GeezerVison

macOS fails a command buffer with "Impacting Interactivity" when its thread
groups do not finish in time for the GPU to be handed to another context.
The kernels launched with a fixed grid loop over their work with the grid as
stride, so with few blocks every thread group runs for as long as the
dispatch. At 30 kHz Pb-Pb that is too long: with 4 sectors of a time frame
the tracklet constructor, running for several sectors at once on 32 or 64
blocks, is killed after 50 to 200 ms in every run, and passes from 128
blocks on; on the full time frame the looper following, 512 groups of one
thread, is killed after 0.5 s. The track fit, 3.1 s in one dispatch of 8236
short blocks, is not.

Launch the tracklet constructor on 1024 blocks, the looper following on
2048 blocks of 32 threads (2 to 3 times faster than 512 single threads),
and the sector refit, which also runs for several sectors at once, on 256.
The compression of the unattached clusters gets 2048 blocks, from its time
on a part of the time frame: it was not reached on a full one yet.