deletions

diff --git a/arm_compute/graph/GraphBuilder.h b/arm_compute/graph/GraphBuilder.h
index e01cfb1bc6..bbbbefbcbb 100644
--- a/arm_compute/graph/GraphBuilder.h
+++ b/arm_compute/graph/GraphBuilder.h
@@ -109,13 +109,17 @@ public:
      * @param[in] method                (Optional) Convolution method to use
      * @param[in] weights_accessor      (Optional) Accessor of the weights node data
      * @param[in] bias_accessor         (Optional) Accessor of the bias node data
+     * @param[in] weights_quant_info    (Optional) Weights quantization info
+     * @param[in] out_quant_info        (Optional) Output quantization info
      *
      * @return Node ID of the created node, EmptyNodeID in case of error
      */
     static NodeID add_convolution_node(Graph &g, NodeParams params, NodeIdxPair input,
                                        Size2D kernel_spatial_extend, unsigned int depth, PadStrideInfo conv_info,
                                        unsigned int num_groups = 1, ConvolutionMethod method = ConvolutionMethod::DEFAULT,
-                                       ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr);
+                                       ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr,
+                                       const QuantizationInfo weights_quant_info = QuantizationInfo(),
+                                       const QuantizationInfo out_quant_info     = QuantizationInfo());
     /** Adds a depth concatenate node to the graph
      *
      * @param[in] g      Graph to add the node to
@@ -135,13 +139,14 @@ public:
      * @param[in] method                (Optional) Convolution method to use
      * @param[in] weights_accessor      (Optional) Accessor of the weights node data
      * @param[in] bias_accessor         (Optional) Accessor of the bias node data
+     * @param[in] quant_info            (Optional) Weights quantization info
      *
      * @return Node ID of the created node, EmptyNodeID in case of error
      */
     static NodeID add_depthwise_convolution_node(Graph &g, NodeParams params, NodeIdxPair input,
                                                  Size2D kernel_spatial_extend, PadStrideInfo conv_info,
                                                  DepthwiseConvolutionMethod method    = DepthwiseConvolutionMethod::DEFAULT,
-                                                 ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr);
+                                                 ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr, const QuantizationInfo quant_info = QuantizationInfo());
     /** Adds an element-wise layer node to the graph
      *
      * @param[in] g         Graph to add the node to
diff --git a/arm_compute/graph/TensorDescriptor.h b/arm_compute/graph/TensorDescriptor.h
index fc0095e3b9..785c493cbc 100644
--- a/arm_compute/graph/TensorDescriptor.h
+++ b/arm_compute/graph/TensorDescriptor.h
@@ -37,18 +37,20 @@ struct TensorDescriptor final
     TensorDescriptor() = default;
     /** Constructor
      *
-     * @param[in] tensor_shape     Tensor shape
-     * @param[in] tensor_data_type Tensor data type
-     * @param[in] tensor_target    Target to allocate the tensor for
+     * @param[in] tensor_shape      Tensor shape
+     * @param[in] tensor_data_type  Tensor data type
+     * @param[in] tensor_quant_info Tensor quantization info
+     * @param[in] tensor_target     Target to allocate the tensor for
      */
-    TensorDescriptor(TensorShape tensor_shape, DataType tensor_data_type, Target tensor_target = Target::UNSPECIFIED)
-        : shape(tensor_shape), data_type(tensor_data_type), target(tensor_target)
+    TensorDescriptor(TensorShape tensor_shape, DataType tensor_data_type, QuantizationInfo tensor_quant_info = QuantizationInfo(), Target tensor_target = Target::UNSPECIFIED)
+        : shape(tensor_shape), data_type(tensor_data_type), quant_info(tensor_quant_info), target(tensor_target)
     {
     }
 
-    TensorShape shape{};                        /**< Tensor shape */
-    DataType    data_type{ DataType::UNKNOWN }; /**< Data type */
-    Target      target{ Target::UNSPECIFIED };  /**< Target */
+    TensorShape      shape{};                        /**< Tensor shape */
+    DataType         data_type{ DataType::UNKNOWN }; /**< Data type */
+    QuantizationInfo quant_info{};                   /**< Quantization info */
+    Target           target{ Target::UNSPECIFIED };  /**< Target */
 };
 } // namespace graph
 } // namespace arm_compute
diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h
index d9bc8376bd..ed578307b5 100644
--- a/arm_compute/graph/TypePrinter.h
+++ b/arm_compute/graph/TypePrinter.h
@@ -308,6 +308,14 @@ inline ::std::ostream &operator<<(::std::ostream &os, const PadStrideInfo &pad_s
 
