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 --- .../RefNormalizationFloat32Workload.cpp | 185 +++++++++++++++++++++ 1 file changed, 185 insertions(+) create mode 100644 src/armnn/backends/RefWorkloads/RefNormalizationFloat32Workload.cpp (limited to 'src/armnn/backends/RefWorkloads/RefNormalizationFloat32Workload.cpp') diff --git a/src/armnn/backends/RefWorkloads/RefNormalizationFloat32Workload.cpp b/src/armnn/backends/RefWorkloads/RefNormalizationFloat32Workload.cpp new file mode 100644 index 0000000000..c743207423 --- /dev/null +++ b/src/armnn/backends/RefWorkloads/RefNormalizationFloat32Workload.cpp @@ -0,0 +1,185 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// See LICENSE file in the project root for full license information. +// + +#include "RefNormalizationFloat32Workload.hpp" + +#include "RefWorkloadUtils.hpp" + +#include "Profiling.hpp" + +#include + +#include +#include + +namespace armnn +{ + +// Helper function to compute "Within" normalization using Krichevsky 2012: Local Brightness Normalization +static void NormalizeWithinUingLbr(const float* inputData, + float* outputData, + const TensorShape& tensorShape, + uint32_t norm_size, + float alpha, + float beta, + float kappa) +{ + const unsigned int batchSize = tensorShape[0]; + const unsigned int depth = tensorShape[1]; + const unsigned int rows = tensorShape[2]; + const unsigned int cols = tensorShape[3]; + + int radius = boost::numeric_cast(norm_size / 2u); /* Strong Assumption on rounding Mode */ + + for (unsigned int n = 0; n < batchSize; n++) + { + for (unsigned int c = 0; c < depth; c++) + { + for (unsigned int h = 0; h < rows; h++) + { + for (unsigned int w = 0; w < cols; w++) + { + float accumulated_scale = 0.0; + for (int y = -radius; y <= radius; y++) + { + for (int x = -radius; x <= radius; x++) + { + int i = boost::numeric_cast(w) + x; + int j = boost::numeric_cast(h) + y; + + if ((i < 0) || (i >= boost::numeric_cast(cols))) + { + continue; + } + + if ((j < 0) || (j >= boost::numeric_cast(rows))) + { + continue; + } + + float inval = inputData[n * cols * rows * depth + + c * cols * rows + + boost::numeric_cast(j) * cols + + boost::numeric_cast(i)]; + + accumulated_scale += inval*inval; + } + } + outputData[n * cols * rows * depth + + c * cols * rows + + h * cols + + w] = inputData[n * cols * rows * depth + + c * cols * rows + + h * cols + + w] / (powf((kappa + (accumulated_scale * alpha)), beta)); + } + } + } + } +} + +// Helper function to compute "Across" normalization using Krichevsky 2012: Local Brightness Normalization +void NormalizeAcrossUingLbr(const float* inputData, + float* outputData, + const TensorShape& tensorShape, + uint32_t norm_size, + float alpha, + float beta, + float kappa) +{ + const unsigned int batchSize = tensorShape[0]; + const unsigned int depth = tensorShape[1]; + const unsigned int rows = tensorShape[2]; + const unsigned int cols = tensorShape[3]; + + int radius = boost::numeric_cast(norm_size / 2u); /* Strong Assumption on rounding Mode */ + + for (unsigned int n = 0; n < batchSize; n++) + { + for (unsigned int c = 0; c < depth; c++) + { + for (unsigned int h = 0; h < rows; h++) + { + for (unsigned int w = 0; w < cols; w++) + { + float accumulated_scale = 0.0; + for (int z = -radius; z <= radius; z++) + { + int k = boost::numeric_cast(c) + z; + + if ((k < 0) || (k >= boost::numeric_cast(depth))) + { + continue; + } + + float inval = inputData[n * cols * rows * depth + + boost::numeric_cast(k) * cols * rows + + h * cols + + w]; + + accumulated_scale += inval*inval; + } + float scale = kappa + (accumulated_scale * alpha); + scale = powf(scale, -beta); + outputData[n * cols * rows * depth + + c * cols * rows + + h * cols + + w] = scale * + inputData[n * cols * rows * depth + + c * cols * rows + + h * cols + + w]; + } + } + } + } +} + +void RefNormalizationFloat32Workload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefNormalizationFloat32Workload_Execute"); + + const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]); + + float* outputData = GetOutputTensorDataFloat(0, m_Data); + const float* inputData = GetInputTensorDataFloat(0, m_Data); + + + if (NormalizationAlgorithmMethod::LocalBrightness == m_Data.m_Parameters.m_NormMethodType) + { + if (NormalizationAlgorithmChannel::Within == m_Data.m_Parameters.m_NormChannelType) + { + NormalizeWithinUingLbr(inputData, + outputData, + inputInfo.GetShape(), + m_Data.m_Parameters.m_NormSize, + m_Data.m_Parameters.m_Alpha, + m_Data.m_Parameters.m_Beta, + m_Data.m_Parameters.m_K); + } + else if (NormalizationAlgorithmChannel::Across == m_Data.m_Parameters.m_NormChannelType) + { + NormalizeAcrossUingLbr(inputData, + outputData, + inputInfo.GetShape(), + m_Data.m_Parameters.m_NormSize, + m_Data.m_Parameters.m_Alpha, + m_Data.m_Parameters.m_Beta, + m_Data.m_Parameters.m_K); + } + else + { + BOOST_LOG_TRIVIAL(warning) << "Illegal NORMALIZATION mode in normalization_f32"; + return; + } + } + else + { + BOOST_LOG_TRIVIAL(warning) << "Lcr method (Jarret 2009: Local Contrast Normalization) not supported yet."; + return; + } +} + +} //namespace armnn -- cgit v1.2.1