1 files changed, 63 insertions, 12 deletions

diff --git a/src/armnn/backends/NeonWorkloads/NeonBatchNormalizationFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonBatchNormalizationFloat32Workload.cpp
index f107c8137f..ca5c8202cd 100644
--- a/src/armnn/backends/NeonWorkloads/NeonBatchNormalizationFloat32Workload.cpp
+++ b/src/armnn/backends/NeonWorkloads/NeonBatchNormalizationFloat32Workload.cpp
@@ -6,40 +6,91 @@
 #include "NeonBatchNormalizationFloat32Workload.hpp"
 #include "backends/CpuTensorHandle.hpp"
 #include "backends/ArmComputeTensorUtils.hpp"
+#include "../../../../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 = BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
+    const arm_compute::TensorInfo aclMeanInfo = BuildArmComputeTensorInfo(mean);
+    const arm_compute::TensorInfo aclVarInfo = BuildArmComputeTensorInfo(var);
+    const arm_compute::TensorInfo aclBetaInfo = BuildArmComputeTensorInfo(beta);
+    const arm_compute::TensorInfo aclGammaInfo = BuildArmComputeTensorInfo(gamma);
+
+    return arm_compute::NEBatchNormalizationLayer::validate(&aclInputInfo,
+                                                            &aclOutputInfo,
+                                                            &aclMeanInfo,
+                                                            &aclVarInfo,
+                                                            &aclBetaInfo,
+                                                            &aclGammaInfo,
+                                                            descriptor.m_Eps);
+}
+
 NeonBatchNormalizationFloat32Workload::NeonBatchNormalizationFloat32Workload(
     const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info)
-    : Float32Workload<BatchNormalizationQueueDescriptor>(descriptor, info)
+    : FloatWorkload<BatchNormalizationQueueDescriptor>(descriptor, info)
 {
     m_Data.ValidateInputsOutputs("NeonBatchNormalizationFloat32Workload", 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();
 
-    BuildArmComputeTensor(m_Mean, m_Data.m_Mean->GetTensorInfo());
-    BuildArmComputeTensor(m_Variance, m_Data.m_Variance->GetTensorInfo());
-    BuildArmComputeTensor(m_Gamma, m_Data.m_Gamma->GetTensorInfo());
-    BuildArmComputeTensor(m_Beta, m_Data.m_Beta->GetTensorInfo());
+    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_Layer.configure(
-        &input, &output, &m_Mean, &m_Variance, &m_Beta, &m_Gamma, m_Data.m_Parameters.m_Eps);
+    m_Gamma = std::make_unique<arm_compute::Tensor>();
+    BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo());
 
-    InitialiseArmComputeTensorData(m_Mean, m_Data.m_Mean->GetConstTensor<float>());
-    InitialiseArmComputeTensorData(m_Variance, m_Data.m_Variance->GetConstTensor<float>());
-    InitialiseArmComputeTensorData(m_Gamma, m_Data.m_Gamma->GetConstTensor<float>());
-    InitialiseArmComputeTensorData(m_Beta, m_Data.m_Beta->GetConstTensor<float>());
+    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);
+
+    InitializeArmComputeTensorDataForFloatTypes(*m_Mean, m_Data.m_Mean);
+    InitializeArmComputeTensorDataForFloatTypes(*m_Variance, m_Data.m_Variance);
+    InitializeArmComputeTensorDataForFloatTypes(*m_Gamma, m_Data.m_Gamma);
+    InitializeArmComputeTensorDataForFloatTypes(*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 NeonBatchNormalizationFloat32Workload::Execute() const
 {
-    ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "NeonBatchNormalizationFloat32Workload_Execute");
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonBatchNormalizationFloat32Workload_Execute");
     m_Layer.run();
 }
 
+void NeonBatchNormalizationFloat32Workload::FreeUnusedTensors()
+{
+    FreeTensorIfUnused(m_Mean);
+    FreeTensorIfUnused(m_Variance);
+    FreeTensorIfUnused(m_Gamma);
+    FreeTensorIfUnused(m_Beta);
+}
+
 } //namespace armnn