[MetaSchedule] Support RewriteLayout postproc on AllocateConst #12991
Conversation
masahi
force-pushed
the
metasch-rewrite-layout-constant
branch
from
bd0ea26 to
a3eba9b
Compare
masahi
force-pushed
the
metasch-rewrite-layout-constant
branch
from
a3eba9b to
1134c5a
Compare
masahi
force-pushed
the
metasch-rewrite-layout-constant
branch
2 times, most recently
from
26df009 to
a25a3d3
Compare
|
There are still some issues with Hexagon tests, debugging. |
| ) | ||
| return bool(result & 1) |
There was a problem hiding this comment.
Fixes in this file are probably necessary after the MS API change PR - without them I got a compile error and segfault using compile_relay. cc @junrushao
masahi
force-pushed
the
metasch-rewrite-layout-constant
branch
from
bf6f2da to
074f48a
Compare
| if executor is None: | ||
| executor = relay.backend.Executor("graph") | ||
|
|
||
| if mod.get_attr("executor") is None: | ||
| mod = mod.with_attr("executor", executor) | ||
| else: | ||
| executor = mod.get_attr("executor") |
There was a problem hiding this comment.
quick question: why do we need the logic on line 97-98? i.e. is non-nullable executor useful in subsequent processing?
There was a problem hiding this comment.
When None executor is passed to relay.build(..., executor=executor), we get segfault. This is certainly due to some sloppy coding somewhere, but I haven't looked into it.
Perhaps the default value of executor should be Executor("graph") rather than None, as in relay.build https://github.com/apache/tvm/blob/main/python/tvm/relay/build_module.py#L281
There was a problem hiding this comment.
I see! This is certainly annoying that it’s engineered this way…Thanks for your explanation!
We probably don’t need to over-engineer it for now. Let’s keep it as it is given vm is not supported.
…e#12991) * [Metaschedule] Support AllocateConst in RewriteLayout * fix * convert constant back to original before trace apply * improved CreatePrimFunc change * use IndexMap to transform constant in RemoveLayoutRewriteBlock * wip * wip * add comments * add test * enable RewriteLayout on Hexagon * lint * improve RewriteLayout postproc * fix compiler warning * fixed CreatePrimFunc test * black * Fix after MS API change * fixed auto tensorize test after MS API PR * add missing Clone method to MultilevelTilingWideVector * fixed hexagon dense MS test * fixed the rest of Hexagon MS tests
RewriteLayoutassumes that the "layout-free" weight to be rewritten is passed as a parameter to prim func. This doesn't hold whenlink-params = Trueis used, since the constant weight is directly embedded in a primfunc asAllocateConstnode.Supporting RewriteLayout on AllocateConst required some nasty hacks to workaround various issues. Detailed comments are provided to explain the problem and my solution.
This lets us enable
RewriteLayoutfor Hexagon.