a blit’s plain copy (XRGB source) from this many bytes on runs in row bands on the loop of
set_parallel (gfxtest –bench-blit: a 1 MB copy takes about a fifth on 12 workers)
… and from this many on with streaming stores (gfx::blend::copy_stream): source and target
together no longer fit the caches, the target is not read first. A 4K frame (32 MB): 1.9 ms on
12 workers and the calling thread, 2.7 ms with plain stores, 5 ms on one thread (Ryzen 5 3600,
DDR4-3200, a loaded machine; the memory’s limit for such a copy is about 1.7 ms there)
the loop large copies run on (Dc::blit, COPY_PAR_MIN); the UI runtime passes a wrapper of
plat::parallel_for – gfx does not depend on plat. Without it they run on the calling thread
(the same bytes)
n zeroed bytes | None when the memory is not there (instead of an abort). A probe
reservation first, then vec![0; n]: that is alloc_zeroed, whose fresh pages need no
clearing pass.