kernel_info_t *gpgpu_sim::select_kernel()
{
    for(unsigned n=0; n < m_running_kernels.size(); n++ ) {
        unsigned idx = (n+m_last_issued_kernel+1)%m_config.max_concurrent_kernel;
        if( m_running_kernels[idx] && !m_running_kernels[idx]->no_more_ctas_to_run() ) {
            m_last_issued_kernel=idx;
            unsigned launch_uid = m_running_kernels[idx]->get_uid();
            if (std::find(m_executed_kernel_uids.begin(), m_executed_kernel_uids.end(), launch_uid) == m_executed_kernel_uids.end()) {
               m_executed_kernel_uids.push_back(launch_uid);
               m_executed_kernel_names.push_back(m_running_kernels[idx]->name());
            }

            return m_running_kernels[idx];
        }
    }
    return NULL;
}