xref: /aosp_15_r20/external/armnn/src/backends/cl/workloads/ClReduceWorkload.cpp (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810)
1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "ClReduceWorkload.hpp"
7 
8 #include <cl/ClTensorHandle.hpp>
9 #include <aclCommon/ArmComputeUtils.hpp>
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 
12 #include "ClWorkloadUtils.hpp"
13 
14 namespace armnn
15 {
16 using namespace armcomputetensorutils;
17 
ClReduceWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const ReduceDescriptor & descriptor)18 arm_compute::Status ClReduceWorkloadValidate(const TensorInfo& input,
19                                              const TensorInfo& output,
20                                              const ReduceDescriptor& descriptor)
21 {
22     if (descriptor.m_vAxis.size() == 1 || descriptor.m_vAxis.empty())
23     {
24         const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
25         const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
26 
27         arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
28                                                                               input.GetNumDimensions(),
29                                                                               descriptor.m_vAxis);
30 
31         return arm_compute::CLReductionOperation::validate(&aclInputInfo,
32                                                            &aclOutputInfo,
33                                                            static_cast<unsigned int>(coords[0]),
34                                                            ConvertReductionOperationToAcl(descriptor),
35                                                            descriptor.m_KeepDims);
36     }
37     else
38     {
39         // Validate layer if there are multiple axes.
40         arm_compute::Status status;
41         IS_MULTI_AXES_REDUCE_SUPPORTED(ClReduceWorkloadValidate, input, descriptor, status);
42         return status;
43     }
44 }
45 
ClReduceWorkload(const ReduceQueueDescriptor & descriptor,const WorkloadInfo & info)46 ClReduceWorkload::ClReduceWorkload(const ReduceQueueDescriptor& descriptor, const WorkloadInfo& info)
47     : ClBaseWorkload<ReduceQueueDescriptor>(descriptor, info)
48 {
49     // Report Profiling Details
50     ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClReduceWorkload_Construct",
51                                          descriptor.m_Parameters,
52                                          info,
53                                          this->GetGuid());
54 
55     m_Data.ValidateInputsOutputs("ClReduceWorkload", 1, 1);
56 
57     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
58     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
59 
60     arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
61                                                                           info.m_InputTensorInfos[0].GetNumDimensions(),
62                                                                           m_Data.m_Parameters.m_vAxis);
63     {
64         ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClReduceWorkload_configure");
65         m_Layer.configure(&input,
66                           &output,
67                           static_cast<unsigned int>(coords[0]),
68                           ConvertReductionOperationToAcl(m_Data.m_Parameters),
69                           m_Data.m_Parameters.m_KeepDims);
70     }
71 }
72 
Execute() const73 void ClReduceWorkload::Execute() const
74 {
75     ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClReduceWorkload_Execute", this->GetGuid());
76     m_Layer.run();
77 }
78 
79 } //namespace armnn
80