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 --- src/graph/backends/CL/CLFunctionsFactory.cpp | 13 +- src/graph/backends/NEON/NEFunctionFactory.cpp | 14 +- src/graph/mutators/NodeFusionMutator.cpp | 208 +++++++++++++++------ src/graph/nodes/ActivationLayerNode.cpp | 6 +- .../FusedConvolutionBatchNormalizationNode.cpp | 152 +++++++++++++++ src/graph/printers/DotGraphPrinter.cpp | 10 +- 6 files changed, 337 insertions(+), 66 deletions(-) create mode 100644 src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp (limited to 'src/graph') diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp index b9e3ddc0a3..7473ff480f 100644 --- a/src/graph/backends/CL/CLFunctionsFactory.cpp +++ b/src/graph/backends/CL/CLFunctionsFactory.cpp @@ -40,7 +40,8 @@ namespace backends /** Target specific information structure used to pass information to the layer templates */ struct CLTargetInfo { - using TensorType = arm_compute::ICLTensor; + using TensorType = arm_compute::ICLTensor; + using TensorConcreteType = CLTensor; static Target TargetType; }; @@ -69,6 +70,14 @@ struct CLEltwiseFunctions using Subtraction = CLArithmeticSubtraction; using Multiplication = CLPixelWiseMultiplication; }; + +/** Function and tensor types to be used inside a CL fused convolution/batch normalization layer */ +struct CLFusedLayerTypes +{ + using ConvolutionLayer = CLConvolutionLayer; + using FuseBatchNormalization = CLFuseBatchNormalization; +}; + // TODO (isagot01): Remove once we support heterogeneous scheduling at function level /** Wrapper for the CPP Function in the OpenCL backend **/ class CPPWrapperFunction : public IFunction @@ -192,6 +201,8 @@ std::unique_ptr CLFunctionFactory::create(INode *node, GraphContext & return detail::create_flatten_layer(*polymorphic_downcast(node)); case NodeType::FullyConnectedLayer: return detail::create_fully_connected_layer(*polymorphic_downcast(node), ctx); + case NodeType::FusedConvolutionBatchNormalizationLayer: + return detail::create_fused_convolution_batch_normalization_layer(*polymorphic_downcast(node)); case NodeType::GenerateProposalsLayer: return detail::create_generate_proposals_layer(*polymorphic_downcast(node), ctx); case NodeType::NormalizationLayer: diff --git a/src/graph/backends/NEON/NEFunctionFactory.cpp b/src/graph/backends/NEON/NEFunctionFactory.cpp index dc987dd86e..f23845c314 100644 --- a/src/graph/backends/NEON/NEFunctionFactory.cpp +++ b/src/graph/backends/NEON/NEFunctionFactory.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,8 @@ namespace backends /** Target specific information structure used to pass information to the layer templates */ struct NETargetInfo { - using TensorType = arm_compute::ITensor; + using TensorType = arm_compute::ITensor; + using TensorConcreteType = arm_compute::Tensor; static Target TargetType; }; @@ -76,6 +77,13 @@ struct NEEltwiseFunctions using Multiplication = NEPixelWiseMultiplication; }; +/** Function and tensor types to be used inside a NEON fused convolution/batch normalization layer */ +struct NEFusedLayerTypes +{ + using ConvolutionLayer = NEConvolutionLayer; + using FuseBatchNormalization = NEFuseBatchNormalization; +}; + namespace detail { // Specialized functions @@ -210,6 +218,8 @@ std::unique_ptr NEFunctionFactory::create(INode *node, GraphContext & return detail::create_flatten_layer(*polymorphic_downcast(node)); case NodeType::FullyConnectedLayer: return detail::create_fully_connected_layer(*polymorphic_downcast(node), ctx); + case NodeType::FusedConvolutionBatchNormalizationLayer: + return detail::create_fused_convolution_batch_normalization_layer(*polymorphic_downcast(node)); case NodeType::NormalizationLayer: return detail::create_normalization_layer(*polymorphic_downcast(node), ctx); case NodeType::PermuteLayer: diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index 9dc02d1ad1..445748caf7 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,9 +23,11 @@ */ #include "arm_compute/graph/mutators/NodeFusionMutator.h" -#include "arm_compute/graph/Graph.h" +#include "arm_compute/graph/GraphBuilder.h" #include "arm_compute/graph/Logger.h" #include "arm_compute/graph/Utils.h" +#include "arm_compute/graph/backends/BackendRegistry.h" +#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h" #include "arm_compute/graph/nodes/Nodes.h" #include "arm_compute/core/utils/misc/Cast.h" @@ -38,69 +40,156 @@ namespace graph { namespace detail { +void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge) +{ + ARM_COMPUTE_ERROR_ON(output_edge == nullptr); + + auto *conv_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->producer()); + auto *bn_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->consumer()); + + // Not fusing if number of groups is greater than 1 + if(conv_node->num_groups() > 1) + { + return; + } + + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing convolution node with ID : " << output_edge->producer_id() + << " with BatchNormalization Layer node with ID : " << output_edge->consumer_id() << std::endl); + + // Prevent fusion if fused node has an output accessor + if(conv_node->output(0)->accessor() == nullptr) + { + const Target assigned_target = conv_node->assigned_target(); + + // Extract conv inputs + const auto conv_input_id = conv_node->input_edge(0)->producer_id(); + const auto conv_weights_id = conv_node->input_edge(1)->producer_id(); + const auto out_quant_info = conv_node->output(0)->desc().quant_info; + const auto conv_info = conv_node->convolution_info(); + const auto conv_method = conv_node->convolution_method(); + const auto num_groups = conv_node->num_groups(); + const auto act_info = bn_node->fused_activation(); + FastMathHint fast_math_hint = conv_node->fast_math_hint(); + + // Extract bn inputs + const auto bn_mean_id = bn_node->input_edge(1)->producer_id(); + const auto bn_var_id = bn_node->input_edge(2)->producer_id(); + const auto bn_beta_id = bn_node->input_edge(3)->producer_id(); + const auto bn_gamma_id = bn_node->input_edge(4)->producer_id(); + const auto epsilon = bn_node->epsilon(); + + // Create the fused node + const NodeID fused_id = g.add_node(epsilon, conv_info, num_groups, conv_method, fast_math_hint, out_quant_info, act_info); + + if(conv_node->input_edge(2) != nullptr) + { + auto conv_bias_id = conv_node->input_edge(2)->producer_id(); + g.add_connection(conv_bias_id, 0, fused_id, 2); + } + + // Add connections from the conv/batch_norm inputs to the fused node + g.add_connection(conv_input_id, 0, fused_id, 0); + g.add_connection(conv_weights_id, 0, fused_id, 1); + g.add_connection(bn_mean_id, 0, fused_id, 3); + g.add_connection(bn_var_id, 0, fused_id, 4); + g.add_connection(bn_beta_id, 0, fused_id, 5); + g.add_connection(bn_gamma_id, 0, fused_id, 6); + + auto fused_node = g.node(fused_id); + std::vector bn_driving_nodes = get_driving_nodes(*bn_node); + + // Extract batch normalization node accessor if any + auto bn_node_accessor = bn_node->output(0)->extract_accessor(); + auto bn_node_name = bn_node->name(); + + // Remove batch normalization node + g.remove_node(bn_node->id()); + + // Get driving nodes of batch normalization node + for(auto &driving_node : bn_driving_nodes) + { + g.add_connection(fused_id, 0, driving_node.node_id, driving_node.index); + configure_tensor(fused_node->output(0)); + } + // Update fused node outputs + fused_node->output(0)->set_accessor(std::move(bn_node_accessor)); + fused_node->set_assigned_target(assigned_target); + fused_node->set_common_node_parameters(NodeParams{ conv_node->name() + "+" + bn_node_name, assigned_target }); + + // Remove convolution node + g.remove_node(conv_node->id()); + } + else + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution with batch normalization due to the presence of an output accessor\n"); + } +} + template -void fuse_node_with_activation(Graph &g, - const std::set &supported_fused_activations, - std::function const &prec) +void fuse_node_with_activation(Graph &g, const Edge *output_edge, const std::set &supported_fused_activations) +{ + ARM_COMPUTE_ERROR_ON(output_edge == nullptr); + + auto *n_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->producer()); + auto *act_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->consumer()); + + ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr || n_node->output(0) == nullptr); + + // Check if activation is supported for fusion + if(supported_fused_activations.count(act_node->activation_info().activation()) == 0) + { + return; + } + + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing node with ID : " << output_edge->producer_id() + << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl); + + // Prevent fusion if fused node has an output accessor + if(n_node->output(0)->accessor() == nullptr) + { + // Get driving nodes of activation node + std::vector act_driving_nodes = get_driving_nodes(*act_node); + + // Set activation info to fused node + n_node->set_fused_activation(act_node->activation_info()); + + // Extract activation node accessor if any + auto act_node_accessor = act_node->output(0)->extract_accessor(); + + // Remove activation node + g.remove_node(act_node->id()); + + // Update fused node outputs + for(auto &driving_node : act_driving_nodes) + { + g.add_connection(n_node->id(), 