diff options
Diffstat (limited to 'src/graph')
-rw-r--r-- | src/graph/DataLayerVisitor.cpp | 10 | ||||
-rw-r--r-- | src/graph/INode.cpp | 15 | ||||
-rw-r--r-- | src/graph/INodeVisitor.cpp | 10 | ||||
-rw-r--r-- | src/graph/backends/CL/CLFunctionsFactory.cpp | 6 | ||||
-rw-r--r-- | src/graph/backends/CL/CLNodeValidator.cpp | 4 | ||||
-rw-r--r-- | src/graph/mutators/NodeFusionMutator.cpp | 442 | ||||
-rw-r--r-- | src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp | 138 | ||||
-rw-r--r-- | src/graph/nodes/FusedConvolutionWithPostOpNode.cpp | 153 | ||||
-rw-r--r-- | src/graph/printers/DotGraphPrinter.cpp | 18 |
9 files changed, 7 insertions, 789 deletions
diff --git a/src/graph/DataLayerVisitor.cpp b/src/graph/DataLayerVisitor.cpp index 85d24b4654..073ffd413d 100644 --- a/src/graph/DataLayerVisitor.cpp +++ b/src/graph/DataLayerVisitor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -131,14 +131,6 @@ void DataLayerVisitor::visit(FusedConvolutionBatchNormalizationNode &n) add_convolution_layer_method<FusedConvolutionBatchNormalizationNode>(_layer_data, n); } -void DataLayerVisitor::visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) -{ - _layer_data.clear(); - add_generic_layer_data<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n); - add_convolution_layer_data<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n); - add_convolution_layer_method<FusedConvolutionBatchNormalizationWithPostOpsNode>(_layer_data, n); -} - void DataLayerVisitor::visit(FusedDepthwiseConvolutionBatchNormalizationNode &n) { _layer_data.clear(); diff --git a/src/graph/INode.cpp b/src/graph/INode.cpp index e5b4adda26..70fe44e134 100644 --- a/src/graph/INode.cpp +++ b/src/graph/INode.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018,2021 Arm Limited. + * Copyright (c) 2018,2021,2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -37,7 +37,6 @@ namespace graph INode::INode() : _graph(nullptr), _id(EmptyNodeID), _common_params({ "", Target::UNSPECIFIED}), _outputs(), _input_edges(), _output_edges(), _assigned_target(Target::UNSPECIFIED) - ,_post_op_info_list(std::list<std::unique_ptr<ConvPostOpInfo>> {}) { } // clang-format on @@ -200,15 +199,5 @@ Target INode::assigned_target() const { return _assigned_target; } - -const std::list<std::unique_ptr<ConvPostOpInfo>> &INode::post_op_info_list() const -{ - return _post_op_info_list; -} - -std::list<std::unique_ptr<ConvPostOpInfo>> &INode::post_op_info_list() -{ - return _post_op_info_list; -} } // namespace graph -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/src/graph/INodeVisitor.cpp b/src/graph/INodeVisitor.cpp index f067d618bd..5369f6f539 100644 --- a/src/graph/INodeVisitor.cpp +++ b/src/graph/INodeVisitor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -85,14 +85,6 @@ void DefaultNodeVisitor::visit(FusedConvolutionBatchNormalizationNode &n) { default_visit(n); } -void DefaultNodeVisitor::visit(FusedConvolutionBatchNormalizationWithPostOpsNode &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 c67f6a538b..882810474e 100644 --- a/src/graph/backends/CL/CLFunctionsFactory.cpp +++ b/src/graph/backends/CL/CLFunctionsFactory.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -274,8 +274,6 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext & return detail::create_fully_connected_layer<CLFullyConnectedLayer, CLTargetInfo>(*polymorphic_downcast<FullyConnectedLayerNode *>(node), ctx); case NodeType::FusedConvolutionBatchNormalizationLayer: return detail::create_fused_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(node), ctx); - case NodeType::FusedConvolutionWithPostOp: - return detail::create_fused_convolution_with_post_op<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionWithPostOpNode *>(node), ctx); case NodeType::FusedDepthwiseConvolutionBatchNormalizationLayer: return detail::create_fused_depthwise_convolution_batch_normalization_layer<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedDepthwiseConvolutionBatchNormalizationNode *>(node), ctx); case NodeType::GenerateProposalsLayer: @@ -318,8 +316,6 @@ std::unique_ptr<IFunction> CLFunctionFactory::create(INode *node, GraphContext & return detail::create_stack_layer<CLStackLayer, CLTargetInfo>(*polymorphic_downcast<StackLayerNode *>(node)); case NodeType::StridedSliceLayer: return detail::create_strided_slice_layer<CLStridedSlice, CLTargetInfo>(*polymorphic_downcast<StridedSliceLayerNode *>(node)); - case NodeType::FusedConvolutionBatchNormalizationLayerWithPostOpsLayer: - return detail::create_fused_convolution_batch_normalization_with_post_op<CLFusedLayerTypes, CLTargetInfo>(*polymorphic_downcast<FusedConvolutionBatchNormalizationWithPostOpsNode *>(node), ctx); default: return nullptr; } diff --git a/src/graph/backends/CL/CLNodeValidator.cpp b/src/graph/backends/CL/CLNodeValidator.cpp index c50782db48..8fd8c14f63 100644 --- a/src/graph/backends/CL/CLNodeValidator.cpp +++ b/src/graph/backends/CL/CLNodeValidator.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -76,8 +76,6 @@ Status CLNodeValidator::validate(INode *node) CLDirectConvolutionLayer, CLGEMMConvolutionLayer, CLWinogradConvolutionLayer>(*polymorphic_downcast<ConvolutionLayerNode *>(node)); - case NodeType::FusedConvolutionWithPostOp: - return detail::validate_fused_convolution_with_post_op<CLGEMMConvolutionLayer>(*polymorphic_downcast<FusedConvolutionWithPostOpNode *>(node)); case NodeType::DepthToSpaceLayer: return detail::validate_depth_to_space_layer<CLDepthToSpaceLayer>(*polymorphic_downcast<DepthToSpaceLayerNode *>(node)); case NodeType::DepthwiseConvolutionLayer: diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index 8eb3e4cb71..38284b93cf 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -29,8 +29,6 @@ #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/FusedConvolutionBatchNormalizationWithPostOpsNode.h" -#include "arm_compute/graph/nodes/FusedConvolutionWithPostOpNode.h" #include "arm_compute/graph/nodes/Nodes.h" #include "src/graph/mutators/MutatorUtils.h" @@ -333,441 +331,6 @@ void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse } } -/** Check valid combinations: - * - * | Main operator | Post operators | - * |:--------------|:---------------------------| - * |conv | add | - * |conv | act + add | - * |conv | add + act | - * |conv | act + add + act | - * -*/ -#define MAX_VALIDE_COMBINATION 4 -#define MAX_POST_OP_NUM 3 -NodeType valide_post_op_type[MAX_VALIDE_COMBINATION][MAX_POST_OP_NUM] = { { EltwiseLayerNode::node_type }, - { EltwiseLayerNode::node_type, ActivationLayerNode::node_type }, - { ActivationLayerNode::node_type, EltwiseLayerNode::node_type }, - { ActivationLayerNode::node_type, EltwiseLayerNode::node_type, ActivationLayerNode::node_type } -}; - -bool check_post_op_type(NodeType *post_op_type, int len) -{ - if(len > MAX_POST_OP_NUM || len <= 0) - { - return false; - } - - bool found = false; - for(int i = 0; i < MAX_VALIDE_COMBINATION; ++i) - { - for(int j = 0; j < len; ++j) - { - if(post_op_type[j] != valide_post_op_type[i][j]) - { - found = false; - break; - } - found = true; - } - if(found) - break; - } - - return found; -} - -void fuse_convolution_with_post_op(Graph &g, INode *fused_node, std::list<INode *> post_op_node_list, int prev_op_dst_pos) -{ - 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<EltwiseLayerNode *>(post_op); - ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); - - fused_node->post_op_info_list().push_back(std::make_unique<ConvPostOpInfoEltwiseAdd>(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<ActivationLayerNode *>(post_op); - ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr); - - fused_node->post_op_info_list().push_back(std::make_unique<ConvPostOpInfoActivation>(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 - { - transfer_driving_nodes_and_remove_old_node(g, fused_node, post_op, true); - } - else - { - // Remove node - g.remove_node(post_op->id()); - } - op_idx++; - } -} - -std::list<INode *> get_post_op_list(Graph &g, int &eltwise_operand_id, int &prev_op_dst_pos, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations) -{ - std::list<INode *> post_op_node_list = {}; - NodeID prev_op_dst_id = conv_node_id; - NodeType post_op_type_list[3] = { NodeType::Dummy, NodeType::Dummy, NodeType::Dummy }; - int post_op_idx = 0; - - // Get list of the connected nodes - auto current_node = g.node(conv_node_id); - - while(post_op_node_list.size() < 3) - { - // This convolution node must have only one output edge, otherwise this function would not have been called - - auto current_output_edge_id = current_node->output_edges().begin(); - auto current_output_edge = g.edge(*current_output_edge_id); - auto post_op_node = current_output_edge->consumer(); - - bool fusable_post_op = false; - if(post_op_node != nullptr && post_op_node->output_edges().size() > 0) - { - switch(post_op_node->type()) - { - case EltwiseLayerNode::node_type: - { - auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast<EltwiseLayerNode *>(post_op_node); - ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); - if(eltwise_node->output(0)->accessor() == nullptr) - { - post_op_node_list.push_back(post_op_node); - fusable_post_op = true; - post_op_type_list[post_op_idx++] = eltwise_node->type(); - - // 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_operand_id = eltwise_input_id_1; - prev_op_dst_pos = 0; - } - else if(eltwise_input_id_1 == prev_op_dst_id) - { - eltwise_operand_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<ActivationLayerNode *>(post_op_node); - 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_node); - fusable_post_op = true; - post_op_type_list[post_op_idx++] = act_node->type(); - prev_op_dst_id = act_node->id(); - } - else - { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops 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) - { - current_node = post_op_node; - } - else - { - break; - } - } - } - - // Check whether it's valid post op list - if(post_op_node_list.size() > 0) - { - bool fuse_with_post_op = check_post_op_type(post_op_type_list, post_op_node_list.size()); - if(!fuse_with_post_op) - { - post_op_node_list.clear(); - } - } - - return post_op_node_list; -} - -/** 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 -*/ -void fuse_convolution_with_post_ops(Graph &g, const Edge *output_edge, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations) -{ - ARM_COMPUTE_ERROR_ON(output_edge == nullptr); - - auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<ConvolutionLayerNode *>(output_edge->producer()); - ARM_COMPUTE_ERROR_ON(conv_node->output(0) == nullptr); - - const ConvolutionMethod conv_algorithm = conv_node->convolution_method(); - if(conv_algorithm != ConvolutionMethod::GEMM) - { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n"); - return; - } - - // 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 is 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; - } - - // Get post op list - int eltwise_operand_id = 0; - int prev_op_dst_pos = 0; // Previous operator dst's postion in current operator - std::list<INode *> post_op_node_list = get_post_op_list(g, eltwise_operand_id, prev_op_dst_pos, conv_node_id, supported_fused_activations); - - if(post_op_node_list.size() == 0) - { - return; - } - else // Do convolution fusion with post op if there're one(elementwise), 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<FusedConvolutionWithPostOpNode>(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); - } - // Adding the Element wise operand in case the post op is element wise operation - auto it = std::find_if(post_op_node_list.begin(), - post_op_node_list.end(), - [&](const INode * nd) - { - return (nd->type() == graph::NodeType::EltwiseLayer); - }); - - if(it != post_op_node_list.end()) - { - g.add_connection(eltwise_operand_id, 0, fused_id, 3); - } - g.remove_node(conv_node->id()); - - // Update fused node outputs - auto fused_node = g.node(fused_id); - fused_node->set_assigned_target(assigned_target); - - // Fuse convolution with post op - fuse_convolution_with_post_op(g, fused_node, post_op_node_list, prev_op_dst_pos); - - 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"); - } -} - -void fuse_convolution_batch_normalization_with_post_ops(Graph &g, const Edge *output_edge, unsigned int conv_node_id, const std::set<Activation> &supported_fused_activations) -{ - ARM_COMPUTE_ERROR_ON(output_edge == nullptr); - - auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<FusedConvolutionBatchNormalizationNode *>(output_edge->producer()); - ARM_COMPUTE_ERROR_ON(conv_node->output(0) == nullptr); - const ConvolutionMethod conv_algorithm = conv_node->convolution_method(); - if(conv_algorithm != ConvolutionMethod::GEMM) - { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n"); - return; - } - - // 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 is 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; - } - - // Get post op list - int eltwise_operand_id = 0; - int prev_op_dst_pos = 0; // Previous operator dst's postion in current operator - std::list<INode *> post_op_node_list = get_post_op_list(g, eltwise_operand_id, prev_op_dst_pos, conv_node_id, supported_fused_activations); - - if(post_op_node_list.size() == 0) - { - return; - } - else // Do convolution fusion with post op if there're one(elementwise), 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 bn_mean_id = conv_node->input_edge(3)->producer_id(); - const auto bn_var_id = conv_node->input_edge(4)->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 float epsilon = conv_node->epsilon(); - const NodeID fused_id = g.add_node<FusedConvolutionBatchNormalizationWithPostOpsNode>(epsilon, conv_info, num_groups, conv_method, fast_math_hint); - - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing FusedConvolutionBatchNormalization 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(bn_mean_id, 0, fused_id, 3); - g.add_connection(bn_var_id, 0, fused_id, 4); - - // Move connections of old FusedConvolutionBatchNormalization to the fused node - if(conv_node->input_edge(5) != nullptr) - { - const auto bn_beta_id = conv_node->input_edge(5)->producer_id(); - g.add_connection(bn_beta_id, 0, fused_id, 5); - } - - if(conv_node->input_edge(6) != nullptr) - { - const auto bn_gamma_id = conv_node->input_edge(6)->producer_id(); - g.add_connection(bn_gamma_id, 0, fused_id, 6); - } - - // Adding the Element wise operand in case the post op is element wise operation - auto it = std::find_if(post_op_node_list.begin(), - post_op_node_list.end(), - [&](const INode * nd) - { - return (nd->type() == graph::NodeType::EltwiseLayer); - }); - - if(it != post_op_node_list.end()) - { - g.add_connection(eltwise_operand_id, 0, fused_id, 7); - } - - // Update fused node outputs - auto fused_node = g.node(fused_id); - fused_node->set_assigned_target(assigned_target); - - auto conv_node_name = conv_node->name(); - - // collect the post ops names - std::string post_ops_name = ""; - for(auto &post_op : post_op_node_list) - { - post_ops_name += post_op->name(); - } - fused_node->set_common_node_parameters(NodeParams{ conv_node->name() + "+" + post_ops_name, assigned_target }); - - // Fuse convolution with post op - fuse_convolution_with_post_op(g, fused_node, post_op_node_list, prev_op_dst_pos); - - post_op_node_list.clear(); - g.remove_node(conv_node->id()); - 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 <typename N1, typename F, typename... Args> void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse_fcn, Args &&... optional_arguments) { @@ -839,10 +402,6 @@ void NodeFusionMutator::mutate(Graph &g) detail::fuse_layer<PadLayerNode, ConvolutionLayerNode>(g, empty_prec, detail::fuse_pad_with_convolution<ConvolutionLayerNode>); detail::fuse_layer<PadLayerNode, DepthwiseConvolutionLayerNode>(g, empty_prec, detail::fuse_pad_with_convolution<DepthwiseConvolutionLayerNode>); - // The fusion of PostOps to ConvolutionLayer: - // It must occur after the fusion of PadLayer into ConvolutionLayer - // It must occur before the fusion of normal ActivationLayer into ConvolutionLayer as it takes precedence - detail::fuse_layer<ConvolutionLayerNode>(g, cl_target_prec, detail::fuse_convolution_with_post_ops, supported_fused_activations); detail::fuse_layer<BatchNormalizationLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<BatchNormalizationLayerNode>, supported_fused_activations); detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations); detail::fuse_layer<DepthwiseConvolutionLayerNode, ActivationLayerNode>(g, qs8_prec, detail::fuse_node_with_activation<DepthwiseConvolutionLayerNode>, supported_fused_activations); @@ -851,7 +410,6 @@ void NodeFusionMutator::mutate(Graph &g) // The fusion of BatchNormalizationLayer must occur after the fusion of ActivationLayer. Because FusedConvolutionBatchNormalizationNode assumes the BatchNormalization is already fused with activation, if any detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization); detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization); - detail::fuse_layer<FusedConvolutionBatchNormalizationNode>(g, cl_target_prec, detail::fuse_convolution_batch_normalization_with_post_ops, supported_fused_activations); } } // namespace graph } // namespace arm_compute diff --git a/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp b/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp deleted file mode 100644 index af81f0369a..0000000000 --- a/src/graph/nodes/FusedConvolutionBatchNormalizationWithPostOpsNode.cpp +++ /dev/null @@ -1,138 +0,0 @@ -/* - * 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/FusedConvolutionBatchNormalizationWithPostOpsNode.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 -{ -FusedConvolutionBatchNormalizationWithPostOpsNode::FusedConvolutionBatchNormalizationWithPostOpsNode(float epsilon, PadStrideInfo info, - unsigned int num_groups, - ConvolutionMethod method, - FastMathHint fast_math_hint) - : _epsilon(epsilon), _info(std::move(info)), _num_groups(num_groups), _method(method), _fast_math_hint(fast_math_hint) -{ - _input_edges.resize(8, EmptyEdgeID); - _outputs.resize(1, NullTensorID); -} - -void FusedConvolutionBatchNormalizationWithPostOpsNode::set_convolution_method(ConvolutionMethod method) -{ - _method = method; -} - -float FusedConvolutionBatchNormalizationWithPostOpsNode::epsilon() const -{ - return _epsilon; -} - -ConvolutionMethod FusedConvolutionBatchNormalizationWithPostOpsNode::convolution_method() const -{ - return _method; -} - -void FusedConvolutionBatchNormalizationWithPostOpsNode::set_fast_math_hint(FastMathHint hint) -{ - _fast_math_hint = hint; -} - -FastMathHint FusedConvolutionBatchNormalizationWithPostOpsNode::fast_math_hint() const -{ - return _fast_math_hint; -} - -PadStrideInfo FusedConvolutionBatchNormalizationWithPostOpsNode::convolution_info() const -{ - return _info; -} - -unsigned int FusedConvolutionBatchNormalizationWithPostOpsNode::num_groups() const -{ - return _num_groups; -} - -TensorDescriptor FusedConvolutionBatchNormalizationWithPostOpsNode::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 FusedConvolutionBatchNormalizationWithPostOpsNode::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 FusedConvolutionBatchNormalizationWithPostOpsNode::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); - - return output_info; -} - -NodeType FusedConvolutionBatchNormalizationWithPostOpsNode::type() const -{ - return FusedConvolutionBatchNormalizationWithPostOpsNode::node_type; -} - -void FusedConvolutionBatchNormalizationWithPostOpsNode::accept(INodeVisitor &v) -{ - v.visit(*this); -} -} // namespace graph -} // namespace arm_compute diff --git a/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp b/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp deleted file mode 100644 index 63341e2760..0000000000 --- a/src/graph/nodes/FusedConvolutionWithPostOpNode.cpp +++ /dev/null @@ -1,153 +0,0 @@ -/* - * 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() -{ - _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; -} - -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 1071d50197..9c7c4248bb 100644 --- a/src/graph/printers/DotGraphPrinter.cpp +++ b/src/graph/printers/DotGraphPrinter.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -85,22 +85,6 @@ void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationNode &n) _info = ss.str(); } -void DotGraphVisitor::visit(FusedConvolutionBatchNormalizationWithPostOpsNode &n) -{ - ARM_COMPUTE_UNUSED(n); - std::stringstream ss; - ss << "FusedConvolutionBatchNormalizationWithPostOpsNode"; - _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); |