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 <cathal.corbett@arm.com>
Change-Id: Ia0fd90c3a6b4b9d29c81106f154617d2e893e26b
diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
index a26442c..766bf2d 100644
--- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
@@ -1,5 +1,5 @@
//
-// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
@@ -9,6 +9,7 @@
#include <armnn/utility/Assert.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
+#include <backendsCommon/SubgraphUtils.hpp>
namespace armnn
{
@@ -20,36 +21,6 @@
// this helper only works if all layers where the inputs connect to are not selected
//
-SubgraphView::IInputSlots CreateIInputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
-{
- SubgraphView::IInputSlots result;
- for (auto&& layer : layers)
- {
- for (unsigned int i = 0 ; i < layer->GetNumInputSlots(); ++i)
- {
- result.push_back(&(layer->GetInputSlot(i)));
- }
- }
- return result;
-}
-
-//
-// this helper only works if all layers where the outputs connect to are not selected
-//
-
-SubgraphView::IOutputSlots CreateIOutputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
-{
- SubgraphView::IOutputSlots result;
- for (auto &&layer: layers)
- {
- for (unsigned int i = 0; i < layer->GetNumOutputSlots(); ++i)
- {
- result.push_back(&(layer->GetOutputSlot(i)));
- }
- }
- return result;
-}
-
bool checkDataTypeInputandOutput(const Layer& layer)
{
auto inputInfo = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
@@ -79,19 +50,6 @@
} // namespace
-inline void ReportUntouchedLayers(OptimizationViews& optimizationViews, std::map<LayerGuid, Layer*> untouched)
-{
- std::vector<Layer*> untouchedVector;
- for (const auto& pair : untouched)
- {
- Layer* layer = pair.second;
- SubgraphView subgraphView({layer},
- CreateIInputsFrom({layer}),
- CreateIOutputsFrom({layer}));
- optimizationViews.AddUntouchedSubgraph(std::move(subgraphView));
- }
-}
-
template<typename LayerType>
LayerType* FuseLayer(OptimizationViews& optimizationViews,
LayerType* baseLayer,
diff --git a/src/backends/backendsCommon/SubgraphUtils.hpp b/src/backends/backendsCommon/SubgraphUtils.hpp
new file mode 100644
index 0000000..bd3d698
--- /dev/null
+++ b/src/backends/backendsCommon/SubgraphUtils.hpp
@@ -0,0 +1,99 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <optimizations/FoldPadIntoLayer2d.hpp>
+
+namespace armnn
+{
+
+namespace
+{
+
+//
+// this helper only works if all layers where the inputs connect to are not selected
+//
+
+SubgraphView::IInputSlots CreateIInputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
+{
+ SubgraphView::IInputSlots result;
+ for (auto&& layer : layers)
+ {
+ for (unsigned int i = 0 ; i < layer->GetNumInputSlots(); ++i)
+ {
+ result.push_back(&(layer->GetInputSlot(i)));
+ }
+ }
+ return result;
+}
+
+//
+// this helper only works if all layers where the outputs connect to are not selected
+//
+
+SubgraphView::IOutputSlots CreateIOutputsFrom(const std::vector<armnn::IConnectableLayer*>& layers)
+{
+ SubgraphView::IOutputSlots result;
+ for (auto &&layer: layers)
+ {
+ for (unsigned int i = 0; i < layer->GetNumOutputSlots(); ++i)
+ {
+ result.push_back(&(layer->GetOutputSlot(i)));
+ }
+ }
+ return result;
+}
+
+}
+
+inline void ReportUntouchedLayers(OptimizationViews& optimizationViews, std::map<LayerGuid, Layer*> untouched)
+{
+ std::vector<Layer*> untouchedVector;
+ for (const auto& pair : untouched)
+ {
+ Layer* layer = pair.second;
+ SubgraphView subgraphView({layer},
+ CreateIInputsFrom({layer}),
+ CreateIOutputsFrom({layer}));
+ optimizationViews.AddUntouchedSubgraph(std::move(subgraphView));
+ }
+}
+
+template<typename LayerType>
+LayerType* FoldPadLayer(OptimizationViews& optimizationViews,
+ LayerType* baseLayer,
+ LayerType* replacementLayer,
+ PadLayer* padLayer)
+{
+ SubgraphView substitutionSubgraph({padLayer, baseLayer},
+ CreateIInputsFrom({padLayer}),
+ CreateIOutputsFrom({baseLayer}));
+ SubgraphView replacementSubgraph(replacementLayer);
+
+ optimizationViews.AddSubstitution({substitutionSubgraph, replacementSubgraph});
+
+ return replacementLayer;
+}
+
+template<typename LayerType>
+LayerType* FoldPadIntoAveragePool2d(OptimizationViews& optimizationViews,
+ Pooling2dLayer* baseLayer,
+ Pooling2dDescriptor& poolDescriptor,
+ PadLayer* padLayer)
+{
+ IConnectableLayer* replacement =
+ optimizationViews.GetINetwork()->AddPooling2dLayer(poolDescriptor, "folded-pad-into-pool2d");
+ LayerType* replacementLayer = PolymorphicDowncast<LayerType*>(replacement);
+
+ FoldPadLayer(optimizationViews,
+ baseLayer,
+ replacementLayer,
+ padLayer);
+
+ return replacementLayer;
+}
+
+} // namespace armnn
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index 1fe53de..d2e8fbf 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -1,5 +1,5 @@
//
-// Copyright © 2017 Arm Ltd. All rights reserved.