     return os;
 }
+
+/** Formatted output of the QuantizationInfo type. */
+inline ::std::ostream &operator<<(::std::ostream &os, const QuantizationInfo &quantization_info)
+{
+    os << "Scale:" << quantization_info.scale << "~"
+       << "Offset:" << quantization_info.offset;
+    return os;
+}
 } // namespace graph
 } // namespace arm_compute
 #endif /* __ARM_COMPUTE_GRAPH_TYPE_PRINTER_H__ */
diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h
index ca01295d15..68a718ab97 100644
--- a/arm_compute/graph/backends/ValidateHelpers.h
+++ b/arm_compute/graph/backends/ValidateHelpers.h
@@ -70,12 +70,18 @@ Status validate_convolution_layer(ConvolutionLayerNode &node)
     ARM_COMPUTE_RETURN_ERROR_ON(node.num_outputs() != 1);
 
     // Extract IO and info
-    arm_compute::ITensorInfo *input          = get_backing_tensor_info(node.input(0));
-    arm_compute::ITensorInfo *weights        = get_backing_tensor_info(node.input(1));
-    arm_compute::ITensorInfo *biases         = get_backing_tensor_info(node.input(2));
-    arm_compute::ITensorInfo *output         = get_backing_tensor_info(node.output(0));
-    const PadStrideInfo       conv_info      = node.convolution_info();
-    const ConvolutionMethod   conv_algorithm = node.convolution_method();
+    arm_compute::ITensorInfo *input   = get_backing_tensor_info(node.input(0));
+    arm_compute::ITensorInfo *weights = get_backing_tensor_info(node.input(1));
+    arm_compute::ITensorInfo *biases  = get_backing_tensor_info(node.input(2));
+    arm_compute::ITensorInfo *output  = get_backing_tensor_info(node.output(0));
+
+    if(is_data_type_quantized_asymmetric(input->data_type()))
+    {
+        biases->set_data_type(DataType::S32);
+    }
+
+    const PadStrideInfo     conv_info      = node.convolution_info();
+    const ConvolutionMethod conv_algorithm = node.convolution_method();
 
     // Validate function
     Status status{};
diff --git a/arm_compute/graph/frontend/Layers.h b/arm_compute/graph/frontend/Layers.h
index 22133b8376..2e7c50e6da 100644
--- a/arm_compute/graph/frontend/Layers.h
+++ b/arm_compute/graph/frontend/Layers.h
@@ -160,28 +160,34 @@ class ConvolutionLayer final : public ILayer
 public:
     /** Construct a convolution layer.
      *
-     * @param[in] conv_width  Convolution width.
-     * @param[in] conv_height Convolution height.
-     * @param[in] ofm         Output feature map.
-     * @param[in] weights     Accessor to get kernel weights from.
-     * @param[in] bias        Accessor to get kernel bias from.
-     * @param[in] conv_info   Padding and stride information.
-     * @param[in] num_groups  (Optional) Number of groups. Default: 1.
+     * @param[in] conv_width         Convolution width.
+     * @param[in] conv_height        Convolution height.
+     * @param[in] ofm                Output feature map.
+     * @param[in] weights            Accessor to get kernel weights from.
+     * @param[in] bias               Accessor to get kernel bias from.
+     * @param[in] conv_info          Padding and stride information.
+     * @param[in] num_groups         (Optional) Number of groups. Default: 1.
+     * @param[in] weights_quant_info (Optional) Weights quantization information
+     * @param[in] out_quant_info     (Optional) Output quantization info
      */
-    ConvolutionLayer(unsigned int        conv_width,
-                     unsigned int        conv_height,
-                     unsigned int        ofm,
-                     ITensorAccessorUPtr weights,
-                     ITensorAccessorUPtr bias,
-                     PadStrideInfo       conv_info,
-                     unsigned int        num_groups = 1)
+    ConvolutionLayer(unsigned int           conv_width,
+                     unsigned int           conv_height,
+                     unsigned int           ofm,
+                     ITensorAccessorUPtr    weights,
+                     ITensorAccessorUPtr    bias,
+                     PadStrideInfo          conv_info,
+                     unsigned int           num_groups         = 1,
+                     const QuantizationInfo weights_quant_info = QuantizationInfo(),
+                     const QuantizationInfo out_quant_info     = QuantizationInfo())
         : _conv_width(conv_width),
           _conv_height(conv_height),
           _ofm(ofm),
           _conv_info(std::move(conv_info)),
           _num_groups(num_groups),
           _weights(std::move(weights)),
-          _bias(std::move(bias))
+          _bias(std::move(bias)),
+          _weights_quant_info(std::move(weights_quant_info)),
+          _out_quant_info(std::move(out_quant_info))
     {
     }
 
