Collapse dims when producer is unpack op #17725
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
IanWood1
added
benchmarks:cuda
Abbreviated Benchmark Summary@ commit 793e4095b4558c71409842768805334c6cf56f1d (no previous benchmark results to compare) Data-Tiling Comparison TableClick to show
Raw Latencies
[Top 3 out of 92 results showed] No improved or regressed compilation metrics 🏖️ For more information: |
IanWood1
force-pushed
the
collapse_dims
branch
2 times, most recently
from
b0942ff
to
307e134
Compare
|
I haven't looked at the code yet, but it does address the compilation issue for #17530. Thanks for pushing on this! |
|
@hanhanW do you mean the large dispatch sizes? I'm not sure currently since the benchmarks are having issues If you can point me to something reproducible locally I can check it out |
There are no issues. This PR is good to me because it fixes my issues. :) I need to check benchmark result though, but at least we can compile all the models in my prototype. |
|
I think your PR breaks some models targeting gfx90a: https://github.com/iree-org/iree/actions/runs/9665068317/job/26661780089?pr=17725 The model and input files and be found at https://github.com/nod-ai/SHARK-TestSuite/tree/main/iree_tests/pytorch/models/sdxl-vae-decode-tank |
Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
ScottTodd
removed
benchmarks:android-cpu
hanhanW
reviewed
Jul 19, 2024
| @@ -507,51 +507,6 @@ util.func public @input_broadcast(%arg0: tensor<4x8xf32>, %arg1: tensor<4xf32>) | |||
|
|
|||
| // ----- | |||
|
|
|||
| // Do nothing if the dispatch is not a single elementwise op (with tensor.empty/linalg.fill producers) | |||
There was a problem hiding this comment.
What's going on in the test? Why do we delete it?
There was a problem hiding this comment.
It was there to make sure that collapsing only happened when there was one operation in the dispatch region. I removed it because I was planning on being able to collapse this, but had to disable collapsing when there is a producer generic.
There was a problem hiding this comment.
Okay, added this test back. Will remove again in a later pr that actually performs collapse here
| @@ -200,15 +227,6 @@ findRootGenericOp(DispatchRegionOp regionOp) { | |||
| } | |||
| } | |||
|
|
|||
| // Check that the operands of the generic op are defined outside the dispatch. | |||
There was a problem hiding this comment.
[no action required]: I don't understand why we had the check. But the changes that you made look reasonable to me.
Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
benvanik
pushed a commit
that referenced
this pull request
Jul 22, 2024
#17594 --------- Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
LLITCHEV
pushed a commit
to LLITCHEV/iree
that referenced
this pull request
Jul 30, 2024
iree-org#17594 --------- Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu> Signed-off-by: Lubo Litchev <lubol@google.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
#17594