Fix operation ordering issue during inlining of operations into dispatch region #5236
Conversation
|
@benvanik This change could potentially be done in the canonicalizer, but I got a bit lost in the interfaces and wasnt sure if it is actually worth having so much stress on the canonicalizers so as to pull in operands with operands into the dispatch region as well. So I aborted that plan, and just fixed the ordering issue while pulling the ops into the dispatch region during creation itself. |
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
| std::swap(nextReadyOps, readyOps); | ||
| } | ||
|
|
||
| LLVM_DEBUG({ |
There was a problem hiding this comment.
this is a lot of debug spew (here and above) - is this enabled with -debug by default?
There was a problem hiding this comment.
I guess so. Is there a way to hide it even further. I can remove it if it is too much.
There was a problem hiding this comment.
Using DEBUG_WITH_TYPE. That makes the debug spew guarded by DEBUG_TYPE.
iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_dynamic.mlir
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/test/dispatch_linalg_on_tensors_dynamic.mlir
Outdated
Show resolved
Hide resolved
There was a problem hiding this comment.
Overall looks fine - can always extract it out later on if we want. Am concerned about the consistent inversion of argument ordering, though - if it's a reasonable thing to fix that'd be really nice!
I think this is because of the readability of the IR. I saw a comment in the code:
// Reverse the values. This is not for correctness, but more for readability
// of the IR.
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
I put this |
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
iree/compiler/Dialect/Flow/Transforms/DispatchLinalgOnTensors.cpp
Outdated
Show resolved
Hide resolved
| llvm::dbgs() << "\n"; | ||
| } | ||
| }); | ||
| assert(orderedOps.size() == ops.size()); |
There was a problem hiding this comment.
Maybe add an error message for the assertion?
| @@ -443,6 +523,18 @@ static void tryToTieOperandsAndResults( | |||
| } | |||
| } | |||
|
|
|||
| void replaceAllUsesWithinDispatchOp(IREE::Flow::DispatchWorkgroupsOp dispatchOp, | |||
There was a problem hiding this comment.
I think this is outside anonymous namespace since other methods are explicitly static. Please also add static to this method.
MaheshRavishankar
force-pushed
the
issue_5151
branch
from
cc18320
to
c701be3
Compare
…tch region. In general the operations cloned into a dispatch region could form a DAG. These operations have to be cloned while keeping the order amongst them consistent to not violate use-def chains. This changes adds a method to clone the operations in the right order. Also cleans up the dispatch region creation code. Fixes iree-org#5151.
MaheshRavishankar
force-pushed
the
issue_5151
branch
from
c701be3
to
afe0ac0
Compare
In general the operations cloned into a dispatch region could form a
DAG. These operations have to be cloned while keeping the order
amongst them consistent to not violate use-def chains. This changes
adds a method to clone the operations in the right order. Also cleans
up the dispatch region creation code.
Fixes #5151.