Fix corner case in if_then_else simplification #8189
Conversation
|
Ready to land? |
|
Hi, I wonder if this fix will eventually be included in binaries that I can download from: https://github.com/halide/Halide/releases If so, is there an estimation when it will become publicly available? Thank you! |
|
There are one or two bugs that I would like to close out before we next do an official release, so it'll be a few more weeks. In the mean-time you can download a nightly build from https://buildbot.halide-lang.org/ These are the same as the releases. For the release artifacts we just grab the latest nightlies. |
|
@abadams Hi Andrew, thank you for your quick response! However, I'm still not very clear which item on https://buildbot.halide-lang.org/ is the nightly build you were referring to. I tried 'Halide-17.0.1-x86-64-linux-52541176253e74467dabc42eeee63d9a62c199f6.tar' on that page, but the issue I reported in #8170 seems still over there. Am I downloading the wrong item on that page? Thank you. |
|
I got more experimental result. |
Fixes the segfault reported in #8170
I have no idea how to repro this in a test. It came about as a result of applying the adams autoscheduler to an hvx pipeline, which produced a schedule with vectors wider than tiles, which turned into predication, which was then turned into something invalid inside the hexagon lowering path.
The if_then_else intrinsic is supposed to evaluate to zero if the condition is false and there's no false value defined.