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
diff --git a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
index a0fca46..9439ddb 100644
--- a/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeSubgraphUtils.hpp
@@ -6,6 +6,9 @@
#pragma once
#include <armnn/backends/OptimizationViews.hpp>
+#include <armnn/utility/Assert.hpp>
+
+#include <aclCommon/ArmComputeUtils.hpp>
namespace armnn
{
@@ -147,4 +150,86 @@
return replacementLayer;
}
+//
+// If reduce layer has multiple axes, add new layer for each axis to simulate the same behaviour
+// as currently only one axis is supported.
+//
+template<typename LayerType>
+void ChainReduceLayers(OptimizationViews& optimizationViews,
+ LayerType* baseLayer,
+ ReduceDescriptor& reduceDescriptor)
+{
+ // If layer has single axis don't chain layers.
+ if (!reduceDescriptor.m_vAxis.empty() && reduceDescriptor.m_vAxis.size() > 1)
+ {
+ // Save base layer output shape to compare against the output of the final layer added.
+ const TensorInfo baseLayerInfo = baseLayer->GetOutputSlot(0).GetTensorInfo();
+
+ // Vector of new chained layers, used for substitution.
+ std::vector<Layer*> layers;
+
+ // Vector of axes so each layer is reshaped correctly.
+ std::vector<uint32_t> reduceAxis;
+ unsigned int recalulateAxis = 0;
+
+ for (unsigned int i = 0; i != reduceDescriptor.m_vAxis.size(); ++i)
+ {
+ // Get TensorInfo to populate subsequent layers with.
+ TensorInfo layerInfoToModify = baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo();
+
+ reduceAxis.emplace_back(reduceDescriptor.m_vAxis[i]);
+
+ // Calculate new shape based on the axes.
+ const TensorShape& reducedShape = ComputeReductionTensorShape(layerInfoToModify,
+ reduceAxis,
+ reduceDescriptor.m_KeepDims);
+ layerInfoToModify.SetShape(reducedShape);
+
+ // Create a vector for the single axis to be assigned to the descriptor.
+ // Update axis if keepDims is set reduce layers correctly.
+ std::vector<uint32_t> singleAxis(1, reduceDescriptor.m_vAxis[i] - recalulateAxis);
+
+ // Create a descriptor and assign single axis.
+ ReduceDescriptor newReduceDescriptor = baseLayer->GetParameters();
+ newReduceDescriptor.m_vAxis.assign(singleAxis.begin(), singleAxis.end());
+
+ // Add new layer to graph.
+ std::string layerName = "reduce_layer_" + std::to_string(i);
+ Layer* replacementLayer = optimizationViews.GetGraph().AddLayer<LayerType>(newReduceDescriptor,
+ layerName.c_str());
+
+ // Connect previous layer with new layer.
+ // The first and last layer will be connected when the subgraph is replaced.
+ if (!layers.empty())
+ {
+ layers[i - 1]->GetOutputSlot(0).Connect(replacementLayer->GetInputSlot(0));
+ }
+
+ // Set updated tensorInfo for new layer.
+ replacementLayer->GetOutputSlot(0).SetTensorInfo(layerInfoToModify);
+
+ if (!reduceDescriptor.m_KeepDims)
+ {
+ recalulateAxis++;
+ }
+
+ layers.emplace_back(replacementLayer);
+ }
+
+ // Check if the TensorInfo from the last layer equals the inferred output from the original layer.
+ ARMNN_ASSERT(baseLayerInfo == layers.back()->GetOutputSlot().GetTensorInfo());
+
+ std::list<Layer*> replacementLayers(layers.begin(), layers.end());
+
+ // Substitute new chained subgraph for original reduce layer.
+ SubgraphView substitutionSubgraph(baseLayer);
+ SubgraphView replacementSubgraph(CreateInputsFrom({replacementLayers.front()}),
+ CreateOutputsFrom({replacementLayers.back()}),
+ std::move(replacementLayers));
+
+ optimizationViews.AddSubstitution({substitutionSubgraph, replacementSubgraph});
+
+ }
+}
+
} // namespace armnn
diff --git a/src/backends/aclCommon/ArmComputeUtils.hpp b/src/backends/aclCommon/ArmComputeUtils.hpp
index d9efab2..5bc5abc 100644
--- a/src/backends/aclCommon/ArmComputeUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeUtils.hpp
@@ -7,6 +7,7 @@
#include <armnn/Descriptors.hpp>
#include <armnn/Tensor.hpp>
#include <armnn/utility/Assert.hpp>
+#include <armnn/utility/NumericCast.hpp>
#include <backendsCommon/WorkloadData.hpp>
#include <arm_compute/core/Types.h>
@@ -267,4 +268,58 @@
}
}
+/// Function to compute the output tensor shape based on the axes and if keepDims is set.
