From f03fcf0dd180ba2c87648a524fcca9214e1f979b Mon Sep 17 00:00:00 2001 From: Aron Virginas-Tar Date: Tue, 9 Jul 2019 17:44:24 +0100 Subject: IVGCVSW-3456 Add support for dynamic output shape in ConvertPrelu Signed-off-by: Aron Virginas-Tar Change-Id: I8fc7a716455be3f51b51177f6896a73790a41fc3 --- 1.2/HalPolicy.cpp | 17 +++++++++++++++-- Android.mk | 1 + ConversionUtils.hpp | 24 ++++++++++++++++++++---- OutputShapeUtils.cpp | 43 +++++++++++++++++++++++++++++++++++++++++++ OutputShapeUtils.hpp | 17 +++++++++++++++++ 5 files changed, 96 insertions(+), 6 deletions(-) create mode 100644 OutputShapeUtils.cpp create mode 100644 OutputShapeUtils.hpp diff --git a/1.2/HalPolicy.cpp b/1.2/HalPolicy.cpp index fe0cfbdc..b194a57a 100644 --- a/1.2/HalPolicy.cpp +++ b/1.2/HalPolicy.cpp @@ -5,6 +5,8 @@ #include "HalPolicy.hpp" +#include "OutputShapeUtils.hpp" + #include "../1.0/HalPolicy.hpp" #include "../1.1/HalPolicy.hpp" @@ -539,7 +541,13 @@ bool HalPolicy::ConvertPrelu(const Operation& operation, const Model& model, Con const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); const armnn::TensorInfo& alphaInfo = alpha.GetTensorInfo(); - const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); + + armnn::TensorInfo outputInfo = GetTensorInfoForOperand(*output); + if (outputInfo.GetNumElements() == 0u) + { + ALOGD("Output shape not set, will infer from inputs"); + outputInfo.SetShape(InferPreluOutputShape(inputInfo.GetShape(), alphaInfo.GetShape())); + } if (!IsLayerSupportedForAnyBackend(__func__, armnn::IsPreluSupported, @@ -560,7 +568,12 @@ bool HalPolicy::ConvertPrelu(const Operation& operation, const Model& model, Con BroadcastTensor(input, alpha, layer, *data.m_Network); - return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data); + return SetupAndTrackLayerOutputSlot(operation, + 0, + *layer, + model, + data, + armnn::Optional(outputInfo)); } bool HalPolicy::ConvertResize(const Operation& operation, diff --git a/Android.mk b/Android.mk index bee57dd0..215b0a84 100644 --- a/Android.mk +++ b/Android.mk @@ -334,6 +334,7 @@ LOCAL_SRC_FILES := \ ConversionUtils.cpp \ DriverOptions.cpp \ ModelToINetworkConverter.cpp \ + OutputShapeUtils.cpp \ RequestThread.cpp \ Utils.cpp diff --git a/ConversionUtils.hpp b/ConversionUtils.hpp index d30b8a4e..c9be0003 100644 --- a/ConversionUtils.hpp +++ b/ConversionUtils.hpp @@ -1028,7 +1028,8 @@ bool SetupAndTrackLayerOutputSlot(const HalOperation& operation, armnn::IConnectableLayer& layer, uint32_t layerOutputIndex, const HalModel& model, - ConversionData& data) + ConversionData& data, + const armnn::Optional& outputInfo = armnn::EmptyOptional()) { using HalOperand = typename HalPolicy::Operand; @@ -1043,7 +1044,15 @@ bool SetupAndTrackLayerOutputSlot(const HalOperation& operation, const uint32_t operandIndex = operation.outputs[operationOutputIndex]; data.m_OutputSlotForOperand[operandIndex] = &outputSlot; - outputSlot.SetTensorInfo(GetTensorInfoForOperand(*outputOperand)); + if (outputInfo.has_value()) + { + outputSlot.SetTensorInfo(outputInfo.value()); + ALOGD("Output info overwritten"); + } + else + { + outputSlot.SetTensorInfo(GetTensorInfoForOperand(*outputOperand)); + } return true; } @@ -1092,9 +1101,16 @@ bool SetupAndTrackLayerOutputSlot(const HalOperation& operation, uint32_t outputIndex, armnn::IConnectableLayer& layer, const HalModel& model, - ConversionData& data) + ConversionData& data, + const armnn::Optional& outputInfo = armnn::EmptyOptional()) { - return SetupAndTrackLayerOutputSlot(operation, outputIndex, layer, outputIndex, model, data); + return SetupAndTrackLayerOutputSlot(operation, + outputIndex, + layer, + outputIndex, + model, + data, + outputInfo); } template + +namespace armnn_driver +{ + +using namespace armnn; + +TensorShape InferPreluOutputShape(const TensorShape& inputShape, const TensorShape& alphaShape) +{ + // NOTE: The inferred PReLU output size will be the maximum size along each dimension + // of input and alpha, starting with the trailing dimensions, and working its way forward. + // + // Example: inputShape={4, 1, 2}, alphaShape={5, 4, 3, 1} => outputShape={5, 4, 3, 2} + + const unsigned int numInputDims = inputShape.GetNumDimensions(); + const unsigned int numAlphaDims = alphaShape.GetNumDimensions(); + + const unsigned int maxNumDims = std::max(numInputDims, numAlphaDims); + + TensorShape outputShape = TensorShape(maxNumDims); + for (unsigned int reverseIdx = 1u; reverseIdx <= maxNumDims; ++reverseIdx) + { + const int inputIdx = numInputDims - reverseIdx; + const int alphaIdx = numAlphaDims - reverseIdx; + + const unsigned int inputDimSize = inputIdx >= 0 ? inputShape[inputIdx] : 0u; + const unsigned int alphaDimSize = alphaIdx >= 0 ? alphaShape[alphaIdx] : 0u; + + const unsigned int outputIdx = maxNumDims - reverseIdx; + outputShape[outputIdx] = std::max(inputDimSize, alphaDimSize); + } + + return outputShape; +} + +} // namespace armnn_driver \ No newline at end of file diff --git a/OutputShapeUtils.hpp b/OutputShapeUtils.hpp new file mode 100644 index 00000000..f314252f --- /dev/null +++ b/OutputShapeUtils.hpp @@ -0,0 +1,17 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +namespace armnn_driver +{ + +armnn::TensorShape InferPreluOutputShape(const armnn::TensorShape& inputShape, const armnn::TensorShape& alphaShape); + +} // namespace armnn_driver + + -- cgit v1.2.1