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/INodeVisitor.h | 11 +- arm_compute/graph/TypePrinter.h | 3 + arm_compute/graph/Types.h | 1 + arm_compute/graph/backends/FunctionHelpers.h | 67 +++++++++- .../FusedConvolutionBatchNormalizationFunction.h | 133 +++++++++++++++++++ arm_compute/graph/mutators/NodeFusionMutator.h | 12 +- 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 + arm_compute/graph/printers/DotGraphPrinter.h | 3 +- 11 files changed, 363 insertions(+), 18 deletions(-) create mode 100644 arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h create mode 100644 arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h (limited to 'arm_compute/graph') diff --git a/arm_compute/graph/INodeVisitor.h b/arm_compute/graph/INodeVisitor.h index 573d642892..842ca4bfb3 100644 --- a/arm_compute/graph/INodeVisitor.h +++ b/arm_compute/graph/INodeVisitor.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -91,6 +91,11 @@ public: * @param[in] n Node to visit. */ virtual void visit(FullyConnectedLayerNode &n) = 0; + /** Visit FusedConvolutionBatchNormalizationNode. + * + * @param[in] n Node to visit. + */ + virtual void visit(FusedConvolutionBatchNormalizationNode &n) = 0; /** Visit InputNode. * * @param[in] n Node to visit. @@ -195,6 +200,10 @@ public: { default_visit(); } + virtual void visit(FusedConvolutionBatchNormalizationNode &n) override + { + default_visit(); + } virtual void visit(InputNode &n) override { default_visit(); diff --git a/arm_compute/graph/TypePrinter.h b/arm_compute/graph/TypePrinter.h index ca62d4ec17..b1cfbcf658 100644 --- a/arm_compute/graph/TypePrinter.h +++ b/arm_compute/graph/TypePrinter.h @@ -98,6 +98,9 @@ inline ::std::ostream &operator<<(::std::ostream &os, const NodeType &node_type) case NodeType::FullyConnectedLayer: os << "FullyConnectedLayer"; break; + case NodeType::FusedConvolutionBatchNormalizationLayer: + os << "FusedConvolutionBatchNormalizationLayer"; + break; case NodeType::GenerateProposalsLayer: os << "GenerateProposalsLayer"; break; diff --git a/arm_compute/graph/Types.h b/arm_compute/graph/Types.h index 8377253338..2905dfcbf6 100644 --- a/arm_compute/graph/Types.h +++ b/arm_compute/graph/Types.h @@ -138,6 +138,7 @@ enum class NodeType EltwiseLayer, FlattenLayer, FullyConnectedLayer, + FusedConvolutionBatchNormalizationLayer, GenerateProposalsLayer, NormalizationLayer, NormalizePlanarYUVLayer, diff --git a/arm_compute/graph/backends/FunctionHelpers.h b/arm_compute/graph/backends/FunctionHelpers.h index 7242bc6ede..d0035d9a84 100644 --- a/arm_compute/graph/backends/FunctionHelpers.h +++ b/arm_compute/graph/backends/FunctionHelpers.h @@ -28,6 +28,7 @@ #include "arm_compute/graph/Tensor.h" #include "arm_compute/graph/TypePrinter.h" #include "arm_compute/graph/Types.h" +#include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h" #include "arm_compute/graph/backends/Utils.h" #include "arm_compute/graph/nodes/Nodes.h" @@ -135,11 +136,12 @@ std::unique_ptr create_batch_normalization_layer(BatchNormalizationLa validate_node(node, 5 /* expected inputs */, 1 /* expected outputs */); // Extract IO and info - typename TargetInfo::TensorType *input = get_backing_tensor(node.input(0)); - typename TargetInfo::TensorType *mean = get_backing_tensor(node.input(1)); - typename TargetInfo::TensorType *var = get_backing_tensor(node.input(2)); - typename TargetInfo::TensorType *beta = get_backing_tensor(node.input(3)); - typename TargetInfo::TensorType *gamma = get_backing_tensor(node.input(4)); + typename TargetInfo::TensorType *input = get_backing_tensor(node.input(0)); + typename TargetInfo::TensorType *mean = get_backing_tensor(node.input(1)); + typename TargetInfo::TensorType *var = get_backing_tensor(node.input(2)); + typename TargetInfo::TensorType *beta = get_backing_tensor(node.input(3)); + typename TargetInfo::TensorType *gamma = get_backing_tensor(node.input(4)); + typename