Read a profile before changing the code

Synthetic English narration.

Start with a shared clock
Begin with one shared clock. The host launches work, the GPU executes it, and the host eventually receives the result. This timeline is schematic.

Waiting can overlap useful work
While the GPU computes, the host can be waiting for that same work. Adding both durations double counts overlapping time. Read the two lanes together.

Locate the expensive region
In the saved mixed-backend profile, FP64 share correction accounts for about sixty-four percent of recorded GPU activity. This is not a wall-time percentage.

Inspect the hotspot
That hotspot tells us where to inspect launches, arithmetic and intermediate memory. It does not say that all its time can be removed.

Turn the profile into an experiment
Change one implementation detail, validate the complete answer, then repeat clean timings. The profile guides the experiment. Unprofiled complete runs determine whether it helped.