[SYCL][ESIMD] Rework -O0 behavior #12612
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
sarnex
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sarnex
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sarnex
force-pushed
the
o0
branch
2 times, most recently
from
cdf0cb2 to
6c6c20a
Compare
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Signed-off-by: Sarnie, Nick <[email protected]>
|
Linux Gen12 failure is not related, failing everywhere: |
| @@ -955,6 +955,8 @@ void EmitAssemblyHelper::RunOptimizationPipeline( | |||
| PB.registerPipelineStartEPCallback([&](ModulePassManager &MPM, | |||
| OptimizationLevel Level) { | |||
| MPM.addPass(ESIMDVerifierPass(LangOpts.SYCLESIMDForceStatelessMem)); | |||
| if (Level == OptimizationLevel::O0) | |||
| MPM.addPass(ESIMDRemoveOptnoneNoinlinePass()); | |||
There was a problem hiding this comment.
I am okay with this ESIMD change.
premanandrao
approved these changes
Feb 14, 2024
mdtoguchi
approved these changes
Feb 16, 2024
smanna12
approved these changes
Feb 16, 2024
|
@intel/dpcpp-esimd-reviewers @intel/dpcpp-tools-reviewers Ping, thanks! |
Signed-off-by: Sarnie, Nick <[email protected]>
v-klochkov
reviewed
Feb 28, 2024
| @@ -1834,6 +1834,15 @@ bool SYCLLowerESIMDPass::prepareForAlwaysInliner(Module &M) { | |||
| continue; | |||
| } | |||
|
|
|||
| // If we are splitting by ESIMD, we are guarenteed the entire | |||
There was a problem hiding this comment.
Is this code still needed after adding the special pass that remove those 2 attributes?
There was a problem hiding this comment.
Yes, I tested many combinations of with/without different changes in this PR and everything we currently have is required.
sarnex
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
sarnex
had a problem deploying
to
WindowsCILock
GitHub Actions
Failure
v-klochkov
approved these changes
Feb 28, 2024
|
Windows fail is not related |
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.
Let me first say I don't like this change and I wish we didn't have to do this, and it should only be temporary.
The root need is that the IGC VC backend does not support O0 code. It gets miscompiled and produces wrong answers and crashes. The team acknowledges this issue but does not have the resources to fix this at the moment.
From the ESIMD user POV, this is really frustrating. We decided it would be okay if they lost debuggability of device code (but retained it for host code) as long as their program runs and they get the right answer.
So the overall approach is to optimize ESIMD(Not SYCL!!) code even in O0 mode. We are actually already doing this, but this change extends it a bit.
So there are a few main changes in this PR:
-O0tosycl-post-link. Note this option has absolutely no effect on SYCL code today, it is only checked in the ESIMD lowering code insycl-post-link.optnone/noinlineearly for ESIMD code inO0.optnone/noinlinefrom all functions if we split out ESIMD/SYCL code, and if we don't don't touch it.All three of these steps are required to fix the tests failing with
O0, if we remove any one of them we see many failures.When running all ESIMD tests with
O0, we see the following improvements:Once IGC VC can actually correctly compile non-optimized code, we should revert this change.