diff options
Diffstat (limited to 'src/backends/neon/workloads/NeonReduceWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonReduceWorkload.cpp | 53 |
1 files changed, 42 insertions, 11 deletions
diff --git a/src/backends/neon/workloads/NeonReduceWorkload.cpp b/src/backends/neon/workloads/NeonReduceWorkload.cpp index 0e1b46a3a1..6125f3609d 100644 --- a/src/backends/neon/workloads/NeonReduceWorkload.cpp +++ b/src/backends/neon/workloads/NeonReduceWorkload.cpp @@ -21,22 +21,52 @@ arm_compute::Status NeonReduceWorkloadValidate(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, - "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 @@ NeonReduceWorkload::NeonReduceWorkload(const ReduceQueueDescriptor& descriptor, 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]), |