Autonomous　driving　poses　high　demands　on　computing　and　communication　resources.　Vehicular　edge　computing　is　presented　to　offload　real-time　computing　tasks　from　connected　and　automated　vehicles　(CAVs)　to　high-performance　edge　servers.　However,　it　brings　additional　communication　overhead　due　to　limited　bandwidth,　and　increases　delay　of　tasks.　To　solve　it,　this　work　first　proposes　an　offloading　architecture　including　multiple　CAVs,　roadside　units　and　cloud.　We　minimize　the　total　cost　of　a　hybrid　system　by　jointly　considering　task　offloading　ratios,　and　allocation　of　communication　and　computing　resources.　Furthermore,　a　mixed　integer　non-linear　program　is　formulated　and　solved　by　a　novel　meta-heuristic　algorithm　called　Self-adaptive　Gray　Wolf　Optimizer　with　Genetic　Operations　(SGWOGO).　SGWOGO　achieves　joint　optimization　of　computation　offloading　among　CAVs,　roadside　units　and　cloud,　and　allocation　of　their　resources.　Finally,　real-life　data-driven　simulation　results　demonstrate　that　SGWOGO　achieves　lower　cost　in　fewer　iterations　compared　with　its　several　state-of-the-art　peers.　　©　2022　IEEE.