@@ -192,17 +198,19 @@ public:
         return GraphBuilder::add_convolution_node(s.graph(), common_params, input,
                                                   Size2D(_conv_width, _conv_height), _ofm, _conv_info, _num_groups,
                                                   s.hints().convolution_method_hint,
-                                                  std::move(_weights), std::move(_bias));
+                                                  std::move(_weights), std::move(_bias), std::move(_weights_quant_info), std::move(_out_quant_info));
     }
 
 private:
-    unsigned int        _conv_width;
-    unsigned int        _conv_height;
-    unsigned int        _ofm;
-    const PadStrideInfo _conv_info;
-    unsigned int        _num_groups;
-    ITensorAccessorUPtr _weights;
-    ITensorAccessorUPtr _bias;
+    unsigned int           _conv_width;
+    unsigned int           _conv_height;
+    unsigned int           _ofm;
+    const PadStrideInfo    _conv_info;
+    unsigned int           _num_groups;
+    ITensorAccessorUPtr    _weights;
+    ITensorAccessorUPtr    _bias;
+    const QuantizationInfo _weights_quant_info;
+    const QuantizationInfo _out_quant_info;
 };
 
 /** Depthwise Convolution Layer */
@@ -216,17 +224,20 @@ public:
      * @param[in] weights     Accessor to get kernel weights from.
      * @param[in] bias        Accessor to get kernel bias from.
      * @param[in] conv_info   Padding and stride information.
+     * @param[in] quant_info  (Optional) Quantization info used for weights
      */
-    DepthwiseConvolutionLayer(unsigned int        conv_width,
-                              unsigned int        conv_height,
-                              ITensorAccessorUPtr weights,
-                              ITensorAccessorUPtr bias,
-                              PadStrideInfo       conv_info)
+    DepthwiseConvolutionLayer(unsigned int           conv_width,
+                              unsigned int           conv_height,
+                              ITensorAccessorUPtr    weights,
+                              ITensorAccessorUPtr    bias,
+                              PadStrideInfo          conv_info,
+                              const QuantizationInfo quant_info = QuantizationInfo())
         : _conv_width(conv_width),
           _conv_height(conv_height),
           _conv_info(std::move(conv_info)),
           _weights(std::move(weights)),
-          _bias(std::move(bias))
+          _bias(std::move(bias)),
+          _quant_info(std::move(quant_info))
     {
     }
 
@@ -237,15 +248,16 @@ public:
         return GraphBuilder::add_depthwise_convolution_node(s.graph(), common_params,
                                                             input, Size2D(_conv_width, _conv_height), _conv_info,
                                                             s.hints().depthwise_convolution_method_hint,
-                                                            std::move(_weights), std::move(_bias));
+                                                            std::move(_weights), std::move(_bias), std::move(_quant_info));
     }
 
 private:
-    unsigned int        _conv_width;
-    unsigned int        _conv_height;
-    const PadStrideInfo _conv_info;
-    ITensorAccessorUPtr _weights;
-    ITensorAccessorUPtr _bias;
+    unsigned int           _conv_width;
+    unsigned int           _conv_height;
+    const PadStrideInfo    _conv_info;
+    ITensorAccessorUPtr    _weights;
+    ITensorAccessorUPtr    _bias;
+    const QuantizationInfo _quant_info;
 };
 
 /** Flatten Layer */
diff --git a/arm_compute/graph/nodes/ConvolutionLayerNode.h b/arm_compute/graph/nodes/ConvolutionLayerNode.h
index 70fefbeeab..d029895609 100644
--- a/arm_compute/graph/nodes/ConvolutionLayerNode.h
+++ b/arm_compute/graph/nodes/ConvolutionLayerNode.h
@@ -36,10 +36,11 @@ class ConvolutionLayerNode final : public INode
 public:
     /** Constructor
      *
-     * @param[in] info   Convolution layer attributes
-     * @param[in] method (Optional) Convolution method to use
+     * @param[in] info           Convolution layer attributes
+     * @param[in] method         (Optional) Convolution method to use
+     * @param[in] out_quant_info (Optional) Output quantization info
      */
-    ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method = ConvolutionMethod::DEFAULT);
+    ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method = ConvolutionMethod::DEFAULT, QuantizationInfo out_quant_info = QuantizationInfo());
     /** Sets the convolution layer method to use
      *
      * @param[in] method Method to use for convolution
@@ -78,6 +79,7 @@ public:
 private:
     PadStrideInfo     _info;
     ConvolutionMethod _method;
+    QuantizationInfo  _out_quant_info;
 };
 } // namespace graph
 } // namespace arm_compute