These are chat archives for halide/Halide

Apr 2018
Steven Johnson
Apr 27 2018 17:44
@abadams re: my comment from yesterday about a tracing bug in vectorizeloops, yeah, it’s definitely bogus in at least one case I’ve seen: we have a sum() of a vectorized func (e.g. kernel_sum_x in apps/resize); the special-casing for Call::trace expands this in a weird way, so that a single begin_realization that should be 1-dimensional becomes n-dimensional, but with all the mins first followed by all the extents. (We can work around this in HTV but it seems wonky.)
(same bug applies to produce/consume events too, we just don’t ever use the coordinates from those)