Matteo Martincigh | 28dcab6 | 2018-10-19 16:40:03 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ClMeanWorkload.hpp" |
| 7 | |
Aron Virginas-Tar | c9cc804 | 2018-11-01 16:15:57 +0000 | [diff] [blame] | 8 | #include <cl/ClTensorHandle.hpp> |
| 9 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
Matteo Martincigh | 28dcab6 | 2018-10-19 16:40:03 +0100 | [diff] [blame] | 10 | |
| 11 | #include "ClWorkloadUtils.hpp" |
| 12 | |
| 13 | namespace |
| 14 | { |
| 15 | |
| 16 | void ConvertArmnnAxesToAclCoordinates(size_t inputDimensions, |
| 17 | unsigned int originalInputRank, |
| 18 | const std::vector<unsigned int>& armnnAxes, |
| 19 | arm_compute::Coordinates& outAclCoords) |
| 20 | { |
| 21 | if (armnnAxes.empty()) |
| 22 | { |
| 23 | // If no reduction axes were provided, then the input must be reduced along all dimensions. |
| 24 | // Since arm_compute::CLReduceMean does not accept an empty vector as the reduction dimensions, we then |
| 25 | // manually create a vector including all the input dimensions (in reversed order) as: |
| 26 | // |
| 27 | // { inputDimensions - 1, inputDimensions - 2, ..., 1, 0 } |
| 28 | // |
| 29 | outAclCoords.set_num_dimensions(inputDimensions); |
| 30 | std::generate(outAclCoords.begin(), outAclCoords.end(), [d = inputDimensions - 1] () mutable { return d--; }); |
| 31 | } |
| 32 | else |
| 33 | { |
| 34 | // Create a vector of reduction dimensions (in reversed order) with the given reduction axes. |
| 35 | // |
| 36 | // Adjust the given reduction axes according to the original rank of the input tensor (before ACL applied any |
| 37 | // dimension correction). |
| 38 | // For example, if the input tensor originally had 4 dimensions, and one of the reduction axes was 2, then the |
| 39 | // new value for that reduction axis should be 1. |
| 40 | // |
| 41 | // Example: |
| 42 | // ArmNN input shape = { 1, 1, 3, 2 } -> ACL input shape = { 2, 3 } |
| 43 | // ArmNN reduction axis = { 2 } -> ACL reduction axis = { 1 } |
| 44 | // ArmNN reduction axis = { 3 } -> ACL reduction axis = { 0 } |
| 45 | // |
| 46 | // The transformation: ACL reduction axis index = original rank - ArmNN reduction axis index - 1 |
| 47 | // |
| 48 | outAclCoords.set_num_dimensions(armnnAxes.size()); |
| 49 | std::transform(armnnAxes.begin(), armnnAxes.end(), |
| 50 | outAclCoords.begin(), |
| 51 | [originalInputRank](unsigned int i){ return originalInputRank - i - 1; }); |
| 52 | } |
| 53 | } |
| 54 | |
| 55 | } // anonymous namespace |
| 56 | |
| 57 | namespace armnn |
| 58 | { |
| 59 | using namespace armcomputetensorutils; |
| 60 | |
| 61 | arm_compute::Status ClMeanValidate(const TensorInfo& input, |
| 62 | const TensorInfo& output, |
| 63 | const MeanDescriptor& desc) |
| 64 | { |
| 65 | const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); |
| 66 | const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); |
| 67 | |
| 68 | arm_compute::Coordinates coords; |
| 69 | ConvertArmnnAxesToAclCoordinates(aclInputInfo.num_dimensions(), |
| 70 | input.GetNumDimensions(), |
| 71 | desc.m_Axis, |
| 72 | coords); |
| 73 | |
| 74 | return arm_compute::CLReduceMean::validate(&aclInputInfo, coords, desc.m_KeepDims, &aclOutputInfo); |
| 75 | } |
| 76 | |
| 77 | ClMeanWorkload::ClMeanWorkload(const MeanQueueDescriptor& descriptor, const WorkloadInfo& info) |
| 78 | : BaseWorkload<MeanQueueDescriptor>(descriptor, info) |
| 79 | { |
| 80 | m_Data.ValidateInputsOutputs("ClMeanWorkload", 1, 1); |
| 81 | |
| 82 | arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); |
| 83 | arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); |
| 84 | |
| 85 | arm_compute::Coordinates coords; |
| 86 | ConvertArmnnAxesToAclCoordinates(input.info()->num_dimensions(), |
| 87 | info.m_InputTensorInfos[0].GetNumDimensions(), |
| 88 | m_Data.m_Parameters.m_Axis, |
| 89 | coords); |
| 90 | |
| 91 | m_Layer.configure(&input, coords, m_Data.m_Parameters.m_KeepDims, &output); |
| 92 | } |
| 93 | |
| 94 | void ClMeanWorkload::Execute() const |
| 95 | { |
| 96 | ARMNN_SCOPED_PROFILING_EVENT_CL("ClMeanWorkload_Execute"); |
| 97 | m_Layer.run(); |
| 98 | } |
| 99 | |
| 100 | } //namespace armnn |