diff options
Diffstat (limited to 'src')
-rw-r--r-- | src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp | 3 | ||||
-rw-r--r-- | src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp | 3 | ||||
-rw-r--r-- | src/graph/mutators/NodeFusionMutator.cpp | 17 |
3 files changed, 18 insertions, 5 deletions
diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp index 5999c66056..4c93fb28bf 100644 --- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp +++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp @@ -64,7 +64,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, { ActivationLayerInfo::ActivationFunction act = act_info.activation(); ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32 && input->data_type() != DataType::F16); - ARM_COMPUTE_RETURN_ERROR_ON(act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU + ARM_COMPUTE_RETURN_ERROR_ON(act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU + && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); ARM_COMPUTE_RETURN_ERROR_ON(act_info.b() > act_info.a()); } diff --git a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp index 6aed41f3aa..abfaa0cd26 100644 --- a/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEBatchNormalizationLayerKernel.cpp @@ -50,7 +50,8 @@ validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const IT { ActivationLayerInfo::ActivationFunction act = act_info.activation(); ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() != DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON(act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU + ARM_COMPUTE_RETURN_ERROR_ON(act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::RELU + && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && act != ActivationLayerInfo::ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); ARM_COMPUTE_RETURN_ERROR_ON(act_info.b() > act_info.a()); } diff --git a/src/graph/mutators/NodeFusionMutator.cpp b/src/graph/mutators/NodeFusionMutator.cpp index 39209d2f49..6677330cec 100644 --- a/src/graph/mutators/NodeFusionMutator.cpp +++ b/src/graph/mutators/NodeFusionMutator.cpp @@ -29,6 +29,8 @@ #include "arm_compute/core/utils/misc/Cast.h" +#include <set> + namespace arm_compute { namespace graph @@ -37,6 +39,9 @@ namespace detail { void fuse_batch_norm_with_activation(Graph &g) { + // Supported activations when fusing + const std::set<Activation> supported_fused_activations = { Activation::RELU, Activation::BOUNDED_RELU, Activation::LU_BOUNDED_RELU }; + // Not interested in the order of nodes for(auto &node : g.nodes()) { @@ -48,14 +53,20 @@ void fuse_batch_norm_with_activation(Graph &g) // Check if following node is an activation layer node if((output_edge != nullptr) && (output_edge->consumer() != nullptr) && (output_edge->consumer()->type() == NodeType::ActivationLayer)) { - ARM_COMPUTE_LOG_GRAPH_VERBOSE("Fusing Batch Normalization node with ID : " << output_edge->producer_id() - << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl); - auto *bn_node = arm_compute::utils::cast::polymorphic_downcast<BatchNormalizationLayerNode *>(output_edge->producer()); auto *act_node = arm_compute::utils::cast::polymorphic_downcast<ActivationLayerNode *>(output_edge->consumer()); ARM_COMPUTE_ERROR_ON(act_node->output(0) == nullptr || bn_node->output(0) == nullptr); + // 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 Batch Normalization node with ID : " << output_edge->producer_id() + << " with Activation Layer node with ID : " << output_edge->consumer_id() << std::endl); + // Prevent fusion if batch normalization node has an output accessor if(bn_node->output(0)->accessor() == nullptr) { |