1 //
2 // Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "RefElementwiseWorkload.hpp"
7
8 #include "Decoders.hpp"
9 #include "ElementwiseFunction.hpp"
10 #include "Encoders.hpp"
11 #include "Profiling.hpp"
12 #include "RefWorkloadUtils.hpp"
13 #include "StringMapping.hpp"
14 #include <ResolveType.hpp>
15 #include <vector>
16
17 namespace armnn
18 {
19
20 template <typename Functor, typename ParentDescriptor, typename armnn::StringMapping::Id DebugString>
RefElementwiseWorkload(const ParentDescriptor & desc,const WorkloadInfo & info)21 RefElementwiseWorkload<Functor, ParentDescriptor, DebugString>::RefElementwiseWorkload(
22 const ParentDescriptor& desc,
23 const WorkloadInfo& info)
24 : RefBaseWorkload<ParentDescriptor>(desc, info)
25 {
26 }
27
28 template <typename Functor, typename ParentDescriptor, typename armnn::StringMapping::Id DebugString>
Execute() const29 void RefElementwiseWorkload<Functor, ParentDescriptor, DebugString>::Execute() const
30 {
31 Execute(m_Data.m_Inputs, m_Data.m_Outputs);
32 }
33
34 template <typename Functor, typename ParentDescriptor, typename armnn::StringMapping::Id DebugString>
ExecuteAsync(ExecutionData & executionData)35 void RefElementwiseWorkload<Functor, ParentDescriptor, DebugString>::ExecuteAsync(
36 ExecutionData& executionData)
37 {
38 WorkingMemDescriptor* workingMemDescriptor = static_cast<WorkingMemDescriptor*>(executionData.m_Data);
39 Execute(workingMemDescriptor->m_Inputs, workingMemDescriptor->m_Outputs);
40 }
41
42 template <typename Functor, typename ParentDescriptor, typename armnn::StringMapping::Id DebugString>
Execute(std::vector<ITensorHandle * > inputs,std::vector<ITensorHandle * > outputs) const43 void RefElementwiseWorkload<Functor, ParentDescriptor, DebugString>::Execute(
44 std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const
45 {
46 ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, StringMapping::Instance().Get(DebugString));
47 const TensorInfo& inputInfo0 = GetTensorInfo(inputs[0]);
48 const TensorInfo& inputInfo1 = GetTensorInfo(inputs[1]);
49 const TensorInfo& outputInfo = GetTensorInfo(outputs[0]);
50
51 const TensorShape& inShape0 = inputInfo0.GetShape();
52 const TensorShape& inShape1 = inputInfo1.GetShape();
53 const TensorShape& outShape = outputInfo.GetShape();
54
55 std::unique_ptr<Decoder<InType>> input0 = MakeDecoder<InType>(inputInfo0, inputs[0]->Map());
56 std::unique_ptr<Decoder<InType>> input1 = MakeDecoder<InType>(inputInfo1, inputs[1]->Map());
57 std::unique_ptr<Encoder<OutType>> output= MakeEncoder<OutType>(outputInfo, outputs[0]->Map());
58
59 ElementwiseBinaryFunction<Functor>(inShape0,
60 inShape1,
61 outShape,
62 *input0,
63 *input1,
64 *output);
65 }
66
67 } //namespace armnn
68
69 template class armnn::RefElementwiseWorkload<std::plus<float>,
70 armnn::AdditionQueueDescriptor,
71 armnn::StringMapping::RefAdditionWorkload_Execute>;
72
73 template class armnn::RefElementwiseWorkload<std::plus<int32_t>,
74 armnn::AdditionQueueDescriptor,
75 armnn::StringMapping::RefAdditionWorkload_Execute>;
76
77 template class armnn::RefElementwiseWorkload<std::minus<float>,
78 armnn::SubtractionQueueDescriptor,
79 armnn::StringMapping::RefSubtractionWorkload_Execute>;
80
81 template class armnn::RefElementwiseWorkload<std::minus<int32_t>,
82 armnn::SubtractionQueueDescriptor,
83 armnn::StringMapping::RefSubtractionWorkload_Execute>;
84
85 template class armnn::RefElementwiseWorkload<std::multiplies<float>,
86 armnn::MultiplicationQueueDescriptor,
87 armnn::StringMapping::RefMultiplicationWorkload_Execute>;
88
89 template class armnn::RefElementwiseWorkload<std::multiplies<int32_t>,
90 armnn::MultiplicationQueueDescriptor,
91 armnn::StringMapping::RefMultiplicationWorkload_Execute>;
92
93 template class armnn::RefElementwiseWorkload<std::divides<float>,
94 armnn::DivisionQueueDescriptor,
95 armnn::StringMapping::RefDivisionWorkload_Execute>;
96
97 template class armnn::RefElementwiseWorkload<std::divides<int32_t>,
98 armnn::DivisionQueueDescriptor,
99 armnn::StringMapping::RefDivisionWorkload_Execute>;
100
101 template class armnn::RefElementwiseWorkload<armnn::maximum<float>,
102 armnn::MaximumQueueDescriptor,
103 armnn::StringMapping::RefMaximumWorkload_Execute>;
104
105 template class armnn::RefElementwiseWorkload<armnn::maximum<int32_t>,
106 armnn::MaximumQueueDescriptor,
107 armnn::StringMapping::RefMaximumWorkload_Execute>;
108
109 template class armnn::RefElementwiseWorkload<armnn::minimum<float>,
110 armnn::MinimumQueueDescriptor,
111 armnn::StringMapping::RefMinimumWorkload_Execute>;
112
113 template class armnn::RefElementwiseWorkload<armnn::minimum<int32_t>,
114 armnn::MinimumQueueDescriptor,
115 armnn::StringMapping::RefMinimumWorkload_Execute>;
116