diff options
22 files changed, 274 insertions, 99 deletions
diff --git a/arm_compute/graph/GraphBuilder.h b/arm_compute/graph/GraphBuilder.h index a2a938b1cc..590e4d9b44 100644 --- a/arm_compute/graph/GraphBuilder.h +++ b/arm_compute/graph/GraphBuilder.h @@ -25,6 +25,7 @@ #define __ARM_COMPUTE_GRAPH_GRAPH_BUILDER_H__ #include "arm_compute/graph/ITensorAccessor.h" +#include "arm_compute/graph/LayerDescriptors.h" #include "arm_compute/graph/Types.h" namespace arm_compute @@ -73,14 +74,16 @@ public: static NodeID add_output_node(Graph &g, NodeParams params, NodeIdxPair input, ITensorAccessorUPtr accessor = nullptr); /** Adds an activation layer node to the graph * - * @param[in] g Graph to add the node to - * @param[in] params Common node parameters - * @param[in] input Input to the activation layer node as a NodeID-Index pair - * @param[in] act_info Activation layer information + * @param[in] g Graph to add the node to + * @param[in] params Common node parameters + * @param[in] input Input to the activation layer node as a NodeID-Index pair + * @param[in] act_info Activation layer information + * @param[in] out_quant_info (Optional) Output quantization info * * @return Node ID of the created node, EmptyNodeID in case of error */ - static NodeID add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info); + static NodeID add_activation_node(Graph &g, NodeParams params, NodeIdxPair input, ActivationLayerInfo act_info, + const QuantizationInfo out_quant_info = QuantizationInfo()); /** Adds a batch normalization layer node to the graph * * @param[in] g Graph to add the node to @@ -163,14 +166,14 @@ public: ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr); /** Adds a depth concatenate node to the graph * - * @param[in] g Graph to add the node to - * @param[in] params Common node parameters - * @param[in] inputs Inputs to the depth concatenate layer node as a NodeID-Index pair - * @param[in] axis Concatenation axis + * @param[in] g Graph to add the node to + * @param[in] params Common node parameters + * @param[in] inputs Inputs to the depth concatenate layer node as a NodeID-Index pair + * @param[in] concat_descriptor Concatenation layer descriptor * * @return Node ID of the created node, EmptyNodeID in case of error */ - static NodeID add_concatenate_node(Graph &g, NodeParams params, std::vector<NodeIdxPair> inputs, DataLayoutDimension axis); + static NodeID add_concatenate_node(Graph &g, NodeParams params, std::vector<NodeIdxPair> inputs, descriptors::ConcatLayerDescriptor concat_descriptor); /** Adds a depth-wise convolution layer node to the graph * * @param[in] g Graph to add the node to @@ -183,13 +186,15 @@ public: * @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 + * @param[in] out_quant_info (Optional) Output 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, int depth_multiplier = 1, DepthwiseConvolutionMethod method = DepthwiseConvolutionMethod::Default, - ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr, const QuantizationInfo quant_info = QuantizationInfo()); + ITensorAccessorUPtr weights_accessor = nullptr, ITensorAccessorUPtr bias_accessor = nullptr, const QuantizationInfo quant_info = QuantizationInfo(), + const QuantizationInfo out_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/INode.h b/arm_compute/graph/INode.h index 4219150f58..edff8379d3 100644 --- a/arm_compute/graph/INode.h +++ b/arm_compute/graph/INode.