diff options
author | Sheri Zhang <sheri.zhang@arm.com> | 2021-11-02 10:45:07 +0000 |
---|---|---|
committer | Sheri Zhang <sheri.zhang@arm.com> | 2021-11-03 17:08:05 +0000 |
commit | fb2280381e7a98ad698ea0c1b2cd635a48ad4acc (patch) | |
tree | e3fab3cff60b806e725ba9c771617e41c654604e /src/graph/mutators/NodeFusionMutator.cpp | |
parent | bc788389dcc7bd682f53a85803f6a202d42ac828 (diff) | |
download | ComputeLibrary-fb2280381e7a98ad698ea0c1b2cd635a48ad4acc.tar.gz |
Add graph level convolution fusion with post operator
Resolves: COMPMID-4701
Signed-off-by: Sheri Zhang <sheri.zhang@arm.com>
Change-Id: I8a0d3c2ed4bf84489d94b8ae6641d6041aadaee5
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6557
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/graph/mutators/NodeFusionMutator.cpp')
-rw-r--r-- | src/graph/mutators/NodeFusionMutator.cpp | 274 |
1 files changed, 269 insertions, 5 deletions
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 <list> #include <set> namespace arm_compute @@ -322,13 +324,13 @@ void fuse_layer(Graph &g, std::function<bool(INode &)> 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<bool(INode &)> 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 <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); + // 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<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) + { + 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<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; + + // 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<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; + } + 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<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); + } + 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<FusedConvolutionWithPostOpNode *>(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<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)); + } + + // 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 <typename N1, typename F, typename... Args> +void fuse_layer(Graph &g, std::function<bool(INode &)> 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<ConvolutionLayerNode>(g, cl_target_prec, detail::fuse_convolution<ConvolutionLayerNode>, supported_fused_activations); + detail::fuse_layer<ConvolutionLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<ConvolutionLayerNode>, supported_fused_activations); + detail::fuse_layer<ConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_convolution_with_batch_normalization); 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>); 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); + detail::fuse_layer<DepthwiseConvolutionLayerNode, BatchNormalizationLayerNode>(g, empty_prec, detail::fuse_depthwise_convolution_with_batch_normalization); detail::fuse_layer<FullyConnectedLayerNode, ActivationLayerNode>(g, empty_prec, detail::fuse_node_with_activation<FullyConnectedLayerNode>, supported_fused_activations); detail::fuse_layer<EltwiseLayerNode, ActivationLayerNode>(g, cl_target_prec, detail::fuse_node_with_activation<EltwiseLayerNode>, supported_fused_activations); - 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); } } // namespace graph } // namespace arm_compute |