TargetInfo::TensorType *output = get_backing_tensor(node.output(0)); const float epsilon = node.epsilon(); const ActivationLayerInfo fused_act = node.fused_activation(); @@ -163,6 +165,61 @@ std::unique_ptr create_batch_normalization_layer(BatchNormalizationLa return std::move(func); } +/** Create a backend batch normalization layer function + * + * @tparam BatchNormalizationLayerFunction Backend batch normalization function + * @tparam TargetInfo Target-specific information + * + * @param[in] node Node to create the backend function for + * + * @return Backend batch normalization layer function + */ +template +std::unique_ptr create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node) +{ + validate_node(node, 7 /* expected inputs */, 1 /* expected outputs */); + + // Extract IO and info + typename TargetInfo::TensorType *input = get_backing_tensor(node.input(0)); + typename TargetInfo::TensorType *weights = get_backing_tensor(node.input(1)); + typename TargetInfo::TensorType *biases = get_backing_tensor(node.input(2)); + typename TargetInfo::TensorType *mean = get_backing_tensor(node.input(3)); + typename TargetInfo::TensorType *var = get_backing_tensor(node.input(4)); + typename TargetInfo::TensorType *beta = get_backing_tensor(node.input(5)); + typename TargetInfo::TensorType *gamma = get_backing_tensor(node.input(6)); + + typename TargetInfo::TensorType *output = get_backing_tensor(node.output(0)); + + const PadStrideInfo conv_info = node.convolution_info(); + const unsigned int num_groups = node.num_groups(); + const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled; + const ActivationLayerInfo fused_act = node.fused_activation(); + const float epsilon = node.epsilon(); + + const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); + if(is_quantized && biases != nullptr) + { + biases->info()->set_data_type(DataType::S32); + } + + // Create and configure function + auto func = support::cpp14::make_unique>(); + func->configure(input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act); + + // Log info + ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " + << node.name() + << " Type: " << node.name() + << " Target: " << TargetInfo::TargetType + << " Data Type: " << input->info()->data_type() + << " Input shape: " << input->info()->tensor_shape() + << " Weights shape: " << weights->info()->tensor_shape() + << " Output shape: " << output->info()->tensor_shape() + << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "") + << std::endl); + return std::move(func); +} + /** Create a backend bounding box transform layer function * * @tparam BoundingBoxTransformLayerFunction Backend bounding box transform function diff --git a/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h new file mode 100644 index 0000000000..92af17b227 --- /dev/null +++ b/arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h @@ -0,0 +1,133 @@ +/* + * 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_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H__ +#define __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H__ + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IFunction.h" + +namespace arm_compute +{ +namespace graph +{ +namespace backends +{ +/** Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE, CL}ConvolutionLayer with the modified weights */ +template +class FusedConvolutionBatchNormalizationFunction : public IFunction +{ +public: + using TensorType = typename TargetInfo::TensorType; + using TensorConcreteType = typename TargetInfo::TensorConcreteType; + + FusedConvolutionBatchNormalizationFunction() + : _conv_layer(), _fused_batch_norm_layer(), _fused_bias(), _is_prepared(false) + { + } + + /** Set the input and output tensors. + * + * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QASYMM8/F16/F32. + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. + * @param[in] bias Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type. + * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input + * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input + * @param[in] epsilon Small value to avoid division with zero. Default value is 0.001f. