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>
diff --git a/arm_compute/graph/GraphBuilder.h b/arm_compute/graph/GraphBuilder.h
index 191848c..5bb1df4 100644
--- a/arm_compute/graph/GraphBuilder.h
+++ b/arm_compute/graph/GraphBuilder.h
@@ -212,17 +212,21 @@
static NodeID add_flatten_node(Graph &g, NodeParams params, NodeIdxPair input);
/** Adds a fully connected layer node to the graph
*
- * @param[in] g Graph to add the layer to
- * @param[in] params Common node parameters
- * @param[in] input Input to the fully connected layer node as a NodeID-Index pair
- * @param[in] num_outputs Number of output neurons
- * @param[in] weights_accessor (Optional) Accessor of the weights node data
- * @param[in] bias_accessor (Optional) Accessor of the bias node data
+ * @param[in] g Graph to add the layer to
+ * @param[in] params Common node parameters
+ * @param[in] input Input to the fully connected layer node as a NodeID-Index pair
+ * @param[in] num_outputs Number of output neurons
+ * @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_fully_connected_layer(Graph &g, NodeParams params, NodeIdxPair input, unsigned int num_outputs,
- 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 normalization layer node to the graph
*
* @param[in] g Graph to add the node to
diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h
index 172f002..b7c9c57 100644
--- a/arm_compute/graph/backends/FunctionHelpers.h
+++ b/arm_compute/graph/backends/FunctionHelpers.h
@@ -541,6 +541,8 @@
ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type()
<< " Target " << TargetInfo::TargetType
<< " 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/arm_compute/graph/frontend/Layers.h b/arm_compute/graph/frontend/Layers.h
index 02ef569..a222c85 100644
--- a/arm_compute/graph/frontend/Layers.h
+++ b/arm_compute/graph/frontend/Layers.h
@@ -381,14 +381,22 @@
public:
/** Construct a fully connected layer.
*
- * @param[in] num_outputs Number of outputs.
- * @param[in] weights Accessor to get weights from.
- * @param[in] bias Accessor to get bias from.
+ * @param[in] num_outputs Number of outputs.
+ * @param[in] weights Accessor to get weights from.
+ * @param[in] bias Accessor to get bias from.
+ * @param[in] weights_quant_info (Optional) Weights quantization information
+ * @param[in] out_quant_info (Optional) Output quantization info
*/
- FullyConnectedLayer(unsigned int num_outputs,
- ITensorAccessorUPtr weights,
- ITensorAccessorUPtr bias)
- : _num_outputs(num_outputs), _weights(std::move(weights)), _bias(std::move(bias))
+ FullyConnectedLayer(unsigned int num_outputs,
+ ITensorAccessorUPtr weights,
+ ITensorAccessorUPtr bias,
+ const QuantizationInfo weights_quant_info = QuantizationInfo(),
+ const QuantizationInfo out_quant_info = QuantizationInfo())
+ : _num_outputs(num_outputs),
+ _weights(std::move(weights)),
+ _bias(std::move(bias)),
+ _weights_quant_info(std::move(weights_quant_info)),
+ _out_quant_info(std::move(out_quant_info))
{
}
@@ -397,13 +405,16 @@
NodeParams common_params = { name(), s.hints().target_hint };
NodeIdxPair input = { s.tail_node(), 0 };
return GraphBuilder::add_fully_connected_layer(s.graph(), common_params, input, _num_outputs,
- 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 _num_outputs;
- ITensorAccessorUPtr _weights;
- ITensorAccessorUPtr _bias;
+ unsigned int _num_outputs;
+ ITensorAccessorUPtr _weights;
+ ITensorAccessorUPtr _bias;
+ const QuantizationInfo _weights_quant_info;
+ const QuantizationInfo _out_quant_info;
};
/** Normalization Layer */
diff --git a/arm_compute/graph/nodes/FullyConnectedLayerNode.h b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
index 79201c8..1bff600 100644
--- a/arm_compute/graph/nodes/FullyConnectedLayerNode.h
+++ b/arm_compute/graph/nodes/FullyConnectedLayerNode.h
@@ -36,30 +36,39 @@
public:
/** Constructor
*
- * @param[in] num_outputs Number of neurons in the layer
- * @param[in] fc_info (Optional) Additional information about the fully connected layer
+ * @param[in] num_outputs Number of neurons in the layer
+ * @param[in] out_quant_info (Optional) Output quantization info
+ * @param[in] fc_info (Optional) Additional information about the fully connected layer
*/
- FullyConnectedLayerNode(unsigned int num_outputs, FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
+ FullyConnectedLayerNode(unsigned int num_outputs,
+ QuantizationInfo out_quant_info = QuantizationInfo(),
+ FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
/** Computes weights descriptor
*
* @warning Works for inputs with 1D batch space
*
- * @param[in] input_descriptor Input descriptor
- * @param[in] num_outputs Number of output neurons
+ * @param[in] input_descriptor Input descriptor
+ * @param[in] num_outputs Number of output neurons
+ * @param[in] weights_quant_info (Optional) Weights quantization info
*
* @return Weights descriptor
*/
- static TensorDescriptor compute_weights_descriptor(const TensorDescriptor &input_descriptor, unsigned int num_outputs);
+ static TensorDescriptor compute_weights_descriptor(const TensorDescriptor &input_descriptor,
+ unsigned int num_outputs,
+ QuantizationInfo weights_quant_info = QuantizationInfo());
/** Computes fully connected layer output descriptor
*
* @warning Works for inputs with 1D batch space
*
* @param[in] input_descriptor Input descriptor
* @param[in] num_outputs Number of output neurons
+ * @param[in] out_quant_info (Optional) Weights quantization info
*
* @return Output descriptor
*/
- static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor, unsigned int num_outputs);
+ static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor,
+ unsigned int num_outputs,
+ QuantizationInfo out_quant_info = QuantizationInfo());
/** Fully connected layer addition information
*
* @return Additional information about the fully connected layer
@@ -74,6 +83,7 @@
private:
unsigned int _num_outputs;
+ QuantizationInfo _out_quant_info;
FullyConnectedLayerInfo _info;
};
} // namespace graph