From fb2280381e7a98ad698ea0c1b2cd635a48ad4acc Mon Sep 17 00:00:00 2001 From: Sheri Zhang Date: Tue, 2 Nov 2021 10:45:07 +0000 Subject: Add graph level convolution fusion with post operator Resolves: COMPMID-4701 Signed-off-by: Sheri Zhang Change-Id: I8a0d3c2ed4bf84489d94b8ae6641d6041aadaee5 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6557 Tested-by: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: SiCong Li Comments-Addressed: Arm Jenkins --- src/graph/INodeVisitor.cpp | 4 + src/graph/backends/CL/CLFunctionsFactory.cpp | 3 + src/graph/backends/CL/CLNodeValidator.cpp | 2 + src/graph/mutators/NodeFusionMutator.cpp | 274 ++++++++++++++++++++- src/graph/nodes/FusedConvolutionWithPostOpNode.cpp | 164 ++++++++++++ src/graph/printers/DotGraphPrinter.cpp | 8 + 6 files changed, 450 insertions(+), 5 deletions(-) create mode 100644 src/graph/nodes/FusedConvolutionWithPostOpNode.cpp (limited to 'src/graph') diff --git a/src/graph/INodeVisitor.cpp b/src/graph/INodeVisitor.cpp index 2bbf519cf5..ceaaeae7f2 100644 --- a/src/graph/INodeVisitor.cpp +++ b/src/graph/INodeVisitor.cpp @@ -85,6 +85,10 @@ void DefaultNodeVisitor::visit(FusedConvolutionBatchNormalizationNode &n) { default_visit(n); } +void DefaultNodeVisitor::visit(FusedConvolutionWithPostOpNode &n) +{ + default_visit(n); +} void DefaultNodeVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) { default_visit(n); diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp index 1cd3f9f9c7..838977e75c 100644 --- a/src/graph/backends/CL/CLFunctionsFactory.cpp +++ b/src/graph/backends/CL/CLFunctionsFactory.cpp @@ -81,6 +81,7 @@ struct CLFusedLayerTypes using ConvolutionLayer = CLConvolutionLayer; using DepthwiseConvolutionLayer = CLDepthwiseConvolutionLayer; using FuseBatchNormalization = CLFuseBatchNormalization; + using GEMMConvolutionLayer = CLGEMMConvolutionLayer; }; /** Wrapper for the CPP Function in the OpenCL backend **/ @@ -273,6 +274,8 @@ std::unique_ptr CLFunctionFactory::create(INode *node, GraphContext & return detail::create_fully_connected_layer(*polymorphic_downcast(node), ctx); case NodeType::FusedConvolutionBatchNormalizationLayer: return detail::create_fused_convolution_batch_normalization_layer(*polymorphic_downcast(node), ctx); + case NodeType::FusedConvolutionWithPostOp: + return detail::create_fused_convolution_with_post_op(*polymorphic_downcast(node), ctx); case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer: return detail::create_fused_depthwise_convolution_batch_normalization_layer(*polymorphic_downcast(node), ctx); case NodeType::GenerateProposalsLayer: diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp index 8e3b4c8705..c50782db48 100644 --- a/src/graph/backends/CL/CLNodeValidator.cpp +++ b/src/graph/backends/CL/CLNodeValidator.cpp @@ -76,6 +76,8 @@ Status CLNodeValidator::validate(INode *node) CLDirectConvolutionLayer, CLGEMMConvolutionLayer, CLWinogradConvolutionLayer>(*polymorphic_downcast(node)); + case NodeType::FusedConvolutionWithPostOp: + return detail::validate_fused_convolution_with_post_op(*polymorphic_downcast(node)); case NodeType::DepthToSpaceLayer: return detail::validate_depth_to_space_layer(*polymorphic_downcast(node)); case NodeType::DepthwiseConvolutionLayer: diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index e37164c60c..463d1cd8f6 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -28,12 +28,14 @@ #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/FusedConvolutionWithPostOpNode.h" #include "arm_compute/graph/nodes/Nodes.h" #include "src/graph/mutators/MutatorUtils.h" #include "support/Cast.h" +#include #include namespace arm_compute @@ -322,13 +324,13 @@ void fuse_layer(Graph &g, std::function const &prec, const F fuse for(unsigned int i = 0; i < g.nodes().size(); ++i) { auto node = g.node(i); - // Check if the node is of type N and not a branching node + // Check if the node is of type N1 and not a branching node if(node && node->type() == N1::node_type && node->output_edges().size() == 1) { 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 + // Check if following node is a type N2 node if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == N2::node_type) && prec(*output_edge->producer())) { fuse_fcn(g, output_edge, optional_arguments...); @@ -336,6 +338,266 @@ void fuse_layer(Graph &g, std::function const &prec, const F fuse } } } + +/** Fuse below operators: + * + * | Main operator | Post operators | + * |:--------------|:---------------------------| + * |conv | add | + * |conv | act + add | + * |conv | add + act | + * |conv | act + add + act | + * + * Notes: currently, only GEMM supports fusion with post operator +*/ +template +void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const std::set &supported_fused_activations) +{ + ARM_COMPUTE_ERROR_ON(output_edge == nullptr); + + auto *conv_node = arm_compute::utils::cast::polymorphic_downcast(output_edge->producer()); + ARM_COMPUTE_ERROR_ON(conv_node->output(0) == nullptr); + // Prevent fusion if fused node has an output accessor + if(conv_node->output(0)->accessor() == nullptr) + { + // If data type is FP32/FP16, data layout is NHWC, and filter size if 1x1, fuse convolution with post op, as Conv1x1 always leads to GEMM. + const Edge *input_edge = conv_node->input_edge(1); + if(input_edge != nullptr && input_edge->tensor() != nullptr) + { + const DataLayout data_layout = input_edge->tensor()->desc().layout; + const DataType data_type = input_edge->tensor()->desc().data_type; + const TensorShape tensor_shape = input_edge->tensor()->desc().shape; + if(data_layout != DataLayout::NHWC || is_data_type_float(data_type) == false || tensor_shape.y() != 1 || tensor_shape.z() != 1) + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n"); + return; + } + } + else + { + return; + } + + std::list post_op_node_list = {}; + int eltwise_adden_input_id = 0; + int prev_op_dst_pos = 0; // Previous operator dst's postion in current operator + NodeID prev_op_dst_id = conv_node->id(); + for(unsigned int i = conv_node_id + 1; i < g.nodes().size(); ++i) + { + auto post_op_node = g.node(i); + bool fusable_post_op = false; + if(post_op_node != nullptr && post_op_node->output_edges().size() > 0) + { + const auto post_op_output_edge_id = *post_op_node->output_edges().begin(); + const auto post_op_output_edge = g.edge(post_op_output_edge_id); + + if(post_op_output_edge != nullptr) + { + switch(post_op_output_edge->producer()->type()) + { + case EltwiseLayerNode::node_type: + { + auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast(post_op_output_edge->producer()); + ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); + if(eltwise_node->output(0)->accessor() == nullptr) + { + post_op_node_list.push_back(post_op_output_edge->producer()); + fusable_post_op = true; + + // Extract elementwise inputs + const auto eltwise_input_id_0 = eltwise_node->input_edge(0)->producer_id(); + const auto eltwise_input_id_1 = eltwise_node->input_edge(1)->producer_id(); + if(eltwise_input_id_0 == prev_op_dst_id) + { + eltwise_adden_input_id = eltwise_input_id_1; + prev_op_dst_pos = 0; + } + else if(eltwise_input_id_1 == prev_op_dst_id) + { + eltwise_adden_input_id = eltwise_input_id_0; + prev_op_dst_pos = 1; + } + } + else + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with elementwise due to the presence of an output accessor\n"); + } + break; + } + case ActivationLayerNode::node_type: + { + auto *act_node = arm_compute::utils::cast::polymorphic_downcast(post_op_output_edge->producer()); + ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr); + // Check if activation is supported for fusion + if(supported_fused_activations.count(act_node->activation_info().activation()) == 0) + { + break; + } + if(act_node->output(0)->accessor() == nullptr) + { + post_op_node_list.push_back(post_op_output_edge->producer()); + fusable_post_op = true; + prev_op_dst_id = act_node->id(); + } + else + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with activation due to the presence of an output accessor\n"); + } + break; + } + default: + break; + } + } + + // Check if the node is not a branching node and current node is fusable + if(post_op_node->output_edges().size() == 1 && fusable_post_op == true) + { + continue; + } + else + { + break; + } + } + } + + if(post_op_node_list.size() == 0) + { + return; + } + else if(post_op_node_list.size() == 1) // Use fusion without post op if post op only contains one activation operator + { + for(const auto &post_op : post_op_node_list) + { + if(post_op->type() == ActivationLayerNode::node_type) + { + post_op_node_list.clear(); + return; + } + } + } + else // Use fusion with post op if there're two or more operators + { + 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 conv_info = conv_node->convolution_info(); + const auto conv_method = conv_node->convolution_method(); + const auto num_groups = conv_node->num_groups(); + FastMathHint fast_math_hint = conv_node->fast_math_hint(); + + // Create the fused node + const NodeID fused_id = g.add_node(conv_info, num_groups, conv_method, fast_math_hint); + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing convolution node with ID : " << conv_node->id()); + + // Add connections from the conv 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); + 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); + } + g.add_connection(eltwise_adden_input_id, 0, fused_id, 3); + g.remove_node(conv_node->id()); + + // Update fused node outputs + auto fused_node = g.node(fused_id); + auto *fused_conv_node = arm_compute::utils::cast::polymorphic_downcast(fused_node); + fused_node->set_assigned_target(assigned_target); + + unsigned int op_idx = 0; + // Fuse post operators with conv + for(const auto &post_op : post_op_node_list) + { + switch(post_op->type()) + { + case EltwiseLayerNode::node_type: + { + auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast(post_op); + ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); + + fused_conv_node->post_op_info_list().push_back(std::make_unique(prev_op_dst_pos, eltwise_node->convert_policy())); + ARM_COMPUTE_LOG_GRAPH_VERBOSE(" with Elementwise Layer node with ID : " << post_op->id()); + break; + } + case ActivationLayerNode::node_type: + { + auto *act_node = arm_compute::utils::cast::polymorphic_downcast(post_op); + ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr); + + fused_conv_node->post_op_info_list().push_back(std::make_unique(act_node->activation_info())); + ARM_COMPUTE_LOG_GRAPH_VERBOSE(" with Activation Layer node with ID : " << post_op->id()); + break; + } + default: + break; + } + + if(op_idx == post_op_node_list.size() - 1) // last fusable node + { + // Get driving nodes of last fusable node + std::vector last_driving_nodes = get_driving_nodes(*post_op); + + // Extract last fusable node accessor if any + auto last_node_accessor = post_op->output(0)->extract_accessor(); + + // Remove node + g.remove_node(post_op->id()); + + // Update fused node outputs + for(auto &driving_node : last_driving_nodes) + { + g.add_connection(fused_id, 0, driving_node.node_id, driving_node.index); + configure_tensor(fused_node->output(0)); + } + + // Update accessor to fused node + fused_node->output(0)->set_accessor(std::move(last_node_accessor)); + } + else + { + // Remove node + g.remove_node(post_op->id()); + } + op_idx++; + } + post_op_node_list.clear(); + ARM_COMPUTE_LOG_GRAPH_VERBOSE(std::endl); + } + } + else + { + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops 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) +{ + // Note that fused nodes may be added to the end of the node list. + // Instead of only looping over the original list of nodes, we loop over the current node list which could be growing. + // This is intentional as it probes the newly added fused nodes for further fusing opportunities. + for(unsigned int i = 0; i < g.nodes().size(); ++i) + { + auto node = g.node(i); + // Check if the node is of type N1 and not a branching node + if(node && node->type() == N1::node_type && node->output_edges().size() == 1) + { + const auto output_edge_id = *node->output_edges().begin(); + const auto output_edge = g.edge(output_edge_id); + + // Check if it's the correct target + if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && prec(*output_edge->producer())) + { + fuse_fcn(g, output_edge, i, optional_arguments...); + } + } + } +} } // namespace detail const char *NodeFusionMutator::name() @@ -381,15 +643,17 @@ void NodeFusionMutator::mutate(Graph &g) }; // Fusion mutations + + detail::fuse_layer(g, cl_target_prec, detail::fuse_convolution, supported_fused_activations); + detail::fuse_layer(g, empty_prec, detail::fuse_node_with_activation, supported_fused_activations); + detail::fuse_layer(g, empty_prec, detail::fuse_convolution_with_batch_normalization); detail::fuse_layer(g, empty_prec, detail::fuse_pad_with_convolution); detail::fuse_layer(g, empty_prec, detail::fuse_pad_with_convolution); 