diff options
author | Sheri Zhang <sheri.zhang@arm.com> | 2021-11-03 21:24:00 +0000 |
---|---|---|
committer | Sheri Zhang <sheri.zhang@arm.com> | 2021-11-04 15:35:58 +0000 |
commit | c65023e01d8e5c8f5af15dd8e2ca4fae3c4889f9 (patch) | |
tree | 313c5c0786971986d92c2e59073eb06406f43990 /src/graph/mutators/NodeFusionMutator.cpp | |
parent | cc859158faff318a81dd9aafcc5b609d7dd5d4a9 (diff) | |
download | ComputeLibrary-c65023e01d8e5c8f5af15dd8e2ca4fae3c4889f9.tar.gz |
Refine code for graph level fusion
Partially resolves: COMPMID-4701
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I11fb89bd31ef5a98ad9a2d6bb5fd0c1e8371b52f
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6572
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/graph/mutators/NodeFusionMutator.cpp')
-rw-r--r-- | src/graph/mutators/NodeFusionMutator.cpp | 450 |
1 files changed, 249 insertions, 201 deletions
diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index 463d1cd8f6..3f979e48ee 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -44,6 +44,40 @@ namespace graph { namespace detail { +void transfer_driving_nodes_and_remove_old_node(Graph &g, INode *new_node, INode *old_node, bool add_output_tensor) +{ + if(new_node == nullptr || old_node == nullptr) + { + return; + } + + // Get driving nodes of last fusable node + std::vector<NodeIdxPair> last_driving_nodes = get_driving_nodes(*old_node); + + // Extract last fusable node accessor if any + if(old_node->output(0) == nullptr) + { + return; + } + auto old_node_accessor = old_node->output(0)->extract_accessor(); + + // Remove node + g.remove_node(old_node->id()); + + // Update fused node outputs + for(auto &driving_node : last_driving_nodes) + { + g.add_connection(new_node->id(), 0, driving_node.node_id, driving_node.index); + if(add_output_tensor) + { + configure_tensor(new_node->output(0)); + } + } + + // Update accessor to fused node + new_node->output(0)->set_accessor(std::move(old_node_accessor)); +} + void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge) { ARM_COMPUTE_ERROR_ON(output_edge == nullptr); @@ -107,24 +141,11 @@ void fuse_convolution_with_batch_normalization(Graph &g, const Edge *output_edge g.add_connection(bn_gamma_id, 0, fused_id, 6); } - auto fused_node = g.node(fused_id); - std::vector<NodeIdxPair> 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(); + auto fused_node = g.node(fused_id); + auto bn_node_name = bn_node->name(); - // Remove batch normalization node - g.remove_node(bn_node->id()); + transfer_driving_nodes_and_remove_old_node(g, fused_node, bn_node, true); - // 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 }); @@ -184,24 +205,11 @@ void fuse_depthwise_convolution_with_batch_normalization(Graph &g, const Edge *o 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<NodeIdxPair> 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(); + auto fused_node = g.node(fused_id); + auto bn_node_name = bn_node->name(); - // Remove batch normalization node - g.remove_node(bn_node->id()); + transfer_driving_nodes_and_remove_old_node(g, fused_node, bn_node, true); - // 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{ depth_conv_node->name() + "+" + bn_node_name, assigned_target }); @@ -242,26 +250,10 @@ void fuse_node_with_activation(Graph &g, const Edge *output_edge, const std::set // Prevent fusion if fused node has an output accessor if(n_node->output(0)->accessor() == nullptr) { - // Get driving nodes of activation node - std::vector<NodeIdxPair> 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)); + transfer_driving_nodes_and_remove_old_node(g, n_node, act_node, false); } else { @@ -339,7 +331,7 @@ void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse } } -/** Fuse below operators: +/** Check valid combinations: * * | Main operator | Post operators | * |:--------------|:---------------------------| @@ -348,136 +340,246 @@ void fuse_layer(Graph &g, std::function<bool(INode &)> const &prec, const F fuse * |conv | add + act | * |conv | act + add + act | * - * Notes: currently, only GEMM supports fusion with post operator */ -template <typename N> -void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const std::set<Activation> &supported_fused_activations) +#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) { - ARM_COMPUTE_ERROR_ON(output_edge == nullptr); + if(len > MAX_POST_OP_NUM || len <= 0) + { + return false; + } - auto *conv_node = arm_compute::utils::cast::polymorphic_downcast<N *>(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) + bool found = false; + for(int i = 0; i < MAX_VALIDE_COMBINATION; ++i) { - // 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) + for(int j = 0; j < len; ++j) { - 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) + if(post_op_type[j] != valide_post_op_type[i][j]) { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n"); - return; + 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 { - return; + // Remove node + g.remove_node(post_op->id()); } + op_idx++; + } +} - std::list<INode *> 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) +std::list<INode *> get_post_op_list(Graph &g, int &eltwise_operand_id, int &prev_op_dst_pos, 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; + 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) { - 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); + 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) + if(post_op_output_edge != nullptr) + { + switch(post_op_output_edge->producer()->type()) { - switch(post_op_output_edge->producer()->type()) + case EltwiseLayerNode::node_type: { - case EltwiseLayerNode::node_type: + auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast<EltwiseLayerNode *>(post_op_output_edge->producer()); + ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); + if(eltwise_node->output(0)->accessor() == nullptr) { - auto *eltwise_node = arm_compute::utils::cast::polymorphic_downcast<EltwiseLayerNode *>(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; + 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) { - 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; - } + eltwise_operand_id = eltwise_input_id_1; + prev_op_dst_pos = 0; } - else + else if(eltwise_input_id_1 == prev_op_dst_id) { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with elementwise due to the presence of an output accessor\n"); + eltwise_operand_id = eltwise_input_id_0; + prev_op_dst_pos = 1; } - break; } - case ActivationLayerNode::node_type: + else { - auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(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; + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with elementwise due to the presence of an output accessor\n"); } - default: + break; + } + case ActivationLayerNode::node_type: + { + auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(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; + 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 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; - } + // 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 && post_op_node_list.size() < 3) + { + continue; + } + else + { + break; } } + } - if(post_op_node_list.size() == 0) + // 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) { - return; + post_op_node_list.clear(); } - else if(post_op_node_list.size() == 1) // Use fusion without post op if post op only contains one activation operator + } + + 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 +*/ +template <typename N> +void fuse_convolution(Graph &g, const Edge *output_edge, 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<N *>(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 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) { - for(const auto &post_op : post_op_node_list) + 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) { - if(post_op->type() == ActivationLayerNode::node_type) - { - post_op_node_list.clear(); - return; - } + ARM_COMPUTE_LOG_GRAPH_VERBOSE("Prevented fusion of convolution node with post ops due to non GEMM convolution\n"); + return; } } - else // Use fusion with post op if there're two or more operators + 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(); @@ -501,70 +603,16 @@ void fuse_convolution(Graph &g, const Edge *output_edge, int conv_node_id, const 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.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); - auto *fused_conv_node = arm_compute::utils::cast::polymorphic_downcast<FusedConvolutionWithPostOpNode *>(fused_node); + auto fused_node = g.node(fused_id); 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<EltwiseLayerNode *>(post_op); - ARM_COMPUTE_ERROR_ON(eltwise_node->output(0) == nullptr); - - fused_conv_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_conv_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 - { - // Get driving nodes of last fusable node - std::vector<NodeIdxPair> 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)); - } + // Fuse convolution with post op + fuse_convolution_with_post_op(g, fused_node, post_op_node_list, prev_op_dst_pos); - // 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); } |