+// Copyright © 2022 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
@@ -566,6 +566,31 @@
untouched.erase(baseLayer->GetGuid());
}
}
+
+ // Special case to fuse padding into average pooling 2d for quantized datatype.
+ // Required to be done as a backend specific optimization as Neon does not support this special case.
+ if (base.GetType() == LayerType::Pooling2d)
+ {
+ Pooling2dLayer* baseLayer = PolymorphicDowncast<Pooling2dLayer*>(&base);
+ Pooling2dDescriptor poolingDescriptor = baseLayer->GetParameters();
+
+ if (baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer().GetType() == LayerType::Pad)
+ {
+ PadLayer* padLayer = PolymorphicDowncast<PadLayer*>(
+ &baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer());
+ if (padLayer->GetOutputSlot(0).GetNumConnections() == 1 &&
+ optimizations::pad_fold::TryFoldPadIntoLayer2d(padLayer->GetParameters(),
+ poolingDescriptor,
+ padLayer->GetOutputSlot().GetTensorInfo(),
+ true))
+ {
+ FoldPadIntoAveragePool2d<Pooling2dLayer>(optimizationViews, baseLayer,
+ poolingDescriptor, padLayer);
+ untouched.erase(baseLayer->GetGuid());
+ untouched.erase(padLayer->GetGuid());
+ }
+ }
+ }
}
if (optimizationViews.GetSubstitutions().empty())
diff --git a/src/backends/reference/RefBackend.cpp b/src/backends/reference/RefBackend.cpp
index a33a775..8c8879c 100644
--- a/src/backends/reference/RefBackend.cpp
+++ b/src/backends/reference/RefBackend.cpp
@@ -14,6 +14,7 @@
#include <armnn/backends/IMemoryManager.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <backendsCommon/DefaultAllocator.hpp>
+#include <backendsCommon/SubgraphUtils.hpp>
#include <Optimizer.hpp>
@@ -70,11 +71,61 @@
return layerSupport;
}
-OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph) const
+OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph,
+ const ModelOptions& modelOptions) const
{
- OptimizationViews optimizationViews;
+ OptimizationViews optimizationViews(modelOptions);
- optimizationViews.AddUntouchedSubgraph(SubgraphView(subgraph));
+ auto it = subgraph.endIConnectable();
+ std::map<LayerGuid, Layer*> untouched;
+
+ while (it != subgraph.beginIConnectable())
+ {
+ --it;
+ Layer& base = *(PolymorphicDowncast<Layer*>(*it));
+ untouched.insert({base.GetGuid(), &base});
+ }
+
+ it = subgraph.endIConnectable();
+ while (it != subgraph.beginIConnectable())
+ {
+ --it;
+ Layer& base = *(PolymorphicDowncast<Layer*>(*it));
+
+ // Special case to fuse padding into average pooling 2d for quantized datatype.
+ // Required to be done as a backend specific optimization as Neon does not support this special case.
+ if (base.GetType() == LayerType::Pooling2d)
+ {
+ Pooling2dLayer* baseLayer = PolymorphicDowncast<Pooling2dLayer*>(&base);
+ Pooling2dDescriptor poolingDescriptor = baseLayer->GetParameters();
+
+ if (baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer().GetType() == LayerType::Pad)
+ {
+ PadLayer* padLayer = PolymorphicDowncast<PadLayer*>(
+ &baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer());
+ if (padLayer->GetOutputSlot(0).GetNumConnections() == 1 &&
+ optimizations::pad_fold::TryFoldPadIntoLayer2d(padLayer->GetParameters(),
+ poolingDescriptor,
+ padLayer->GetOutputSlot().GetTensorInfo(),
+ true))
+ {
+ FoldPadIntoAveragePool2d<Pooling2dLayer>(optimizationViews, baseLayer,
+ poolingDescriptor, padLayer);
+ untouched.erase(baseLayer->GetGuid());
+ untouched.erase(padLayer->GetGuid());
+ }
+ }
+ }
+ }
+
+ if (optimizationViews.GetSubstitutions().empty())
+ {
+ optimizationViews.AddUntouchedSubgraph(SubgraphView(subgraph));
+ }
+ else
+ {
+ ReportUntouchedLayers(optimizationViews, untouched);
+ }
return optimizationViews;
}
diff --git a/src/backends/reference/RefBackend.hpp b/src/backends/reference/RefBackend.hpp
index 9828d09..ecbe4d5 100644
--- a/src/backends/reference/RefBackend.hpp
+++ b/src/backends/reference/RefBackend.hpp
@@ -50,7 +50,8 @@
IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
- OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph) const override;
+ OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph,
+ const ModelOptions& modelOptions) const override;
std::vector<ITensorHandleFactory::FactoryId> GetHandleFactoryPreferences() const override;