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,7 @@ #define __ARM_COMPUTE_GRAPH_INODE_H__ #include "arm_compute/core/Error.h" +#include "arm_compute/graph/LayerDescriptors.h" #include "arm_compute/graph/TensorDescriptor.h" #include "arm_compute/graph/Types.h" diff --git a/arm_compute/graph/LayerDescriptors.h b/arm_compute/graph/LayerDescriptors.h new file mode 100644 index 0000000000..79099326ec --- /dev/null +++ b/arm_compute/graph/LayerDescriptors.h @@ -0,0 +1,69 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_CONCAT_DESCRIPTOR_H__ +#define __ARM_COMPUTE_CONCAT_DESCRIPTOR_H__ + +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace graph +{ +namespace descriptors +{ +/** Common node parameters */ +struct ConcatLayerDescriptor +{ + /** Default constructor */ + ConcatLayerDescriptor() + : axis(DataLayoutDimension::CHANNEL), output_qinfo() + { + } + + /** Constructor concatenate layer descriptor + * + * @param[in] axis Axis. + */ + ConcatLayerDescriptor(DataLayoutDimension axis) + : axis(axis), output_qinfo() + { + } + + /** Constructor concatenate layer descriptor + * + * @param[in] axis Axis. + * @param[in] output_qinfo Output quantization info. + */ + ConcatLayerDescriptor(DataLayoutDimension axis, QuantizationInfo output_qinfo) + : axis(axis), output_qinfo(output_qinfo) + { + } + + const DataLayoutDimension axis; /**< Concatenation Axis */ + const QuantizationInfo output_qinfo; /**< Output quantizazion info */ +}; +} // namespace descriptor +} // namespace graph +} // namespace arm_compute +#endif /* __ARM_COMPUTE_CONCAT_DESCRIPTOR_H__ */
\ No newline at end of file diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h index d0035d9a84..4a423d2490 100644 --- a/arm_compute/graph/backends/FunctionHelpers.h +++ b/arm_compute/graph/backends/FunctionHelpers.h @@ -109,7 +109,7 @@ std::unique_ptr<IFunction> create_activation_layer(ActivationLayerNode &node) ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.name() << " Type: " << node.type() - << " Target " << TargetInfo::TargetType + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() << " Activation function: " << act_info.activation() @@ -245,8 +245,10 @@ std::unique_ptr<IFunction> create_bounding_box_transform_layer(BoundingBoxTransf func->configure(input, output, deltas, bbox_info); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Shape: " << input->info()->tensor_shape() << " BoundingBox Info img W: " << bbox_info.img_width() << " " @@ -326,6 +328,12 @@ std::unique_ptr<arm_compute::IFunction> create_concatenate_layer(ConcatenateLaye func->configure(inputs, output, concat_axis); // Log info + const bool is_quantized = is_data_type_quantized_asymmetric(output->info()->data_type()); + std::ostringstream qss; + if(is_quantized) + { + qss << " Output QuantInfo: " << output->info()->quantization_info(); + } ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.name() << " Type: " << node.type() @@ -334,6 +342,7 @@ std::unique_ptr<arm_compute::IFunction> create_concatenate_layer(ConcatenateLaye << " Shape: " << output->info()->tensor_shape() << " Num Inputs: " << inputs.size() << " Axis: " << concat_axis + << qss.str() << std::endl); return std::move(func); @@ -421,10 +430,10 @@ std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node, << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Groups: " << num_groups - << qss.str() << " Input shape: " << input->info()->tensor_shape() << " Weights shape: " << weights->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() + << qss.str() << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") << std::endl); return func; @@ -536,11 +545,11 @@ std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvoluti << " Type: " << func_name << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() - << qss.str() << " Input shape: " << input->info()->tensor_shape() << " Weights shape: " << weights->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() << " Depth multiplier: " << depth_multiplier + << qss.str() << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") << std::endl); return func; @@ -1177,8 +1186,10 @@ std::unique_ptr<IFunction> create_roi_align_layer(ROIAlignLayerNode &node) func->configure(input, rois, output, pool_info); // Log info - ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type() - << " Target " << TargetInfo::TargetType + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.type() + << " Target: " << TargetInfo::TargetType << " Data Type: " << input->info()->data_type() << " Input shape: " << input->info()->tensor_shape() << " Output shape: " << output->info()->tensor_shape() diff --git a/arm_compute/graph/backends/ValidateHelpers.h b/arm_compute/graph/backends/ValidateHelpers.h index 8942be2da8..dbf8f35121 100644 --- a/arm_compute/graph/backends/ValidateHelpers.h +++ b/arm_compute/graph/backends/ValidateHelpers.h @@ -203,6 +203,7 @@ Status validate_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node) return status; } + /** Validates a detection output layer node * * @tparam DetectionOutputLayer DetectionOutput layer type @@ -372,6 +373,29 @@ Status validate_reorg_layer(ReorgLayerNode &node) return ReorgLayer::validate(input, output, node.stride()); } +/** Validates a Reshape layer node + * + * @tparam ReshapeLayer Reshape layer type + * + * @param[in] node Node to validate + * + * @return Status + */ +template <typename ReshapeLayer> +Status validate_reshape_layer(ReshapeLayerNode &node) +{ + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Validating ReshapeLayer node with ID : " << node.id() << " and Name: " << node.name() << std::endl); + ARM_COMPUTE_RETURN_ERROR_ON(node.num_inputs() != 1); + ARM_COMPUTE_RETURN_ERROR_ON(node.num_outputs() != 1); + + // Extract input and output + arm_compute::ITensorInfo *input = detail::get_backing_tensor_info(node.input(0)); + arm_compute::ITensorInfo *output = detail::get_backing_tensor_info(node.output(0)); + + // Validate function + return ReshapeLayer::validate(input, output); +} + /** Validates a ROI Align layer node * * @tparam ROIAlignLayer ROIAlign layer type diff --git a/arm_compute/graph/frontend/Layers.h b/arm_compute/graph/frontend/Layers.h index d062d5a53e..67dc06c878 100644 --- a/arm_compute/graph/frontend/Layers.h +++ b/arm_compute/graph/frontend/Layers.h @@ -72,22 +72,24 @@ class OutputLayer final : public ILayer public: /** Construct an output layer. * - * @param[in] accessor Accessor to give output tensor data to. + * @param[in] accessor Accessor to give output tensor data to. + * @param[in] connection_idx (Optional) Input connection index */ - OutputLayer(ITensorAccessorUPtr accessor) - : _accessor(std::move(accessor)) + OutputLayer(ITensorAccessorUPtr accessor, unsigned int connection_idx = 0) + : _accessor(std::move(accessor)), _connection_idx(connection_idx) { } NodeID create_layer(IStream &s) override { NodeParams common_params = { name(), s.hints().target_hint }; - NodeIdxPair input = { s.tail_node(), 0 }; + NodeIdxPair input = { s.tail_node(), _connection_idx }; return GraphBuilder::add_output_node(s.graph(), common_params, input, std::move(_accessor)); } private: ITensorAccessorUPtr _accessor; + unsigned int _connection_idx; }; /** Activation Layer */ @@ -96,10 +98,13 @@ class ActivationLayer final : public ILayer public: /** Construct an activation layer. * - * @param[in] act_info Activation information + * @param[in] act_info Activation information + * @param[in] out_quant_info (Optional) Output quantization info */ - ActivationLayer(ActivationLayerInfo act_info) - : _act_info(act_info) + ActivationLayer(ActivationLayerInfo act_info, + const QuantizationInfo out_quant_info = QuantizationInfo()) + : _act_info(act_info), + _out_quant_info(std::move(out_quant_info)) { } @@ -107,11 +112,12 @@ public: { NodeParams common_params = { name(), s.hints().target_hint }; NodeIdxPair input = { s.tail_node(), 0 }; - return GraphBuilder::add_activation_node(s.graph(), common_params, input, _act_info); + return GraphBuilder::add_activation_node(s.graph(), common_params, input, _act_info, std::move(_out_quant_info)); } private: - ActivationLayerInfo _act_info; + ActivationLayerInfo _act_info; + const QuantizationInfo _out_quant_info; }; /** Batchnormalization Layer */ @@ -225,7 +231,7 @@ public: */ template <typename... Ts> ConcatLayer(SubStream &&sub_stream1, SubStream &&sub_stream2, Ts &&... rest_sub_streams) - : _sub_streams(), _axis(DataLayoutDimension::CHANNEL) + : _sub_streams(), _concat_descriptor(DataLayoutDimension::CHANNEL) { _sub_streams.push_back(arm_compute::support::cpp14::make_unique<SubStream>(std::move(sub_stream1))); _sub_streams.push_back(arm_compute::support::cpp14::make_unique<SubStream>(std::move(sub_stream2))); @@ -238,14 +244,14 @@ public: } /** Construct a concatenation layer * - * @param[in] axis Axis over the concatenation will be performed - * @param[in] sub_stream1 First graph branch - * @param[in] sub_stream2 Second graph branch - * @param[in] rest_sub_streams Rest sub-graph branches + * @param[in] concat_descriptor Concat layer descriptor + * @param[in] sub_stream1 First graph branch + * @param[in] sub_stream2 Second graph branch + * @param[in] rest_sub_streams Rest sub-graph branches */ template <typename... Ts> - ConcatLayer(DataLayoutDimension axis, SubStream &&sub_stream1, SubStream &&sub_stream2, Ts &&... rest_sub_streams) - : _sub_streams(), _axis(axis) + ConcatLayer(descriptors::ConcatLayerDescriptor concat_descriptor, SubStream &&sub_stream1, SubStream &&sub_stream2, Ts &&... rest_sub_streams) + : _sub_streams(), _concat_descriptor(concat_descriptor) { _sub_streams.push_back(arm_compute::support::cpp14::make_unique<SubStream>(std::move(sub_stream1))); _sub_streams.push_back(arm_compute::support::cpp14::make_unique<SubStream>(std::move(sub_stream2))); @@ -262,7 +268,7 @@ public: */ template <typename... Ts> ConcatLayer(SubStream &&sub_stream) - : _sub_streams(), _axis(DataLayoutDimension::CHANNEL) + : _sub_streams(), _concat_descriptor(DataLayoutDimension::CHANNEL) { _sub_streams.push_back(arm_compute::support::cpp14::make_unique<SubStream>(std::move(sub_stream))); } @@ -289,14 +295,14 @@ public: } } } - nid = GraphBuilder::add_concatenate_node(s.graph(), common_params, nodes, _axis); + nid = GraphBuilder::add_concatenate_node(s.graph(), common_params, nodes, _concat_descriptor); } return nid; } private: std::vector<std::unique_ptr<SubStream>> _sub_streams; - DataLayoutDimension _axis; + descriptors::ConcatLayerDescriptor _concat_descriptor; }; /** Convolution Layer */ @@ -414,28 +420,31 @@ class DepthwiseConvolutionLayer final : public ILayer public: /** Construct a depthwise convolution layer. * - * @param[in] conv_width Convolution width. - * @param[in] conv_height Convolution height. - * @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] depth_multiplier (Optional) Depth multiplier parameter. - * @param[in] quant_info (Optional) Quantization info used for weights + * @param[in] conv_width Convolution width. + * @param[in] conv_height Convolution height. + * @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] depth_multiplier (Optional) Depth multiplier parameter. + * @param[in] weights_quant_info (Optional) Quantization info used for weights + * @param[in] out_quant_info (Optional) Output quantization info */ DepthwiseConvolutionLayer(unsigned int conv_width, unsigned int conv_height, ITensorAccessorUPtr weights, ITensorAccessorUPtr bias, PadStrideInfo conv_info, - int depth_multiplier = 1, - const QuantizationInfo quant_info = QuantizationInfo()) + int depth_multiplier = 1, + const QuantizationInfo weights_quant_info = QuantizationInfo(), + const QuantizationInfo out_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)), _depth_multiplier(depth_multiplier), - _quant_info(std::move(quant_info)) + _weights_quant_info(std::move(weights_quant_info)), + _out_quant_info(std::move(out_quant_info)) { } @@ -446,7 +455,7 @@ public: return GraphBuilder::add_depthwise_convolution_node(s.graph(), common_params, input, Size2D(_conv_width, _conv_height), _conv_info, _depth_multiplier, s.hints().depthwise_convolution_method_hint, - std::move(_weights), std::move(_bias), std::move(_quant_info)); + std::move(_weights), std::move(_bias), std::move(_weights_quant_info), std::move(_out_quant_info)); } private: @@ -456,7 +465,8 @@ private: ITensorAccessorUPtr _weights; ITensorAccessorUPtr _bias; int _depth_multiplier; - const QuantizationInfo _quant_info; + const QuantizationInfo _weights_quant_info; + const QuantizationInfo _out_quant_info; }; /** DetectionOutput Layer */ class DetectionOutputLayer final : public ILayer diff --git a/arm_compute/graph/nodes/ActivationLayerNode.h b/arm_compute/graph/nodes/ActivationLayerNode.h index 723120655b..a17b0103e4 100644 --- a/arm_compute/graph/nodes/ActivationLayerNode.h +++ b/arm_compute/graph/nodes/ActivationLayerNode.h @@ -36,9 +36,11 @@ class ActivationLayerNode final : public INode public: /** Constructor * - * @param[in] info Activation Layer information + * @param[in] info Activation Layer information + * @param[in] out_quant_info (Optional) Output quantization info */ - ActivationLayerNode(ActivationLayerInfo info); + ActivationLayerNode(ActivationLayerInfo info, + QuantizationInfo out_quant_info = QuantizationInfo()); /** Activation metadata accessor * * @return The activation info of the layer @@ -56,6 +58,7 @@ public: private: ActivationLayerInfo _info; + QuantizationInfo _out_quant_info; }; } // namespace graph } // namespace arm_compute diff --git a/arm_compute/graph/nodes/ConcatenateLayerNode.h b/arm_compute/graph/nodes/ConcatenateLayerNode.h index 20c8523752..fc122845e8 100644 --- a/arm_compute/graph/nodes/ConcatenateLayerNode.h +++ b/arm_compute/graph/nodes/ConcatenateLayerNode.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -36,10 +36,10 @@ class ConcatenateLayerNode final : public INode public: /** Constructor * - * @param[in] total_nodes Number of nodes that will get concatenated - * @param[in] axis Concatenation axis + * @param[in] total_nodes Number of nodes that will get concatenated + * @param[in] concat_descriptor Concatenate Layer Descriptor */ - ConcatenateLayerNode(unsigned int total_nodes, DataLayoutDimension axis); + ConcatenateLayerNode(unsigned int total_nodes, descriptors::ConcatLayerDescriptor concat_descriptor); /** Computes concatenations output descriptor * * @param[in] input_descriptors Input descriptors @@ -68,6 +68,12 @@ public: */ DataLayoutDimension concatenation_axis() const; + /** Concatenation output quantization info accessor + * + * @return Output quantization info + */ + QuantizationInfo output_quantization_info() const; + // Inherited overridden methods: NodeType type() const override; bool forward_descriptors() override; @@ -75,9 +81,9 @@ public: void accept(INodeVisitor &v) override; private: - unsigned int _total_nodes; - DataLayoutDimension _axis; - bool _is_enabled; + unsigned int _total_nodes; + descriptors::ConcatLayerDescriptor _concat_descriptor; + bool _is_enabled; }; } // namespace graph } // namespace arm_compute diff --git a/arm_compute/graph/nodes/DepthwiseConvolutionLayerNode.h b/arm_compute/graph/nodes/DepthwiseConvolutionLayerNode.h index 8c0aae13c9..fd0273416e 100644 --- a/arm_compute/graph/nodes/DepthwiseConvolutionLayerNode.h +++ b/arm_compute/graph/nodes/DepthwiseConvolutionLayerNode.