What is the main function of ggml_backend_graph_compute() in clip_image_batch_encode?

[perp]

In llama cpp the function ggml_backend_graph_compute() cost the most time of load time, What this function is actually doing?