From 3a7d3a70d99cbe22f5e4711d5dbbea2a245da7ed Mon Sep 17 00:00:00 2001 From: Rob Hughes Date: Tue, 24 Sep 2019 16:59:56 +0100 Subject: NNXSW-1826 Add an optimization step which combines Permute and BatchToSpace into DepthToSpace This is only possible in some limited cases, but removes an extra layer from the graph and so should improve performance in all cases. Change-Id: I7b3e6ba5dacb4fdb816ad270edaecda1436ab4cf Signed-off-by: Rob Hughes --- src/armnn/optimizations/All.hpp | 1 + .../PermuteAndBatchToSpaceAsDepthToSpace.cpp | 87 ++++++++++++++++++++++ .../PermuteAndBatchToSpaceAsDepthToSpace.hpp | 27 +++++++ 3 files changed, 115 insertions(+) create mode 100644 src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.cpp create mode 100644 src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.hpp (limited to 'src/armnn/optimizations') diff --git a/src/armnn/optimizations/All.hpp b/src/armnn/optimizations/All.hpp index 68965fd23c..4ea3f7f2d4 100644 --- a/src/armnn/optimizations/All.hpp +++ b/src/armnn/optimizations/All.hpp @@ -14,3 +14,4 @@ #include "ConvertFp32NetworkToFp16.hpp" #include "AddDebug.hpp" #include "FoldPadIntoConvolution2d.hpp" +#include "PermuteAndBatchToSpaceAsDepthToSpace.hpp" \ No newline at end of file diff --git a/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.cpp b/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.cpp new file mode 100644 index 0000000000..c42162b6c1 --- /dev/null +++ b/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.cpp @@ -0,0 +1,87 @@ +// +// Copyright © 2019 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "PermuteAndBatchToSpaceAsDepthToSpace.hpp" + +using namespace armnn; +using namespace armnn::optimizations; + +void PermuteAndBatchToSpaceAsDepthToSpaceImpl::Run(Graph& graph, InputSlot& connection) const +{ + // Validate base layer (the Permute) is compatible + Layer& base = connection.GetConnectedOutputSlot()->GetOwningLayer(); + BOOST_ASSERT(base.GetType() == LayerType::Permute); + const TensorInfo& inputInfo = base.GetInputSlot(0).GetConnection()->GetTensorInfo(); + const TensorInfo& intermediateInfo = base.GetOutputSlot(0).GetTensorInfo(); + if (intermediateInfo.GetNumDimensions() != 4) + { + // Must be 4D, otherwise the below checks do not make sense + return; + } + if (!static_cast(base).GetParameters().m_DimMappings.IsEqual(PermutationVector{ 3, 1, 2, 0 })) + { + // Must swap batch and channels dimensions, otherwise it is not the (original) channels dimension + // that is being decomposed. + return; + } + + // Validate child layer (the BatchToSpace) is compatible + Layer& child = connection.GetOwningLayer(); + BOOST_ASSERT(child.GetType() == LayerType::BatchToSpaceNd); + const TensorInfo& outputInfo = child.GetOutputSlot(0).GetTensorInfo(); + const BatchToSpaceNdDescriptor& batchToSpaceDesc = static_cast(child).GetParameters(); + if (batchToSpaceDesc.m_DataLayout != DataLayout::NHWC) + { + // The rest of this function assumes NHWC, although in future this restriction could be lifted. + return; + } + if (batchToSpaceDesc.m_Crops != std::vector>{ { 0, 0 }, { 0, 0 } }) + { + // Cropping is not supported in DepthToSpace + return; + } + if (batchToSpaceDesc.m_BlockShape.size() != 2 || + batchToSpaceDesc.m_BlockShape[0] != batchToSpaceDesc.m_BlockShape[1]) + { + // Asymmetric or non-2D block sizes are not supported by DepthToSpace + return; + } + uint32_t blockSize = batchToSpaceDesc.m_BlockShape[0]; + if (outputInfo.GetShape()[0] != 1 || outputInfo.GetShape()[3] != 1) + { + // The final output must have 1 batch and 1 channel because these dimensions will be swapped around + // once we make the substitution, and it needs to be equivalent. + return; + } + + // Validate the intermediate tensor quantization params. + // These must be identical to either the input or output quantization params, otherwise the intermediate tensor + // may not have sufficient range/precision to preserve the values. + // This would mean that once we perform the substitution this loss of precision will no longer occur, + // so we would have changed the meaning of the network. + bool isIntermediateQuantParamsSameAsInput = + intermediateInfo.GetQuantizationScale() == inputInfo.GetQuantizationScale() && + intermediateInfo.GetQuantizationOffset() == inputInfo.GetQuantizationOffset(); + bool isIntermediateQuantParamsSameAsOutput = + intermediateInfo.GetQuantizationScale() == outputInfo.GetQuantizationScale() && + intermediateInfo.GetQuantizationOffset() == outputInfo.GetQuantizationOffset(); + if (!isIntermediateQuantParamsSameAsInput && !isIntermediateQuantParamsSameAsOutput) + { + return; + } + + // Insert equivalent DepthToSpace layer + const std::string name = std::string("merged-") + base.GetName() + std::string("-with-") + child.GetName(); + + // Inserts equivalent reshape before base layer. + const DepthToSpaceDescriptor depthToSpaceDesc(blockSize, DataLayout::NHWC); + auto& depthToSpace = *graph.InsertNewLayer(base.GetInputSlot(0), depthToSpaceDesc, name.c_str()); + depthToSpace.GetOutputHandler().SetTensorInfo(outputInfo); + + // Moves connections from child output to new layer. + // Child layer will be removed as it's left unconnected. + // Base layer will be removed if left unconnected. + child.GetOutputSlot().MoveAllConnections(depthToSpace.GetOutputSlot()); +} diff --git a/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.hpp b/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.hpp new file mode 100644 index 0000000000..4a73efca40 --- /dev/null +++ b/src/armnn/optimizations/PermuteAndBatchToSpaceAsDepthToSpace.hpp @@ -0,0 +1,27 @@ +// +// Copyright © 2019 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// +#pragma once + +#include "Optimization.hpp" + +namespace armnn +{ +namespace optimizations +{ + +/// Replaces Permute leading into BatchToSpace with a DepthToSpace +/// in the case where the Permute swaps the batch and channels dimensions +/// such that the replacement is valid. +class PermuteAndBatchToSpaceAsDepthToSpaceImpl +{ +public: + void Run(Graph& graph, InputSlot& connection) const; +}; + +using PermuteAndBatchToSpaceAsDepthToSpace = + OptimizeForConnection; + +} // namespace optimizations +} // namespace armnn -- cgit v1.2.1