+inline const TensorShape ComputeReductionTensorShape(const armnn::TensorInfo& input,
+ const std::vector<uint32_t>& vAxis,
+ const bool keepDims)
+{
+ unsigned int rank = input.GetNumDimensions();
+ unsigned int outputRank = 0;
+
+ // Calculate output dimension
+ if (keepDims)
+ {
+ outputRank = rank;
+ }
+ else if (vAxis.empty())
+ {
+ outputRank = 1;
+ }
+ else if (vAxis.size() > input.GetNumDimensions())
+ {
+ throw LayerValidationException("ReduceLayer: Dimensions to reduce can not be bigger than input dimensions");
+ }
+ else
+ {
+ outputRank = input.GetNumDimensions() - armnn::numeric_cast<unsigned int>(vAxis.size());
+ if (outputRank == 0)
+ {
+ outputRank = 1;
+ }
+ }
+
+ std::vector<unsigned int> dimSizes(outputRank, 1);
+ if (!vAxis.empty())
+ {
+ // Skip the dimension that has been reduced unless keepDims is true.
+ unsigned int outputIndex = 0;
+ for (unsigned int i = 0; i < input.GetNumDimensions(); ++i)
+ {
+ if (std::find(vAxis.begin(), vAxis.end(), i) == vAxis.end())
+ {
+ dimSizes[outputIndex] = armnn::numeric_cast<unsigned int>(input.GetShape()[i]);
+ ++outputIndex;
+ }
+ else if (keepDims)
+ {
+ dimSizes[outputIndex] = 1;
+ ++outputIndex;
+ }
+ }
+ }
+
+ const TensorShape inferredShape = TensorShape(outputRank, dimSizes.data());
+ return inferredShape;
+}
+
} // namespace armnn
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index f97cb4b..92a06aa 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -29,6 +29,7 @@
#include "workloads/ClDivisionWorkload.hpp"
#include "workloads/ClFullyConnectedWorkload.hpp"
#include "workloads/ClMultiplicationWorkload.hpp"
+#include "workloads/ClReduceWorkload.hpp"
#include "workloads/ClSubtractionWorkload.hpp"
#include <Optimizer.hpp>
@@ -188,7 +189,8 @@
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)
@@ -412,6 +414,26 @@
}
}
}
+
+ // 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 = ClReduceWorkloadValidate(
+ baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(),
+ baseLayer->GetOutputSlot(0).GetTensorInfo(),
+ reduceDescriptor);
+
+ if (status)
+ {
+ ChainReduceLayers<ReduceLayer>(optimizationViews, baseLayer, reduceDescriptor);
+ untouched.erase(baseLayer->GetGuid());
+ }
+ }
}
}
}
diff --git a/src/backends/cl/workloads/ClReduceWorkload.cpp b/src/backends/cl/workloads/ClReduceWorkload.cpp
index 6f594ff..0ad6259 100644
--- a/src/backends/cl/workloads/ClReduceWorkload.cpp
+++ b/src/backends/cl/workloads/ClReduceWorkload.cpp
@@ -20,23 +20,52 @@
const ReduceDescriptor& desc)
{
const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
- const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
- if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1)
- {
- return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
- "ClReduceWorkload: Reduction is supported only on 1 axis.");
- }
arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
input.GetNumDimensions(),
desc.m_vAxis);
+ // As ACL only support one axis, validate the layer for each axis if more than one is present.
+ if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1)
+ {
+ arm_compute::Status status;
- return arm_compute::CLReductionOperation::validate(&aclInputInfo,
- &aclOutputInfo,
- static_cast<unsigned int>(coords[0]),
- ConvertReductionOperationToAcl(desc),
- desc.m_KeepDims);
+ for (unsigned int i = 0; i != desc.m_vAxis.size(); ++i)
+ {
+ TensorInfo inputToModify = input;
+ std::vector<uint32_t> singleAxis(1, desc.m_vAxis[i]);
+
+ // Calculate the output shape using the input shape for a single axis.
