We used to use the blitter for that and then there was a memory location to write to to initiate execution of what was uploaded. The Jaguar developer docs are available online if you haven't seen them.

Once the GPU or the DSP was running. You could then chain overlays together by running to a location in the GPU itself from which you called the blitter to upload the next segment and then you could just jump to its first instruction. I believe later compilers enabled just running the GPU from main RAM but it has been decades.

It was unwittingly extremely useful mental preparation for programming in cuda a decade later.