0, driving_node.node_id, driving_node.index); + } + + // Update accessor to fused node + n_node->output(0)->set_accessor(std::move(act_node_accessor)); + } + else + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of node with activation due to the presence of an output accessor\n"); + } +} + +template +void fuse_layer(Graph &g, std::function const &prec, const F fuse_fcn, Args &&... optional_arguments) { // Not interested in the order of nodes for(auto &node : g.nodes()) { // Check if the node is of type N and not a branching node - if(node && node->type() == N::node_type && node->output_edges().size() == 1) + if(node && node->type() == N1::node_type && node->output_edges().size() == 1) { - auto output_edge_id = *node->output_edges().begin(); - auto output_edge = g.edge(output_edge_id); + const auto output_edge_id = *node->output_edges().begin(); + const auto output_edge = g.edge(output_edge_id); + // Check if following node is an activation layer node - if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == NodeType::ActivationLayer)) + if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == N2::node_type) && prec(*output_edge->producer())) { - auto *n_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->producer()); - auto *act_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->consumer()); - - ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr || n_node->output(0) == nullptr); - - // Check given precondition - if(!prec(*n_node)) - { - continue; - } - // Check if activation is supported for fusion - if(supported_fused_activations.count(act_node->activation_info().activation()) == 0) - { - continue; - } - - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing node with ID : " << output_edge->producer_id() - << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl); - - // Prevent fusion if fused node has an output accessor - if(n_node->output(0)->accessor() == nullptr) - { - // Get driving nodes of activation node - std::vector act_driving_nodes = get_driving_nodes(*act_node); - - // Set activation info to fused node - n_node->set_fused_activation(act_node->activation_info()); - - // Extract activation node accessor if any - auto act_node_accessor = act_node->output(0)->extract_accessor(); - - // Remove activation node - g.remove_node(act_node->id()); - - // Update fused node outputs - for(auto &driving_node : act_driving_nodes) - { - g.add_connection(n_node->id(), 0, driving_node.node_id, driving_node.index); - } - - // Update accessor to fused node - n_node->output(0)->set_accessor(std::move(act_node_accessor)); - } - else - { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of node with activation due to the presence of an output accessor\n"); - } + fuse_fcn(g, output_edge, optional_arguments...); } } } @@ -129,9 +218,10 @@ void NodeFusionMutator::mutate(Graph &g) }; // Fusion mutations - detail::fuse_node_with_activation(g, supported_fused_activations, empty_prec); - detail::fuse_node_with_activation(g, supported_fused_activations, empty_prec); - detail::fuse_node_with_activation(g, supported_fused_activations, qs8_prec); + detail::fuse_layer(g, empty_prec, detail::fuse_node_with_activation, supported_fused_activations); + detail::fuse_layer(g, empty_prec, detail::fuse_node_with_activation, supported_fused_activations); + detail::fuse_layer(g, qs8_prec, detail::fuse_node_with_activation, supported_fused_activations); + detail::fuse_layer(g, empty_prec, detail::fuse_convolution_with_batch_normalization); } } // namespace graph } // namespace arm_compute diff --git a/src/graph/nodes/ActivationLayerNode.cpp b/src/graph/nodes/ActivationLayerNode.cpp index 414684cf30..85cb10bbdb 100644 --- a/src/graph/nodes/ActivationLayerNode.cpp +++ b/src/graph/nodes/ActivationLayerNode.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -67,7 +67,7 @@ TensorDescriptor ActivationLayerNode::configure_output(size_t idx) const NodeType ActivationLayerNode::type() const { - return NodeType::ActivationLayer; + return ActivationLayerNode::node_type; } void ActivationLayerNode::accept(INodeVisitor &v) @@ -75,4 +75,4 @@ void ActivationLayerNode::accept(INodeVisitor &v) v.visit(*this); } } // namespace graph -} // namespace arm_compute \ No newline at end of file +} // namespace arm_compute diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp new file mode 100644 index 0000000000..27a348fa69 --- /dev/null +++ b/src/graph/nodes/FusedConvolutionBatchNormalizationNode.cpp @@ -0,0 +1,152 @@ +/* + * 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. + */ +#include "arm_compute/graph/nodes/FusedConvolutionBatchNormalizationNode.h" + +#include "arm_compute/core/Utils.h" +#include "arm_compute/graph/Graph.h" +#include "arm_compute/graph/INodeVisitor.h" +#include "arm_compute/graph/Utils.h" + +namespace arm_compute +{ +namespace graph +{ +FusedConvolutionBatchNormalizationNode::FusedConvolutionBatchNormalizationNode(float epsilon, PadStrideInfo info, + unsigned int num_groups, + ConvolutionMethod method, + FastMathHint fast_math_hint, + QuantizationInfo out_quant_info, ActivationLayerInfo fused_activation) + : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(out_quant_info), _fused_activation(fused_activation) +{ + _input_edges.resize(7, EmptyEdgeID); + _outputs.resize(1, NullTensorID); +} + +void FusedConvolutionBatchNormalizationNode::set_convolution_method(ConvolutionMethod method) +{ + _method = method; +} + +float FusedConvolutionBatchNormalizationNode::epsilon() const +{ + return _epsilon; +} + +ConvolutionMethod FusedConvolutionBatchNormalizationNode::convolution_method() const +{ + return _method; +} + +void FusedConvolutionBatchNormalizationNode::set_fast_math_hint(FastMathHint hint) +{ + _fast_math_hint = hint; +} + +FastMathHint FusedConvolutionBatchNormalizationNode::fast_math_hint() const +{ + return _fast_math_hint; +} + +PadStrideInfo FusedConvolutionBatchNormalizationNode::convolution_info() const +{ + return _info; +} + +unsigned int FusedConvolutionBatchNormalizationNode::num_groups() const +{ + return _num_groups; +} + +ActivationLayerInfo FusedConvolutionBatchNormalizationNode::fused_activation() const +{ + return _fused_activation; +} + +void FusedConvolutionBatchNormalizationNode::set_fused_activation(ActivationLayerInfo fused_activation) +{ + _fused_activation = fused_activation; +} + +TensorDescriptor FusedConvolutionBatchNormalizationNode::compute_output_descriptor(const TensorDescriptor &input_descriptor, + const TensorDescriptor &weights_descriptor, + const PadStrideInfo &info) +{ + unsigned int output_width = 0; + unsigned int output_height = 0; + + const unsigned int input_width = get_dimension_size(input_descriptor, DataLayoutDimension::WIDTH); + const unsigned int input_height = get_dimension_size(input_descriptor, DataLayoutDimension::HEIGHT); + const unsigned int kernel_width = get_dimension_size(weights_descriptor, DataLayoutDimension::WIDTH); + const unsigned int kernel_height = get_dimension_size(weights_descriptor, DataLayoutDimension::HEIGHT); + + std::tie(output_width, output_height) = scaled_dimensions(input_width, input_height, kernel_width, kernel_height, info); + + TensorDescriptor output_descriptor = input_descriptor; + output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::WIDTH), output_width); + output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::HEIGHT), output_height); + output_descriptor.shape.set(get_dimension_idx(output_descriptor, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]); + + return output_descriptor; +} + +bool FusedConvolutionBatchNormalizationNode::forward_descriptors() +{ + if((input_id(0) != NullTensorID) && (input_id(1) != NullTensorID) && (output_id(0) != NullTensorID)) + { + Tensor *dst = output(0); + ARM_COMPUTE_ERROR_ON(dst == nullptr); + dst->desc() = configure_output(0); + return true; + } + return false; +} + +TensorDescriptor FusedConvolutionBatchNormalizationNode::configure_output(size_t idx) const +{ + ARM_COMPUTE_UNUSED(idx); + const Tensor *src = input(0); + const Tensor *weights = input(1); + + ARM_COMPUTE_ERROR_ON(src == nullptr || weights == nullptr); + + TensorDescriptor output_info = compute_output_descriptor(src->desc(), weights->desc(), _info); + if(!_out_quant_info.empty()) + { + output_info.quant_info = _out_quant_info; + } + + return output_info; +} + +NodeType FusedConvolutionBatchNormalizationNode::type() const +{ + return FusedConvolutionBatchNormalizationNode::node_type; +} + +void FusedConvolutionBatchNormalizationNode::accept(INodeVisitor &v) +{ + v.visit(*this); +} +} // namespace graph +} // namespace arm_compute diff --git a/src/graph/printers/DotGraphPrinter.cpp b/src/graph/printers/DotGraphPrinter.cpp index ef156ea252..c939de1b64 100644 --- a/src/graph/printers/DotGraphPrinter.cpp +++ b/src/graph/printers/DotGraphPrinter.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -77,6 +77,14 @@ void DotGraphVisitor::visit(EltwiseLayerNode &n) _info = ss.str(); } +void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n) +{ + ARM_COMPUTE_UNUSED(n); + std::stringstream ss; + ss << "FusedConvolutionBatchNormalizationNode"; + _info = ss.str(); +} + void DotGraphVisitor::visit(NormalizationLayerNode &n) { std::stringstream ss; -- cgit v1.2.1