+ // Currently the output TensorInfo inferred will be reduced upon multiple axis
+ // which will fail validation as only one axis is supported.
+ const TensorShape& reducedShape = ComputeReductionTensorShape(inputToModify, singleAxis, desc.m_KeepDims);
+ inputToModify.SetShape(reducedShape);
+
+ const arm_compute::TensorInfo aclOutputInfoModified =
+ armcomputetensorutils::BuildArmComputeTensorInfo(inputToModify);
+
+ status = arm_compute::CLReductionOperation::validate(&aclInputInfo,
+ &aclOutputInfoModified,
+ static_cast<unsigned int>(coords[i]),
+ ConvertReductionOperationToAcl(desc),
+ desc.m_KeepDims);
+ if (!status)
+ {
+ break;
+ }
+ }
+ return status;
+ }
+ else
+ {
+ const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ return arm_compute::CLReductionOperation::validate(&aclInputInfo,
+ &aclOutputInfo,
+ static_cast<unsigned int>(coords[0]),
+ ConvertReductionOperationToAcl(desc),
+ desc.m_KeepDims);
+ }
}
ClReduceWorkload::ClReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info)
diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp
index a1299fb..6d5eab0 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 @@
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 @@
}
}
}
+
+ // 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());
+ }
+ }
}
}
}
diff --git a/src/backends/neon/workloads/NeonReduceWorkload.cpp b/src/backends/neon/workloads/NeonReduceWorkload.cpp
index 0e1b46a..6125f36 100644
--- a/src/backends/neon/workloads/NeonReduceWorkload.cpp
+++ b/src/backends/neon/workloads/NeonReduceWorkload.cpp
@@ -21,22 +21,52 @@
const ReduceDescriptor& desc)
{
const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
- const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
- if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1)
- {
- return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
- "NeonReduceWorkload: Reduction is supported only on 1 axis.");
- }
arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
input.GetNumDimensions(),
desc.m_vAxis);
- return arm_compute::NEReductionOperation::validate(&aclInputInfo,
- &aclOutputInfo,
- static_cast<unsigned int>(coords[0]),
- ConvertReductionOperationToAcl(desc),
- desc.m_KeepDims);
+ // As ACL only support one axis, validate the layer for each axis if more than one is present.
+ if (!desc.m_vAxis.empty() && desc.m_vAxis.size() > 1)
+ {
+ arm_compute::Status status;
+
+ for (unsigned int i = 0; i != desc.m_vAxis.size(); ++i)
+ {
+ TensorInfo inputToModify = input;
+ std::vector<uint32_t> singleAxis(1, desc.m_vAxis[i]);
+
+ // Calculate the output shape using the input shape for a single axis.
+ // Currently the output TensorInfo inferred will be reduced upon multiple axis
+ // which will fail validation as only one axis is supported.
+ const TensorShape& reducedShape = ComputeReductionTensorShape(inputToModify, singleAxis, desc.m_KeepDims);
+ inputToModify.SetShape(reducedShape);
+
+ const arm_compute::TensorInfo aclOutputInfoModified =
+ armcomputetensorutils::BuildArmComputeTensorInfo(inputToModify);
+
+ status = arm_compute::NEReductionOperation::validate(&aclInputInfo,
+ &aclOutputInfoModified,
+ static_cast<unsigned int>(coords[i]),
+ ConvertReductionOperationToAcl(desc),
+ desc.m_KeepDims);
+ if (!status)
+ {
+ break;
+ }
+ }
+ return status;
+ }
+ else
+ {
+ const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+ return arm_compute::NEReductionOperation::validate(&aclInputInfo,
+ &aclOutputInfo,
+ static_cast<unsigned int>(coords[0]),
+ ConvertReductionOperationToAcl(desc),
+ desc.m_KeepDims);
+ }
}
NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info)
@@ -50,6 +80,7 @@
arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
info.m_InputTensorInfos[0].GetNumDimensions(),
m_Data.m_Parameters.m_vAxis);
+
m_Layer.configure(&input,
&output,
static_cast<unsigned int>(coords[0]),