src/backends/neon/workloads/NeonGatherNdWorkload.cpp

//
// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "NeonGatherNdWorkload.hpp"
#include "NeonWorkloadUtils.hpp"
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include "backendsCommon/WorkloadUtils.hpp"

namespace armnn
{
arm_compute::Status NeonGatherNdWorkloadValidate(const TensorInfo& paramInfo,
                                                 const TensorInfo& indicesInfo,
                                                 const TensorInfo& outputInfo)
{
    // Calculate ND, K, W, C.
    std::map<std::string, unsigned int> keyIndices = CalculateGatherNdKeyIndices(paramInfo, indicesInfo);

    /// Call Gather with adequate shapes
    // Reshape params into { K, C }
    armnn::TensorInfo params_K_C_Info =  paramInfo;
    params_K_C_Info.SetShape({ keyIndices["K"], keyIndices["C"] });

    // Reshape indices into { W }
    armnn::TensorInfo indices_W_Info = indicesInfo;
    indices_W_Info.SetShape({ keyIndices["W"] });

    // Reshape output to have the shape given by gather { W, C }
    // (the original outputInfo has the shape given by gatherNd)
    armnn::TensorInfo outputGather_Info = outputInfo;
    outputGather_Info.SetShape({ keyIndices["W"], keyIndices["C"] });

    const arm_compute::TensorInfo aclParamsInfo  = BuildArmComputeTensorInfo(params_K_C_Info);
    const arm_compute::TensorInfo aclIndicesInfo = BuildArmComputeTensorInfo(indices_W_Info);
    const arm_compute::TensorInfo aclOutputInfo  = BuildArmComputeTensorInfo(outputGather_Info);

    auto aclAxis = ComputeAclAxis(0, params_K_C_Info);
    return arm_compute::NEGather::validate(&aclParamsInfo, &aclIndicesInfo, &aclOutputInfo, aclAxis);
}

NeonGatherNdWorkload::NeonGatherNdWorkload(const GatherNdQueueDescriptor& descriptor,
                                           const WorkloadInfo& info)
        : NeonBaseWorkload<GatherNdQueueDescriptor>(descriptor, info)
{
    m_Data.ValidateInputsOutputs("NeonGatherNdWorkload", 2, 1);

    TensorInfo paramsInfo  = info.m_InputTensorInfos[0];
    TensorInfo indicesInfo = info.m_InputTensorInfos[1];
    TensorInfo outputInfo  = info.m_OutputTensorInfos[0];

    arm_compute::ITensor& input   = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& indices = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ITensor& output  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();

    // Calculate ND, K, W, C.
    std::map<std::string, unsigned int> keyIndices = CalculateGatherNdKeyIndices(paramsInfo, indicesInfo);

    /// Calculate flattened indices: m_FlattenedIndices = indices * m_FlattenedCoeff.
    /// This could be done using MatMul instead of multiplication followed by reduce sum operation,
    /// but GeMM does not support s32 at the moment.

    // Prepare the tensor to store the output of the reduce_sum operation
    armnn::TensorInfo flattenedIndices_Info = indicesInfo;
    flattenedIndices_Info.SetShape({ keyIndices["W"] });
    BuildArmComputeTensor(m_FlattenedIndices, flattenedIndices_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_FlattenedIndices);

    // Reshape indices into { W, ND }
    indices.info()->set_tensor_shape(BuildArmComputeTensorShape({ keyIndices["W"], keyIndices["ND"] }));

    // Calculate the m_FlattenedCoeff
    TensorShape paramsShape = paramsInfo.GetShape();
    std::vector<unsigned int> flattenedCoeff(keyIndices["ND"], 1);
    for (unsigned int i = 1; i < keyIndices["ND"]; ++i)
    {
        flattenedCoeff[i - 1] = paramsShape[i];
    }
    for (unsigned int i = keyIndices["ND"] - 1; i > 0; --i)
    {
        flattenedCoeff[i - 1] *= flattenedCoeff[i];
    }
    armnn::TensorInfo flattenedCoeff_Info = indicesInfo;
    flattenedCoeff_Info.SetShape({ keyIndices["ND"] });
    BuildArmComputeTensor(m_FlattenedCoeff, flattenedCoeff_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_FlattenedCoeff);
    CopyArmComputeITensorData(flattenedCoeff.data(), m_FlattenedCoeff);

    // Prepare the tensor to store the output of the multiplication
    armnn::TensorInfo outputMul_Info = indicesInfo;
    outputMul_Info.SetShape({ keyIndices["W"], keyIndices["ND"] });
    BuildArmComputeTensor(m_outputMul, outputMul_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_outputMul);

    // Multiply
    auto convertPolicy = (IsQuantizedType(info.m_InputTensorInfos[0].GetDataType()) ||
                          IsQuantizedType(info.m_InputTensorInfos[1].GetDataType())) ?
                          arm_compute::ConvertPolicy::SATURATE :
                          arm_compute::ConvertPolicy::WRAP;

    m_MulLayer.configure(&indices,
                         &m_FlattenedCoeff,
                         &m_outputMul,
                         1.0f,
                         convertPolicy,
                         arm_compute::RoundingPolicy::TO_ZERO,
                         arm_compute::ActivationLayerInfo());

    // Reduce Sum
    const std::vector<unsigned int> armnnReduceAxes(1, 1);
    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(m_outputMul.info()->num_dimensions(),
                                                                          outputMul_Info.GetNumDimensions(),
                                                                          armnnReduceAxes);
    m_ReduceSumLayer.configure(&m_outputMul,
                               &m_FlattenedIndices,
                               static_cast<unsigned int>(coords[0]),
                               arm_compute::ReductionOperation::SUM,
                               false);

    /// Call Gather with adequate shapes
    // Reshape params into { K, C }
    paramsInfo.SetShape({ keyIndices["K"], keyIndices["C"] });
    input.info()->set_tensor_shape(BuildArmComputeTensorShape(paramsInfo.GetShape()));

    // Reshape output to have the shape given by gather { W, C }
    // (the original outputInfo has the shape given by gatherNd)
    armnn::TensorInfo outputGather_Info = outputInfo;
    outputGather_Info.SetShape({ keyIndices["W"], keyIndices["C"] });
    BuildArmComputeTensor(m_outputGather, outputGather_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_outputGather);

    m_GatherLayer.configure(&input, &m_FlattenedIndices, &m_outputGather, ComputeAclAxis(0, paramsInfo));

    // Reshape output to the original output shape
    m_ReshapeLayer.configure(&m_outputGather, &output);
}

void NeonGatherNdWorkload::Execute() const
{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonGatherNdWorkload_Execute", this->GetGuid());
    m_MulLayer.run();
    m_ReduceSumLayer.run();
    m_GatherLayer.run();
    m_ReshapeLayer.run();
}
} //namespace armnn