xref: /aosp_15_r20/external/armnn/src/backends/reference/workloads/RefGatherNdWorkload.cpp (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810)
1 //
2 // Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "RefGatherNdWorkload.hpp"
7 
8 #include "Gather.hpp"
9 #include "Profiling.hpp"
10 #include "RefWorkloadUtils.hpp"
11 #include "backendsCommon/WorkloadUtils.hpp"
12 
13 namespace armnn
14 {
15 
Execute() const16 void RefGatherNdWorkload::Execute() const
17 {
18     Execute(m_Data.m_Inputs, m_Data.m_Outputs);
19 }
20 
ExecuteAsync(ExecutionData & executionData)21 void RefGatherNdWorkload::ExecuteAsync(ExecutionData& executionData)
22 {
23     WorkingMemDescriptor* workingMemDescriptor = static_cast<WorkingMemDescriptor*>(executionData.m_Data);
24     Execute(workingMemDescriptor->m_Inputs, workingMemDescriptor->m_Outputs);
25 }
26 
Execute(std::vector<ITensorHandle * > inputs,std::vector<ITensorHandle * > outputs) const27 void RefGatherNdWorkload::Execute(std::vector<ITensorHandle*> inputs, std::vector<ITensorHandle*> outputs) const
28 {
29     ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "RefGatherNdWorkload_Execute");
30 
31     const TensorInfo& inputInfo0 = GetTensorInfo(inputs[0]);
32     const TensorInfo& inputInfo1 = GetTensorInfo(inputs[1]);
33     const TensorInfo& outputInfo = GetTensorInfo(outputs[0]);
34 
35     std::unique_ptr<Decoder<float>> params_decoderPtr = MakeDecoder<float>(inputInfo0, inputs[0]->Map());
36 
37     const int32_t* indicesDataPtr = reinterpret_cast<int32_t*>(inputs[1]->Map());
38     std::vector<int32_t> indices(indicesDataPtr, indicesDataPtr + inputInfo1.GetNumElements());
39 
40     std::unique_ptr<Encoder<float>> output_encoderPtr = MakeEncoder<float>(outputInfo, outputs[0]->Map());
41 
42     std::map<std::string, unsigned int> keyIndices = CalculateGatherNdKeyIndices(inputInfo0, inputInfo1);
43 
44     /// Calculate flattened indices: flattenedIndices = indices * flattenedCoefficients
45     // Calculate the flattened coefficients to use in the multiplication
46     // to calculate the flattened indices needed by gather
47     TensorShape paramsShape = inputInfo0.GetShape();
48     std::vector<unsigned int> flattenedCoeff(keyIndices["ND"], 1);
49     for (unsigned int i = 1; i < keyIndices["ND"]; ++i)
50     {
51         flattenedCoeff[i-1] = paramsShape[i];
52     }
53     for (unsigned int i = keyIndices["ND"]-1; i > 0; --i)
54     {
55         flattenedCoeff[i-1] *= flattenedCoeff[i];
56     }
57 
58     // Prepare the vector to store the output of the matrix multiplication,
59     // which will represent the flattened indices needed by gather
60     armnn::TensorInfo flattenedIndices_Info = inputInfo1;
61     flattenedIndices_Info.SetShape({ keyIndices["W"] });
62     std::vector<int32_t> flattenedIndices(flattenedIndices_Info.GetNumElements(), 0);
63 
64     // Multiplication to calculate the flattened indices, which are the indices needed by gather.
65     for (unsigned int i = 0; i < keyIndices["W"]; ++i)
66     {
67         for (unsigned int j = 0; j < keyIndices["ND"]; ++j)
68         {
69             flattenedIndices[i] += indices[i * keyIndices["ND"] + j] * static_cast<int32_t>(flattenedCoeff[j]);
70         }
71     }
72 
73     /// Call Gather with adequate shapes
74     // Reshape params into {K, C}
75     armnn::TensorInfo params_K_C_Info =  inputInfo0;
76     params_K_C_Info.SetShape({ keyIndices["K"], keyIndices["C"] });
77 
78     // Reshape indices into {N, W}
79     armnn::TensorInfo indices_N_W_Info = inputInfo1;
80     indices_N_W_Info.SetShape({ keyIndices["N"], keyIndices["W"] });
81 
82     // Reshape output to have the shape given by gather {N, W, C}
83     // (the original outputInfo has the shape given by gatherNd)
84     armnn::TensorInfo outputGather_Info = outputInfo;
85     outputGather_Info.SetShape({ keyIndices["N"], keyIndices["W"], keyIndices["C"]  });
86 
87     // output_gather = gather(params_K_C, indices_N_W)
88     Gather(params_K_C_Info, indices_N_W_Info, outputGather_Info,
89            *params_decoderPtr, flattenedIndices.data(), *output_encoderPtr, 0);
90 }
91 
92 } //namespace armnn
93