From acce504ec4aebe5e5da470c1cfc3cee401ff11f3 Mon Sep 17 00:00:00 2001 From: giuros01 Date: Thu, 21 Feb 2019 17:32:34 +0000 Subject: COMPMID-1740: Fuse batch normalization with Convolution Layer at graph level Change-Id: I77ca51c2c72783cc26a099a6a9c3210cdbbe822d Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/797 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Reviewed-by: Georgios Pinitas --- arm_compute/graph/nodes/ActivationLayerNode.h | 5 +- .../nodes/FusedConvolutionBatchNormalizationNode.h | 144 +++++++++++++++++++++ arm_compute/graph/nodes/Nodes.h | 1 + arm_compute/graph/nodes/NodesFwd.h | 1 + 4 files changed, 150 insertions(+), 1 deletion(-) create mode 100644 arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h (limited to 'arm_compute/graph/nodes') diff --git a/arm_compute/graph/nodes/ActivationLayerNode.h b/arm_compute/graph/nodes/ActivationLayerNode.h index 570351bb94..723120655b 100644 --- a/arm_compute/graph/nodes/ActivationLayerNode.h +++ b/arm_compute/graph/nodes/ActivationLayerNode.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,6 +51,9 @@ public: TensorDescriptor configure_output(size_t idx) const override; void accept(INodeVisitor &v) override; +public: + static constexpr NodeType node_type = NodeType::ActivationLayer; + private: ActivationLayerInfo _info; }; diff --git a/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h new file mode 100644 index 0000000000..9b0f5b7ade --- /dev/null +++ b/arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h @@ -0,0 +1,144 @@ +/* + * 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_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__ +#define __ARM_COMPUTE_GRAPH_FUSED_CONVOLUTION_BATCH_NORMALIZATION_NODE_H__ + +#include "arm_compute/graph/INode.h" + +namespace arm_compute +{ +namespace graph +{ +/** Batch Normalization node */ +class FusedConvolutionBatchNormalizationNode final : public INode +{ +public: + /** Constructor + * + * @param[in] epsilon Epsilon parameter. + * @param[in] info Convolution layer attributes. + * @param[in] num_groups (Optional) Number of groups (Defaults to 1) + * @param[in] method (Optional) Convolution method to use + * @param[in] fast_math_hint (Optional) Fast math hint + * @param[in] out_quant_info (Optional) Output quantization info + * @param[in] fused_activation (Optional) Fused activation layer. Disabled if not specified + */ + FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info, + unsigned int num_groups = 1, + ConvolutionMethod method = ConvolutionMethod::Default, + FastMathHint fast_math_hint = FastMathHint::Disabled, + QuantizationInfo out_quant_info = QuantizationInfo(), ActivationLayerInfo fused_activation = ActivationLayerInfo()); + + /** Epsilon parameter accessor + * + * @return Epsilon parameter + */ + float epsilon() const; + + /** Returns fused activation + * + * @return Fused activation + */ + ActivationLayerInfo fused_activation() const; + + /** Sets fused activation + * + * @param[in] fused_activation Fused activation to set + */ + void set_fused_activation(ActivationLayerInfo fused_activation); + + /** Computes convolution output descriptor + * + * @param[in] input_descriptor Input descriptor + * @param[in] weights_descriptor Weights descriptor + * @param[in] info Convolution operation attributes + * + * @return Output descriptor + */ + static TensorDescriptor compute_output_descriptor(const TensorDescriptor &input_descriptor, + const TensorDescriptor &weights_descriptor, + const PadStrideInfo &info); + + /** Sets the convolution layer method to use + * + * @param[in] method Method to use for convolution + */ + void set_convolution_method(ConvolutionMethod method); + + /** Number of groups in convolution accessor + * + * @return Number of groups in convolution + */ + unsigned int num_groups() const; + + /** Convolution layer method accessor + * + * @note This is an indication on which convolution layer implementation to use, + * if it fails to be created the library's heuristic approach will be used + * + * @return Convolution layer method to be used by the node + */ + ConvolutionMethod convolution_method() const; + + /** Sets the fast math fast hint + * + * @param[in] hint Hint to use for convolution + */ + void set_fast_math_hint(FastMathHint hint); + + /** Fast math hint accessor + * + * @return Fast math hint to be used by the node + */ + FastMathHint fast_math_hint() const; + + /** Convolution metadata accessor + * + * @return Convolution information + */ + PadStrideInfo convolution_info() const; + + // Inherited overridden methods: + NodeType type() const override; + bool forward_descriptors() override; + TensorDescriptor configure_output(size_t idx) const override; + void accept(INodeVisitor &v) override; + +public: + static constexpr NodeType node_type = NodeType::FusedConvolutionBatchNormalizationLayer; + +private: + float _epsilon; + + PadStrideInfo _info; + unsigned int _num_groups; + ConvolutionMethod _method; + FastMathHint _fast_math_hint; + QuantizationInfo _out_quant_info; + ActivationLayerInfo _fused_activation; +}; + +} // namespace graph +} // namespace arm_compute +#endif /* __ARM_COMPUTE_GRAPH_BATCH_NORMALIZATION_LAYER_NODE_H__ */ diff --git a/arm_compute/graph/nodes/Nodes.h b/arm_compute/graph/nodes/Nodes.h index 24064855e8..e23b2b9897 100644 --- a/arm_compute/graph/nodes/Nodes.h +++ b/arm_compute/graph/nodes/Nodes.h @@ -38,6 +38,7 @@ #include "arm_compute/graph/nodes/EltwiseLayerNode.h" #include "arm_compute/graph/nodes/FlattenLayerNode.h" #include "arm_compute/graph/nodes/FullyConnectedLayerNode.h" +#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h" #include "arm_compute/graph/nodes/GenerateProposalsLayerNode.h" #include "arm_compute/graph/nodes/InputNode.h" #include "arm_compute/graph/nodes/NormalizationLayerNode.h" diff --git a/arm_compute/graph/nodes/NodesFwd.h b/arm_compute/graph/nodes/NodesFwd.h index cbda3092fd..80576d4608 100644 --- a/arm_compute/graph/nodes/NodesFwd.h +++ b/arm_compute/graph/nodes/NodesFwd.h @@ -44,6 +44,7 @@ class DummyNode; class EltwiseLayerNode; class FlattenLayerNode; class FullyConnectedLayerNode; +class FusedConvolutionBatchNormalizationNode; class GenerateProposalsLayerNode; class InputNode; class NormalizationLayerNode; -- cgit v1.2.1