| Nikhil Raj | 8599a41 | 2018-11-19 14:51:07 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #include "ClMergerWorkload.hpp" |
| 6 | #include "ClWorkloadUtils.hpp" |
| 7 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 8 | #include <backendsCommon/CpuTensorHandle.hpp> |
| 9 | #include <cl/ClTensorHandle.hpp> |
| 10 | #include <cl/ClLayerSupport.hpp> |
| 11 | |
| 12 | #include <boost/polymorphic_pointer_cast.hpp> |
| 13 | |
| 14 | namespace armnn |
| 15 | { |
| 16 | using namespace armcomputetensorutils; |
| 17 | |
| 18 | arm_compute::Status ClMergerWorkloadValidate(const std::vector<const TensorInfo*>& inputs, |
| 19 | const TensorInfo& output, |
| 20 | const MergerDescriptor& descriptor) |
| 21 | |
| 22 | { |
| 23 | std::vector<arm_compute::TensorInfo> aclInputs; |
| 24 | for (const TensorInfo* input : inputs) |
| 25 | { |
| 26 | arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW); |
| 27 | aclInputs.emplace_back(aclInputInfo); |
| 28 | } |
| 29 | const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); |
| 30 | arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH; |
| 31 | |
| 32 | std::vector<arm_compute::ITensorInfo*> aclInputPtrs; |
| 33 | for (arm_compute::ITensorInfo& input : aclInputs) |
| 34 | { |
| 35 | aclInputPtrs.emplace_back(&input); |
| 36 | } |
| 37 | |
| 38 | return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis); |
| 39 | |
| 40 | } |
| 41 | |
| 42 | ClMergerWorkload::ClMergerWorkload(const MergerQueueDescriptor& descriptor, const WorkloadInfo& info) |
| 43 | : BaseWorkload<MergerQueueDescriptor>(descriptor, info) |
| 44 | { |
| 45 | m_Execute = true; |
| 46 | |
| 47 | unsigned int innerAxisOrder = descriptor.m_Parameters.GetNumDimensions() - descriptor.m_Parameters.GetConcatAxis(); |
| 48 | |
| 49 | if (innerAxisOrder != 1) |
| 50 | { |
| 51 | m_Execute = false; |
| 52 | return; |
| 53 | } |
| 54 | |
| 55 | std::vector<arm_compute::ICLTensor *> aclInputs; |
| 56 | arm_compute::DataLayout aclDataLayout = ConvertDataLayout(armnn::DataLayout::NCHW); |
| 57 | for (auto input : m_Data.m_Inputs) |
| 58 | { |
| 59 | arm_compute::ICLTensor& aclInput = boost::polymorphic_pointer_downcast<IClTensorHandle>(input)->GetTensor(); |
| 60 | aclInput.info()->set_data_layout(aclDataLayout); |
| 61 | aclInputs.emplace_back(&aclInput); |
| 62 | } |
| 63 | arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>( |
| 64 | m_Data.m_Outputs[0])->GetTensor(); |
| 65 | output.info()->set_data_layout(aclDataLayout); |
| 66 | |
| 67 | arm_compute::DataLayoutDimension aclAxis = arm_compute::DataLayoutDimension::WIDTH; |
| 68 | |
| 69 | m_Layer.configure(aclInputs, &output, aclAxis); |
| 70 | |
| 71 | m_Layer.prepare(); |
| 72 | |
| 73 | } |
| 74 | |
| 75 | void ClMergerWorkload::Execute() const |
| 76 | { |
| 77 | if (m_Execute) |
| 78 | { |
| 79 | ARMNN_SCOPED_PROFILING_EVENT_CL("ClMergerWorkload_Execute"); |
| 80 | m_Layer.run(); |
| 81 | } |
| 82 | |
| 83 | } |
| 84 | |
| 85 | } //namespace armnn |