These are chat archives for halide/Halide

7th
Nov 2018
Steven Johnson
@steven-johnson
Nov 07 17:25
@abadams — syncing llvm-trunk all the way back to a mid-July commit still crashes the arm32 jit at recent Halide top of tree. Now I’m wondering if the injection is actually in Halide, not LLVM. Gonna try bisecting there.
Andrew Adams
@abadams
Nov 07 17:25
I think it has been broken for at least that long
Steven Johnson
@steven-johnson
Nov 07 17:26
really? ok, I’ll try going back further
Andrew Adams
@abadams
Nov 07 17:28
Searching gitter history, on May 8 I said "Anyone know what's going on with the arm-32 buildbot? It has been failing for trunk for a month."
and I think it was never resolved
Steven Johnson
@steven-johnson
Nov 07 17:29
yow
how did we miss that
Andrew Adams
@abadams
Nov 07 17:29
So try early april I guess?
It was in the pile of buildbot shit that nobody wanted to deal with
Given that it was trunk, I guess I was hoping that if I ignored it long enough the problem would be fixed upstream
Pranav Bhandarkar
@pranavb-ca
Nov 07 17:53
@steven-johnson to build Halide so that I have distrib generated, is this the prescribed method? MSBuild.exe /m /t:distrib /p:Configuration=Release .\distrib.vcxproj I couldn't really tell where the "hand-rolled" approach starts in the buildbot.
Steven Johnson
@steven-johnson
Nov 07 17:53
that looks right
(but it’s been a while so I may be wrong)
Pranav Bhandarkar
@pranavb-ca
Nov 07 17:53
what is the hand-rolled way?
Steven Johnson
@steven-johnson
Nov 07 17:54
Look at master.cfg starting at line 543 or so (in the buildbot)
Pranav Bhandarkar
@pranavb-ca
Nov 07 17:55
Got it, thanks!
so, it looks like the hand-rolled version doesn't bother about halide_config.*
Steven Johnson
@steven-johnson
Nov 07 17:59
yeah, it’s clearly long out of date
Pranav Bhandarkar
@pranavb-ca
Nov 07 17:59
how do halide/apps work on windows then? because AFAICT they need HALIDE_DISTRIB_PATH/halide_config.make
Steven Johnson
@steven-johnson
Nov 07 18:00
Windows uses CMake, not make
And CMake builds all the apps as part of a single monolithic build step, basically
Pranav Bhandarkar
@pranavb-ca
Nov 07 18:01
ok
Steven Johnson
@steven-johnson
Nov 07 18:31
On my mac, recent simd_op_check is failing with recent-ish llvm trunk with “psubusw did not generate”. Haven’t investigated yet.
Steven Johnson
@steven-johnson
Nov 07 21:20
@abadams fyi: an llvm-trunk from mid-March also fails.
Andrew Adams
@abadams
Nov 07 22:20
@shoaibkamil did you get anywhere with fixing licm? It's now messing up one of my pipelines in the autoscheduler project so I was going to take a look if you don't already have a fix.
Steven Johnson
@steven-johnson
Nov 07 22:33
Injection point for the simd_op_check failure is https://reviews.llvm.org/rL345602
Andrew Adams
@abadams
Nov 07 22:35
@shoaibkamil nope, the smoking gun test case I thought I had evaporated when I removed a hack I was testing, so now I have no repro for LICM problems. LMK if you have a simple repro
Steven Johnson
@steven-johnson
Nov 07 22:38
(I’m sure it’s just another variant of halide/Halide#3281 — disabling with a TODO is likely the only short-term fix)