The same three decode steps without and with overlap
Two timelines with a GPU lane and a CPU lane. Without overlap, each GPU step is followed by a hatched gap while the CPU processes its results, and three steps end at 500. With overlap, the GPU runs steps back to back while the CPU processes the previous step under the next one, and both lanes finish three steps at 420.
WITHOUT OVERLAP: EVENT_LOOP_NORMAL
GPU
CPU
WITH OVERLAP: EVENT_LOOP_OVERLAP
GPU
CPU
t
step 1
step 2
step 3
three steps by 500
the GPU waits for the CPU
step 1
step 2
step 3
three steps by 420
the CPU processes step N under step N+1