1 files changed, 104 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/NeonBatchNormalizationWorkload.cpp b/src/backends/neon/workloads/NeonBatchNormalizationWorkload.cpp
new file mode 100644
index 0000000000..44d5035431
--- /dev/null
+++ b/src/backends/neon/workloads/NeonBatchNormalizationWorkload.cpp
@@ -0,0 +1,104 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonBatchNormalizationWorkload.hpp"
+#include <backendsCommon/CpuTensorHandle.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+#include <armnn/ArmNN.hpp>
+
+namespace armnn
+{
+using namespace armcomputetensorutils;
+
+
+arm_compute::Status NeonBatchNormalizationValidate(const TensorInfo& input,
+                                                   const TensorInfo& output,
+                                                   const TensorInfo& mean,
+                                                   const TensorInfo& var,
+                                                   const TensorInfo& beta,
+                                                   const TensorInfo& gamma,
+                                                   const BatchNormalizationDescriptor& descriptor)
+{
+    const arm_compute::TensorInfo aclInputInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclOutputInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclMeanInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(mean, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclVarInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(var, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclBetaInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(beta, descriptor.m_DataLayout);
+    const arm_compute::TensorInfo aclGammaInfo =
+          armcomputetensorutils::BuildArmComputeTensorInfo(gamma, descriptor.m_DataLayout);
+
+    return arm_compute::NEBatchNormalizationLayer::validate(&aclInputInfo,
+                                                            &aclOutputInfo,
+                                                            &aclMeanInfo,
+                                                            &aclVarInfo,
+                                                            &aclBetaInfo,
+                                                            &aclGammaInfo,
+                                                            descriptor.m_Eps);
+}
+
+NeonBatchNormalizationWorkload::NeonBatchNormalizationWorkload(
+    const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info)
+    : BaseWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonBatchNormalizationWorkload", 1, 1);
+
+    arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
+    input.info()->set_data_layout(aclDataLayout);
+    output.info()->set_data_layout(aclDataLayout);
+
+    m_Mean = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo());
+
+    m_Variance = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo());
+
+    m_Gamma = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
+
+    m_Beta = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo());
+
+    m_Layer.configure(&input,
+                      &output,
+                      m_Mean.get(),
+                      m_Variance.get(),
+                      m_Beta.get(),
+                      m_Gamma.get(),
+                      m_Data.m_Parameters.m_Eps);
+
+    InitializeArmComputeTensorData(*m_Mean, m_Data.m_Mean);
+    InitializeArmComputeTensorData(*m_Variance, m_Data.m_Variance);
+    InitializeArmComputeTensorData(*m_Gamma, m_Data.m_Gamma);
+    InitializeArmComputeTensorData(*m_Beta, m_Data.m_Beta);
+
+    // Force Compute Library to perform the necessary copying and reshaping, after which
+    // delete all the input tensors that will no longer be needed
+    m_Layer.prepare();
+    FreeUnusedTensors();
+}
+
+void NeonBatchNormalizationWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonBatchNormalizationWorkload_Execute");
+    m_Layer.run();
+}
+
+void NeonBatchNormalizationWorkload::FreeUnusedTensors()
+{
+    FreeTensorIfUnused(m_Mean);
+    FreeTensorIfUnused(m_Variance);
+    FreeTensorIfUnused(m_Gamma);
+    FreeTensorIfUnused(m_Beta);
+}
+
+} //namespace armnn