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] num_groups Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout + * @param[in] fast_math Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation + * available which may introduce a drop of accuracy as well. Default is false + * @param[in] fused_act Activation layer information in case of a fused activation. + * + */ + void configure(TensorType *input, + TensorType *weights, + TensorType *bias, + TensorType *output, + const TensorType *mean, + const TensorType *var, + const TensorType *beta, + const TensorType *gamma, + float epsilon, const PadStrideInfo &conv_info, unsigned int num_groups, bool fast_math, ActivationLayerInfo const &fused_act) + { + // We don't run any validate, as we assume that the layers have been already validated + const bool has_bias = (bias != nullptr); + const TensorType *bias_to_use; + + // We check if the layer has a bias. If yes, use it in-place. If not, we need to create one + // as batch normalization might end up with a bias != 0 + if(has_bias) + { + _fused_batch_norm_layer.configure(weights, mean, var, nullptr, nullptr, bias, beta, gamma, epsilon); + bias_to_use = bias; + } + else + { + _fused_batch_norm_layer.configure(weights, mean, var, nullptr, &_fused_bias, nullptr, beta, gamma, epsilon); + bias_to_use = &_fused_bias; + } + + _conv_layer.configure(input, weights, bias_to_use, output, conv_info, WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups); + + if(!has_bias) + { + _fused_bias.allocator()->allocate(); + } + } + + // Inherited methods overridden: + void run() + { + prepare(); + _conv_layer.run(); + } + + void prepare() + { + if(!_is_prepared) + { + _fused_batch_norm_layer.run(); + _is_prepared = true; + } + } + +private: + typename FusedLayerTypes::ConvolutionLayer _conv_layer; + typename FusedLayerTypes::FuseBatchNormalization _fused_batch_norm_layer; + TensorConcreteType _fused_bias; + bool _is_prepared; +}; +} // namespace backends +} // namespace graph +} // namespace arm_compute + +#endif /* __ARM_COMPUTE_GRAPH_BACKENDS_FUSED_CONVOLUTION_BATCH_NORMAZLIZATION_FUNCTION_H__ */ diff --git a/arm_compute/graph/mutators/NodeFusionMutator.h b/arm_compute/graph/mutators/NodeFusionMutator.h index 8f16c65dfa..b9ca464822 100644 --- a/arm_compute/graph/mutators/NodeFusionMutator.h +++ b/arm_compute/graph/mutators/NodeFusionMutator.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,21 +24,13 @@ #ifndef __ARM_COMPUTE_GRAPH_NODE_FUSION_MUTATOR_H__ #define __ARM_COMPUTE_GRAPH_NODE_FUSION_MUTATOR_H__ +#include "arm_compute/graph/Graph.h" #include "arm_compute/graph/IGraphMutator.h" namespace arm_compute { namespace graph { -namespace detail -{ -/** Fused batch normalization with activation - * - * @param[in] g Graph to perform operation fusion on - */ -void fuse_batch_norm_with_activation(Graph &g); -} // namespace detail - /** Mutation pass to fuss nodes */ class NodeFusionMutator final : public IGraphMutator { 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; diff --git a/arm_compute/graph/printers/DotGraphPrinter.h b/arm_compute/graph/printers/DotGraphPrinter.h index d4cf6928e5..9d2ea46fde 100644 --- a/arm_compute/graph/printers/DotGraphPrinter.h +++ b/arm_compute/graph/printers/DotGraphPrinter.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -56,6 +56,7 @@ public: void visit(ConvolutionLayerNode &n) override; void visit(DepthwiseConvolutionLayerNode &n) override; void visit(EltwiseLayerNode &n) override; + void visit(FusedConvolutionBatchNormalizationNode &n) override; void visit(NormalizationLayerNode &n) override; void visit(PoolingLayerNode &n) override; void default_visit() override; -- cgit v1.2.1