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_depthwise_convolution_with_batch_normalization); detail::fuse_layer(g, empty_prec, detail::fuse_node_with_activation, supported_fused_activations); detail::fuse_layer(g, cl_target_prec, detail::fuse_node_with_activation, supported_fused_activations); - detail::fuse_layer(g, empty_prec, detail::fuse_convolution_with_batch_normalization); - detail::fuse_layer(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization); } } // namespace graph } // namespace arm_compute diff --git a/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp b/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp new file mode 100644 index 0000000000..c7ca59d863 --- /dev/null +++ b/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp @@ -0,0 +1,164 @@ +/* + * Copyright (c) 2021 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/FusedConvolutionWithPostOpNode.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 +{ +FusedConvolutionWithPostOpNode::FusedConvolutionWithPostOpNode(PadStrideInfo info, + unsigned int num_groups, + ConvolutionMethod method, + FastMathHint fast_math_hint, + QuantizationInfo out_quant_info) + : _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(std::move(out_quant_info)), _fused_activation(), + _post_op_info_list(std::list> {}) +{ + _input_edges.resize(4, EmptyEdgeID); + _outputs.resize(1, NullTensorID); +} + +void FusedConvolutionWithPostOpNode::set_convolution_method(ConvolutionMethod method) +{ + _method = method; +} + +ConvolutionMethod FusedConvolutionWithPostOpNode::convolution_method() const +{ + return _method; +} + +void FusedConvolutionWithPostOpNode::set_fast_math_hint(FastMathHint hint) +{ + _fast_math_hint = hint; +} + +FastMathHint FusedConvolutionWithPostOpNode::fast_math_hint() const +{ + return _fast_math_hint; +} + +const std::list> &FusedConvolutionWithPostOpNode::post_op_info_list() const +{ + return _post_op_info_list; +} + +std::list> &FusedConvolutionWithPostOpNode::post_op_info_list() +{ + return _post_op_info_list; +} + +PadStrideInfo FusedConvolutionWithPostOpNode::convolution_info() const +{ + return _info; +} + +unsigned int FusedConvolutionWithPostOpNode::num_groups() const +{ + return _num_groups; +} + +ActivationLayerInfo FusedConvolutionWithPostOpNode::fused_activation() const +{ + return _fused_activation; +} + +void FusedConvolutionWithPostOpNode::set_fused_activation(ActivationLayerInfo fused_activation) +{ + _fused_activation = fused_activation; +} + +void FusedConvolutionWithPostOpNode::set_convolution_info(PadStrideInfo info) +{ + _info = info; +} + +TensorDescriptor FusedConvolutionWithPostOpNode::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); + + const DataLayout data_layout = input_descriptor.layout; + TensorDescriptor output_descriptor = input_descriptor; + output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::WIDTH), output_width); + output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::HEIGHT), output_height); + output_descriptor.shape.set(get_dimension_idx(data_layout, DataLayoutDimension::CHANNEL), weights_descriptor.shape[3]); + + return output_descriptor; +} + +bool FusedConvolutionWithPostOpNode::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 FusedConvolutionWithPostOpNode::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 FusedConvolutionWithPostOpNode::type() const +{ + return FusedConvolutionWithPostOpNode::node_type; +} + +void FusedConvolutionWithPostOpNode::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 08f7f25ee5..47371e34d5 100644 --- a/src/graph/printers/DotGraphPrinter.cpp +++ b/src/graph/printers/DotGraphPrinter.cpp @@ -85,6 +85,14 @@ void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n) _info = ss.str(); } +void DotGraphVisitor::visit(FusedConvolutionWithPostOpNode &n) +{ + ARM_COMPUTE_UNUSED(n); + std::stringstream ss; + ss << "FusedConvolutionWithPostOpNode"; + _info = ss.str(); +} + void DotGraphVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) { ARM_COMPUTE_UNUSED(n); -- cgit v1.2.1