From 3883b2776cec33f16f0ea9a2d795de2b7c766df7 Mon Sep 17 00:00:00 2001 From: Cathal Corbett Date: Fri, 22 Jul 2022 16:03:36 +0100 Subject: GitHub #667: Neon fold padding into average pool 2D quantization bug fix. * Originated from a GitHub issue: https://github.com/ARM-software/armnn/issues/667 * Initially, Arm NN supports the pool 2D operation because there is no padding on the pool2d. Neon failure occurs when padding is followed by average pool 2D due to folding optimization. * Here we prevent the folding optimization from happening for the above special case and add it in as a backend specific optimization. Signed-off-by: Cathal Corbett Change-Id: Ia0fd90c3a6b4b9d29c81106f154617d2e893e26b --- src/armnn/optimizations/FoldPadIntoLayer2d.hpp | 43 +++++--- .../FoldPadIntoQuantizedAveragePooling2DTests.cpp | 114 +++++++++++++++++++++ src/armnn/test/optimizations/FoldPadTests.cpp | 64 +++++++++++- 3 files changed, 206 insertions(+), 15 deletions(-) create mode 100644 src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp (limited to 'src/armnn') diff --git a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp b/src/armnn/optimizations/FoldPadIntoLayer2d.hpp index eb6bc90afd..4c4bd80d41 100644 --- a/src/armnn/optimizations/FoldPadIntoLayer2d.hpp +++ b/src/armnn/optimizations/FoldPadIntoLayer2d.hpp @@ -1,5 +1,5 @@ // -// Copyright © 2017 Arm Ltd. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -73,6 +73,17 @@ inline bool IsNeutralElement( : tensorValue == GetZeroElement(tensorInfo); } +inline bool IsPooling2dPadded(const Pooling2dDescriptor& poolDescriptor) +{ + const auto poolingPadValues = std::make_tuple(poolDescriptor.m_PadLeft, poolDescriptor.m_PadRight, + poolDescriptor.m_PadTop, poolDescriptor.m_PadBottom); + if (poolingPadValues != std::make_tuple(0U, 0U, 0U, 0U)) + { + return true; + } + return false; +} + template bool TryFoldPadIntoLayer2d( const PadDescriptor& padDescriptor, Descriptor& layerDescriptor, const TensorInfo& tensorInfo) @@ -101,25 +112,29 @@ bool TryFoldPadIntoLayer2d( return true; } -inline bool TryFoldPadIntoLayer2d( - const PadDescriptor& padDescriptor, Pooling2dDescriptor& poolDescriptor, const TensorInfo& tensorInfo) +inline bool TryFoldPadIntoLayer2d(const PadDescriptor& padDescriptor, + Pooling2dDescriptor& poolDescriptor, + const TensorInfo& tensorInfo, + bool isBackendOptimization = false) { - const auto poolingPadValues = std::make_tuple(poolDescriptor.m_PadLeft, poolDescriptor.m_PadRight, - poolDescriptor.m_PadTop, poolDescriptor.m_PadBottom); - bool poolHasPadding = false; - if (poolingPadValues != std::make_tuple(0U, 0U, 0U, 0U)) + // Cannot fold Average or L2 pooling if padding exists and the padding method is Exclude. + if (poolDescriptor.m_PoolType != PoolingAlgorithm::Max && + IsPooling2dPadded(poolDescriptor) && + poolDescriptor.m_PaddingMethod == PaddingMethod::Exclude) { - poolHasPadding = true; + return false; } - // We cannot fold Average or L2 pooling if there's is already padding and that padding method is Exclude. - if (poolDescriptor.m_PoolType != PoolingAlgorithm::Max) // PoolingAlgorithm::Average or PoolingAlgorithm::L2 + // Cannot fold Average pooling if data type is quantized and layout is NHWC in Neon backend. + // Therefore, this specific case will become a backend specific optimization. + if (!isBackendOptimization && + tensorInfo.IsQuantized() && + poolDescriptor.m_PoolType == PoolingAlgorithm::Average && + poolDescriptor.m_DataLayout == DataLayout::NHWC) { - if ((poolHasPadding) && (poolDescriptor.m_PaddingMethod == PaddingMethod::Exclude)) - { - return false; - } + return false; } + poolDescriptor.m_PaddingMethod = PaddingMethod::IgnoreValue; return TryFoldPadIntoLayer2d(padDescriptor, poolDescriptor, tensorInfo); diff --git a/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp b/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp new file mode 100644 index 0000000000..32627c62f7 --- /dev/null +++ b/src/armnn/test/optimizations/FoldPadIntoQuantizedAveragePooling2DTests.cpp @@ -0,0 +1,114 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include +#include + +#include + +#include + +using namespace armnn; + +namespace +{ +#if defined(ARMNNREF_ENABLED)||defined(ARMCOMPUTECL_ENABLED) +void FoldPadIntoQuantizedAvgPoolTest(Compute backendId) +{ + // Create a network + INetworkPtr network = INetwork::Create(); + + const unsigned int inputShape[] = {1, 2, 2, 3}; + const unsigned int paddedShape[] = {1, 4, 4, 3}; + const unsigned int outputShape[] = {1, 2, 2, 3}; + + TensorInfo inputInfo(4, inputShape, DataType::QAsymmU8, 1.0f, 0.0f); + TensorInfo paddedInfo(4, paddedShape, DataType::QAsymmU8, 1.0f, 0.0f); + TensorInfo outputInfo(4, outputShape, DataType::QAsymmU8, 1.0f, 0.0f); + + IConnectableLayer* input = network->AddInputLayer(0, "input"); + input->GetOutputSlot(0).SetTensorInfo(inputInfo); + + PadDescriptor padDescriptor({{0, 0}, + {1, 1}, + {1, 1}, + {0, 0}}); + + IConnectableLayer* padLayer = network->AddPadLayer(padDescriptor, "pad"); + padLayer->GetOutputSlot(0).SetTensorInfo(paddedInfo); + + Pooling2dDescriptor pooling2dDescriptor; + pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Average; + pooling2dDescriptor.m_PoolWidth = 3; + pooling2dDescriptor.m_PoolHeight = 3; + pooling2dDescriptor.m_StrideX = 1; + pooling2dDescriptor.m_StrideY = 1; + pooling2dDescriptor.m_DataLayout = DataLayout::NHWC; + + IConnectableLayer* pool2dLayer = network->AddPooling2dLayer(pooling2dDescriptor, "pool2d"); + pool2dLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); + + IConnectableLayer* output = network->AddOutputLayer(0, "output"); + + // Connect up layers - input -> pad -> pool2d -> output + input->GetOutputSlot(0).Connect(padLayer->GetInputSlot(0)); + padLayer->GetOutputSlot(0).Connect(pool2dLayer->GetInputSlot(0)); + pool2dLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + // Create ArmNN runtime + IRuntimePtr run = IRuntime::Create(IRuntime::CreationOptions()); + + // Optimise ArmNN network + IOptimizedNetworkPtr optNet = Optimize(*network, {backendId}, run->GetDeviceSpec()); + + auto checkPadFoldedIntoPool2d = [&](const Layer* const layer) { + if (!IsLayerOfType(layer) || (layer->GetNameStr() != "folded-pad-into-pool2d")) + { + return false; + } + + const auto pool2dLayer = static_cast(layer); + const Pooling2dDescriptor pool2dLayerParams = pool2dLayer->GetParameters(); + + Pooling2dDescriptor pool2dLayerParamsNoPad = pool2dLayerParams; + pool2dLayerParamsNoPad.m_PadLeft = 0; + pool2dLayerParamsNoPad.m_PadRight = 0; + pool2dLayerParamsNoPad.m_PadTop = 0; + pool2dLayerParamsNoPad.m_PadBottom = 0; + // If we fold then PaddingMethod will be set to Ignore. The original will be Exclude. + pool2dLayerParamsNoPad.m_PaddingMethod = PaddingMethod::Exclude; + + return (pool2dLayerParamsNoPad == pooling2dDescriptor) && (pool2dLayerParams.m_PadLeft == 1) && + (pool2dLayerParams.m_PadRight == 1) && (pool2dLayerParams.m_PadTop == 1) && + (pool2dLayerParams.m_PadBottom == 1) && (pool2dLayerParams.m_PaddingMethod == PaddingMethod::IgnoreValue); + }; + + Graph& graph = GetGraphForTesting(optNet.get()); + CHECK(CheckSequence(graph.cbegin(), graph.cend(), + &IsLayerOfType, + checkPadFoldedIntoPool2d, + &IsLayerOfType)); +} +#endif +} + + +TEST_SUITE("Optimizer_FoldPadIntoQuantizedAvgPoolCpuRef") +{ +TEST_CASE("FoldPadIntoQuantizedAvgPoolCpuRefTest") +{ + FoldPadIntoQuantizedAvgPoolTest(Compute::CpuRef); +} +} + +#if defined(ARMCOMPUTECL_ENABLED) +TEST_SUITE("Optimizer_FoldPadIntoQuantizedAvgPoolGpuAcc") +{ +TEST_CASE("FoldPadIntoQuantizedAvgPoolGpuAccTest") +{ + FoldPadIntoQuantizedAvgPoolTest(Compute::GpuAcc); +} +} +#endif diff --git a/src/armnn/test/optimizations/FoldPadTests.cpp b/src/armnn/test/optimizations/FoldPadTests.cpp index 4d7defcabe..b2672ea584 100644 --- a/src/armnn/test/optimizations/FoldPadTests.cpp +++ b/src/armnn/test/optimizations/FoldPadTests.cpp @@ -1,5 +1,5 @@ // -// Copyright © 2021 Arm Ltd and Contributors. All rights reserved. +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. // SPDX-License-Identifier: MIT // @@ -474,6 +474,68 @@ TEST_CASE("FoldPadLayerIntoPooling2dLayer_MaxPoolingLayerWithLargePadValueShould &IsLayerOfType)); } +TEST_CASE("FoldPadLayerIntoPooling2dLayer_QuantizedAveragePoolingShouldNotBeFolded") +{ + Graph graph; + const unsigned int inputShape[] = {1, 2, 2, 3}; + const unsigned int paddedShape[] = {1, 4, 4, 3}; + const unsigned int outputShape[] = {1, 2, 2, 3}; + + TensorInfo inputInfo(4, inputShape, DataType::QAsymmU8); + TensorInfo paddedInfo(4, paddedShape, DataType::QAsymmU8); + TensorInfo outputInfo(4, outputShape, DataType::QAsymmU8); + + Layer* input = graph.AddLayer(0, "input"); + input->GetOutputSlot().SetTensorInfo(inputInfo); + + PadDescriptor padDescriptor({{0, 0}, + {1, 1}, + {1, 1}, + {0, 0}}); + + PadLayer* padLayer = graph.AddLayer(padDescriptor, "pad"); + padLayer->GetOutputSlot().SetTensorInfo(paddedInfo); + + Pooling2dDescriptor pooling2dDescriptor; + pooling2dDescriptor.m_PoolType = PoolingAlgorithm::Average; + pooling2dDescriptor.m_PoolWidth = 3; + pooling2dDescriptor.m_PoolHeight = 3; + pooling2dDescriptor.m_StrideX = 1; + pooling2dDescriptor.m_StrideY = 1; + pooling2dDescriptor.m_DataLayout = DataLayout::NHWC; + + Pooling2dLayer* pool2dLayer = graph.AddLayer(pooling2dDescriptor, "pool2d"); + pool2dLayer->GetOutputSlot().SetTensorInfo(outputInfo); + + Layer* output = graph.AddLayer(0, "output"); + + // Connect up layers - input -> pad -> pool2d -> output + input->GetOutputSlot().Connect(padLayer->GetInputSlot(0)); + padLayer->GetOutputSlot().Connect(pool2dLayer->GetInputSlot(0)); + pool2dLayer->GetOutputSlot().Connect(output->GetInputSlot(0)); + + auto checkSimplePool2d = [&](const Layer* const layer) { + const auto pool2dLayer = static_cast(layer); + return IsLayerOfType(layer) && (layer->GetNameStr() == "pool2d") && + (pool2dLayer->GetParameters() == pooling2dDescriptor); + }; + + CHECK(CheckSequence(graph.cbegin(), graph.cend(), + &IsLayerOfType, + &IsLayerOfType, + checkSimplePool2d, + &IsLayerOfType)); + + armnn::Optimizer::Pass(graph, MakeOptimizations(FoldPadIntoPooling2d())); + + // The optimization should not have modified the graph. + CHECK(CheckSequence(graph.cbegin(), graph.cend(), + &IsLayerOfType, + &IsLayerOfType, + checkSimplePool2d, + &IsLayerOfType)); +} + #if defined(ARMNNREF_ENABLED) TEST_CASE("FoldPadLayerIntoPooling2dLayer_ExecuteInferenceWithAndWithoutOptimization") { -- cgit v1.2.1