Method run

src/runtime/NEON/functions/NELSTMLayer.cpp:774–881

}

void NELSTMLayer::run()
{
    ARM_COMPUTE_TRACE_EVENT(ARM_COMPUTE_PROF_CAT_CPU, ARM_COMPUTE_PROF_LVL_CPU, "NELSTMLayer::run");
    prepare();

    MemoryGroupResourceScope scope_mg(_memory_group);

    _concat_inputs_forget_gate.run();
    _fully_connected_forget_gate.run();

    if (_run_peephole_opt)
    {
        _pixelwise_mul_forget_gate.run();
        _accum_forget_gate1.run();
    }
    if (_is_layer_norm_lstm)
    {
        _mean_std_norm_forget_gate.run();
        _pixelwise_mul_forget_gate_coeff.run();
        _accum_forget_gate_bias.run();
    }
    _activation_forget_gate.run();

    if (_run_cifg_opt)
    {
        if (_ones.info()->data_type() == DataType::F16)
        {
            std::fill_n(reinterpret_cast<half *>(_ones.buffer()),
                        _ones.info()->total_size() / _ones.info()->element_size(), 1);
        }
        else
        {
            std::fill_n(reinterpret_cast<float *>(_ones.buffer()),
                        _ones.info()->total_size() / _ones.info()->element_size(), 1);
        }
        _subtract_input_gate.run();
    }
    else
    {
        _fully_connected_input_gate.run();

        if (_run_peephole_opt)
        {
            _pixelwise_mul_input_gate.run();
            _accum_input_gate1.run();
        }

        if (_is_layer_norm_lstm)
        {
            _mean_std_norm_input_gate.run();
            _pixelwise_mul_input_gate_coeff.run();
            _accum_input_gate_bias.run();
        }
        _activation_input_gate.run();
    }

    _fully_connected_cell_state.run();
    _transpose_cell_state.run();