diff options
author | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-05-03 12:22:03 +0100 |
---|---|---|
committer | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-05-06 17:58:26 +0000 |
commit | d905decd256558bbee165e636ce4242ac3b9c917 (patch) | |
tree | 86f51622399553d1741b66ff232a429de8fc43f8 /src/backends/neon/NeonBackend.cpp | |
parent | 1f58f03d82c482626b1b4673b6c0e25da4338fb5 (diff) | |
download | armnn-d905decd256558bbee165e636ce4242ac3b9c917.tar.gz |
MLCE-418 Reduce layer does not support multiple axes
* Added backend specific optimization to chain new reduces layers
for each axis to simulate behaviour of a layer with multiple axes.
* Added function to calculate reduced output shape.
* Added unit tests.
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I180b0b111b7bcf3d0c283f1db0b82d5f17757682
Diffstat (limited to 'src/backends/neon/NeonBackend.cpp')
-rw-r--r-- | src/backends/neon/NeonBackend.cpp | 24 |
1 files changed, 23 insertions, 1 deletions
diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp index a1299fb458..6d5eab0ddf 100644 --- a/src/backends/neon/NeonBackend.cpp +++ b/src/backends/neon/NeonBackend.cpp @@ -29,6 +29,7 @@ #include "workloads/NeonDivisionWorkload.hpp" #include "workloads/NeonFullyConnectedWorkload.hpp" #include "workloads/NeonMultiplicationWorkload.hpp" +#include "workloads/NeonReduceWorkload.hpp" #include "workloads/NeonSubtractionWorkload.hpp" #include <Optimizer.hpp> @@ -164,7 +165,8 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph if ((base.GetType() == LayerType::DepthwiseConvolution2d || base.GetType() == LayerType::Convolution2d || base.GetType() == LayerType::BatchNormalization || base.GetType() == LayerType::FullyConnected || base.GetType() == LayerType::Addition || base.GetType() == LayerType::Multiplication - || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division) + || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division + || base.GetType() == LayerType::Reduce) && (base.GetAdditionalInformation<ActivationDescriptor>() == nullptr)) { for (auto output = base.BeginOutputSlots(); output != base.EndOutputSlots(); ++output) @@ -389,6 +391,26 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph } } } + + // Separate check for Reduce as we aren't fusing with activation layer + if (base.GetType() == LayerType::Reduce) + { + ReduceLayer* baseLayer = PolymorphicDowncast<ReduceLayer*>(&base); + + // Get params from base layer + ReduceDescriptor reduceDescriptor = baseLayer->GetParameters(); + + arm_compute::Status status = NeonReduceWorkloadValidate( + baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(), + baseLayer->GetOutputSlot(0).GetTensorInfo(), + reduceDescriptor); + + if (status) + { + ChainReduceLayers<ReduceLayer>(optimizationViews, baseLayer, reduceDescriptor); + untouched.erase(baseLayer->GetGuid()); + } + } } } } |