1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "ClConcatWorkload.hpp"
6 #include "ClWorkloadUtils.hpp"
7 #include <aclCommon/ArmComputeTensorUtils.hpp>
8 #include <armnn/utility/PolymorphicDowncast.hpp>
9 #include <backendsCommon/CpuTensorHandle.hpp>
10 #include <cl/ClTensorHandle.hpp>
11 #include <cl/ClLayerSupport.hpp>
12
13 #include <arm_compute/core/Types.h>
14 #include <arm_compute/runtime/CL/functions/CLConcatenateLayer.h>
15
16 namespace armnn
17 {
18 using namespace armcomputetensorutils;
19
20 namespace
21 {
CalcAxis(const OriginsDescriptor & desc)22 size_t CalcAxis(const OriginsDescriptor& desc)
23 {
24 return (desc.GetNumDimensions() - desc.GetConcatAxis()) - 1;
25 }
26 } //namespace
27
ClConcatWorkloadValidate(const std::vector<const TensorInfo * > & inputs,const TensorInfo & output,const OriginsDescriptor & descriptor)28 arm_compute::Status ClConcatWorkloadValidate(const std::vector<const TensorInfo*>& inputs,
29 const TensorInfo& output,
30 const OriginsDescriptor& descriptor)
31 {
32 std::vector<arm_compute::TensorInfo> aclInputs;
33 for (const TensorInfo* input : inputs)
34 {
35 arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW);
36 aclInputs.emplace_back(aclInputInfo);
37 }
38 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
39 std::vector<const arm_compute::ITensorInfo*> aclInputPtrs;
40 for (arm_compute::ITensorInfo& input : aclInputs)
41 {
42 aclInputPtrs.emplace_back(&input);
43 }
44
45 size_t aclAxis = CalcAxis(descriptor);
46 return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis);
47 }
48
ClConcatWorkload(const ConcatQueueDescriptor & descriptor,const WorkloadInfo & info)49 ClConcatWorkload::ClConcatWorkload(const ConcatQueueDescriptor& descriptor, const WorkloadInfo& info)
50 : BaseWorkload<ConcatQueueDescriptor>(descriptor, info)
51 {
52 bool allInputsAreSubtensors = true;
53
54 // Check that all inputs are sub-tensors
55 for (auto input : descriptor.m_Inputs)
56 {
57 if (!input->GetParent())
58 {
59 // Non sub-tensor input found so we need to execute the concat function
60 allInputsAreSubtensors = false;
61 break;
62 }
63 }
64
65 if (allInputsAreSubtensors)
66 {
67 // Can skip configuring the concat function since it's not executed
68 return;
69 }
70
71 std::vector<const arm_compute::ICLTensor *> aclInputs;
72 for (auto input : m_Data.m_Inputs)
73 {
74 arm_compute::ICLTensor& aclInput = armnn::PolymorphicPointerDowncast<IClTensorHandle>(input)->GetTensor();
75 aclInputs.emplace_back(&aclInput);
76 }
77
78 arm_compute::ICLTensor& output =
79 armnn::PolymorphicPointerDowncast<IClTensorHandle>(m_Data.m_Outputs[0])->GetTensor();
80
81 // Create the layer function
82 auto layer = std::make_unique<arm_compute::CLConcatenateLayer>();
83
84 // Configure input and output tensors
85 size_t aclAxis = CalcAxis(descriptor.m_Parameters);
86 layer->configure(aclInputs, &output, aclAxis);
87
88 // Prepare
89 layer->prepare();
90 m_Layer = std::move(layer);
91 }
92
Execute() const93 void ClConcatWorkload::Execute() const
94 {
95 if (m_Layer)
96 {
97 ARMNN_SCOPED_PROFILING_EVENT_CL("ClConcatWorkload_Execute");
98 m_Layer->run();
99 }
100 }
101
102 } //namespace armnn
103