From 4fcda0101ec3d110c1d6d7bee5c83416b645528a Mon Sep 17 00:00:00 2001 From: telsoa01 Date: Fri, 9 Mar 2018 14:13:49 +0000 Subject: Release 18.02 Change-Id: Id3c11dc5ee94ef664374a988fcc6901e9a232fa6 --- .../RefL2NormalizationFloat32Workload.cpp | 61 ++++++++++++++++++++++ 1 file changed, 61 insertions(+) create mode 100644 src/armnn/backends/RefWorkloads/RefL2NormalizationFloat32Workload.cpp (limited to 'src/armnn/backends/RefWorkloads/RefL2NormalizationFloat32Workload.cpp') diff --git a/src/armnn/backends/RefWorkloads/RefL2NormalizationFloat32Workload.cpp b/src/armnn/backends/RefWorkloads/RefL2NormalizationFloat32Workload.cpp new file mode 100644 index 0000000000..82c1ecd32e --- /dev/null +++ b/src/armnn/backends/RefWorkloads/RefL2NormalizationFloat32Workload.cpp @@ -0,0 +1,61 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// See LICENSE file in the project root for full license information. +// + +#include "RefL2NormalizationFloat32Workload.hpp" + +#include "RefWorkloadUtils.hpp" +#include "TensorBufferArrayView.hpp" + +#include "Profiling.hpp" + +#include + +namespace armnn +{ + +void RefL2NormalizationFloat32Workload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefL2NormalizationFloat32Workload_Execute"); + + const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]); + const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]); + + TensorBufferArrayView input(inputInfo.GetShape(), GetInputTensorDataFloat(0, m_Data)); + TensorBufferArrayView output(outputInfo.GetShape(), GetOutputTensorDataFloat(0, m_Data)); + + const unsigned int batchSize = inputInfo.GetShape()[0]; + const unsigned int depth = inputInfo.GetShape()[1]; + const unsigned int rows = inputInfo.GetShape()[2]; + const unsigned int cols = inputInfo.GetShape()[3]; + + for (unsigned int n = 0; n < batchSize; ++n) + { + for (unsigned int d = 0; d < depth; ++d) + { + for (unsigned int h = 0; h < rows; ++h) + { + for (unsigned int w = 0; w < cols; ++w) + { + float reduction = 0.0; + for (unsigned int c = 0; c < depth; ++c) + { + const float value = input.Get(n, c, h, w); + reduction += value * value; + } + + // Using std::max(reduction, epsilon) below would prevent against division by 0. + // However, at the time of writing: + // - This is not supported by the ACL functions used to implement L2Normalization in the CL + // backend. + // - The reference semantics for this operator do not include this parameter. + const float scale = 1.0f / sqrtf(reduction); + output.Get(n, d, h, w) = input.Get(n, d, h, w) * scale; + } + } + } + } +} + +} //namespace armnn -- cgit v1.2.1