diff options
author | Cathal Corbett <cathal.corbett@arm.com> | 2022-07-22 16:03:36 +0100 |
---|---|---|
committer | Nikhil Raj <nikhil.raj@arm.com> | 2022-08-05 15:50:57 +0100 |
commit | 3883b2776cec33f16f0ea9a2d795de2b7c766df7 (patch) | |
tree | 6842e15904037d73426d814d5751945b3d9c2376 /src/backends/reference/RefBackend.cpp | |
parent | 9d63fee68081b65bd72de3a70da76c2696c6c6ed (diff) | |
download | armnn-3883b2776cec33f16f0ea9a2d795de2b7c766df7.tar.gz |
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
Diffstat (limited to 'src/backends/reference/RefBackend.cpp')
-rw-r--r-- | src/backends/reference/RefBackend.cpp | 59 |
1 files changed, 55 insertions, 4 deletions
diff --git a/src/backends/reference/RefBackend.cpp b/src/backends/reference/RefBackend.cpp index a33a7756a0..8c8879c8be 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 @@ IBackendInternal::ILayerSupportSharedPtr RefBackend::GetLayerSupport() const return layerSupport; } -OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph) const +OptimizationViews RefBackend::OptimizeSubgraphView(const SubgraphView& subgraph, + const ModelOptions& modelOptions) const { - OptimizationViews optimizationViews; - - optimizationViews.AddUntouchedSubgraph(SubgraphView(subgraph)); + OptimizationViews optimizationViews(modelOptions); + + 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; } |