diff options
Diffstat (limited to 'src/backends/cl')
| -rw-r--r-- | src/backends/cl/ClBackend.cpp | 24 | ||||
| -rw-r--r-- | src/backends/cl/workloads/ClReduceWorkload.cpp | 51 |
2 files changed, 12 insertions, 63 deletions
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp index 92a06aa8e1..f97cb4bba8 100644 --- a/src/backends/cl/ClBackend.cpp +++ b/src/backends/cl/ClBackend.cpp @@ -29,7 +29,6 @@ #include "workloads/ClDivisionWorkload.hpp" #include "workloads/ClFullyConnectedWorkload.hpp" #include "workloads/ClMultiplicationWorkload.hpp" -#include "workloads/ClReduceWorkload.hpp" #include "workloads/ClSubtractionWorkload.hpp" #include <Optimizer.hpp> @@ -189,8 +188,7 @@ OptimizationViews ClBackend::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::Reduce) + || base.GetType() == LayerType::Subtraction || base.GetType() == LayerType::Division) && (base.GetAdditionalInformation<ActivationDescriptor>() == nullptr)) { for (auto output = base.BeginOutputSlots(); output != base.EndOutputSlots(); ++output) @@ -414,26 +412,6 @@ OptimizationViews ClBackend::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 = 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 0ad6259cc2..6f594ff7a9 100644 --- a/src/backends/cl/workloads/ClReduceWorkload.cpp +++ b/src/backends/cl/workloads/ClReduceWorkload.cpp @@ -20,52 +20,23 @@ arm_compute::Status ClReduceWorkloadValidate(const TensorInfo& input, 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; - - 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); - } + 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) |