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -39,10 +39,12 @@ public: * @param[in] info Convolution layer attributes * @param[in] depth_multiplier (Optional) Depth multiplier parameter. * @param[in] method (Optional) Depthwise convolution method to use + * @param[in] out_quant_info (Optional) Output quantization info */ DepthwiseConvolutionLayerNode(PadStrideInfo info, int depth_multiplier = 1, - DepthwiseConvolutionMethod method = DepthwiseConvolutionMethod::Default); + DepthwiseConvolutionMethod method = DepthwiseConvolutionMethod::Default, + QuantizationInfo out_quant_info = QuantizationInfo()); /** Sets the depthwise convolution method to use * * @param[in] method Depthwise convolution method to use @@ -103,6 +105,7 @@ private: PadStrideInfo _info; int _depth_multiplier; DepthwiseConvolutionMethod _method; + QuantizationInfo _out_quant_info; ActivationLayerInfo _fused_activation; }; } // namespace graph 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 diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp index b714c55136..26ea02a9ff 100644 --- a/utils/GraphUtils.cpp +++ b/utils/GraphUtils.cpp @@ -153,24 +153,29 @@ NumPyAccessor::NumPyAccessor(std::string npy_path, TensorShape shape, DataType d } template <typename T> -void NumPyAccessor::access_numpy_tensor(ITensor &tensor) +void NumPyAccessor::access_numpy_tensor(ITensor &tensor, T tolerance) { const int num_elements = tensor.info()->tensor_shape().total_size(); - int num_mismatches = utils::compare_tensor<T>(tensor, _npy_tensor); + int num_mismatches = utils::compare_tensor<T>(tensor, _npy_tensor, tolerance); float percentage_mismatches = static_cast<float>(num_mismatches) / num_elements; _output_stream << "Results: " << 100.f - (percentage_mismatches * 100) << " % matches with the provided output[" << _filename << "]." << std::endl; + _output_stream << " " << num_elements - num_mismatches << " out of " << num_elements << " matches with the provided output[" << _filename << "]." << std::endl + << std::endl; } bool NumPyAccessor::access_tensor(ITensor &tensor) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::F32, DataType::QASYMM8); ARM_COMPUTE_ERROR_ON(_npy_tensor.info()->dimension(0) != tensor.info()->dimension(0)); switch(tensor.info()->data_type()) { + case DataType::QASYMM8: + access_numpy_tensor<qasymm8_t>(tensor, 0); + break; case DataType::F32: - access_numpy_tensor<float>(tensor); + access_numpy_tensor<float>(tensor, 0.0001f); break; default: ARM_COMPUTE_ERROR("NOT SUPPORTED!"); diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h index 131378e5bd..47656766a6 100644 --- a/utils/GraphUtils.h +++ b/utils/GraphUtils.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -160,7 +160,7 @@ public: private: template <typename T> - void access_numpy_tensor(ITensor &tensor); + void access_numpy_tensor(ITensor &tensor, T tolerance); Tensor _npy_tensor; const std::string _filename; diff --git a/utils/Utils.h b/utils/Utils.h index 04ccc3e812..788ae4eeb7 100644 --- a/utils/Utils.h +++ b/utils/Utils.h @@ -782,15 +782,16 @@ void init_sgemm_output(T &dst, T &src0, T &src1, arm_compute::DataType dt) */ uint64_t get_mem_free_from_meminfo(); -/** Compare to tensor +/** Compare two tensors * - * @param[in] tensor1 First tensor to be compared. - * @param[in] tensor2 Second tensor to be compared. + * @param[in] tensor1 First tensor to be compared. + * @param[in] tensor2 Second tensor to be compared. + * @param[in] tolerance Tolerance used for the comparison. * * @return The number of mismatches */ template <typename T> -int compare_tensor(ITensor &tensor1, ITensor &tensor2) +int compare_tensor(ITensor &tensor1, ITensor &tensor2, T tolerance) { ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(&tensor1, &tensor2); ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(&tensor1, &tensor2); @@ -807,7 +808,7 @@ int compare_tensor(ITensor &tensor1, ITensor &tensor2) execute_window_loop(window, [&](const Coordinates & id) { - if(std::abs(*reinterpret_cast<T *>(itensor1.ptr()) - *reinterpret_cast<T *>(itensor2.ptr())) > 0.0001) + if(std::abs(*reinterpret_cast<T *>(itensor1.ptr()) - *reinterpret_cast<T *>(itensor2.ptr())) > tolerance) { ++num_mismatches; } |