feat(ptodsl): expose pto.section("cube") and pto.section("vector") interface as an explicit section hint - #1041
Open
Zhendong404 wants to merge 7 commits into
Open
Conversation
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
from
July 29, 2026 06:46
21bd65b to
b52e71a
Compare
jimmychou0
reviewed
Jul 29, 2026
|
|
||
| if (*source == AddressSpace::ACC) | ||
| return InferredSectionKind::Cube; | ||
| if (*source == AddressSpace::VEC || *destination == AddressSpace::VEC) |
Contributor
There was a problem hiding this comment.
UB -> L1 和 L1 -> UB 会归为 Vector， 是否会有问题
| outer = Module.create() | ||
| outer.operation.attributes["pto.target_arch"] = StringAttr.get(spec.target_arch) | ||
|
|
||
| if spec.kernel_kind is None: |
Collaborator
Author
There was a problem hiding this comment.
已增加对simt kernel的适配
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
3 times, most recently
from
July 30, 2026 06:48
07ffcc9 to
5f34398
Compare
Collaborator
Author
|
/run a3 |
Collaborator
Author
|
/run a5 |
|
已接收
页面会自动刷新，可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
日志尾部 |
|
已接收
页面会自动刷新，可以直接看当前阶段、排队情况和最近结果。 |
A3 板测失败
失败用例
|
A3 板测失败详情：PR #1041syncall_binding
comm_p2p_binding_variants
comm_p2p
|
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
2 times, most recently
from
July 30, 2026 15:09
ede4431 to
3906634
Compare
Collaborator
Author
|
/run a5 |
|
已接收
页面会自动刷新，可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
日志尾部 |
Collaborator
Author
|
/run a5 |
|
已接收
页面会自动刷新，可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
日志尾部 |
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
from
July 31, 2026 00:45
1818aa8 to
c93e90d
Compare
Collaborator
Author
|
/run a5 |
|
已接收
页面会自动刷新，可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
日志尾部 |
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
from
July 31, 2026 01:05
c93e90d to
5b7fa20
Compare
jimmychou0
approved these changes
Jul 31, 2026
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
2 times, most recently
from
July 31, 2026 07:19
c37805f to
2b7feda
Compare
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
2 times, most recently
from
July 31, 2026 14:20
9d0d6a5 to
8fe48c3
Compare
Zhendong404
force-pushed
the
feature-dsl-section-hint
branch
from
August 1, 2026 15:24
8fe48c3 to
bb5966f
Compare
| for (func::FuncOp funcOp : childModule.getOps<func::FuncOp>()) | ||
| funcs.push_back(funcOp); | ||
| }); | ||
| moduleOp.walk([&](func::FuncOp funcOp) { funcs.push_back(funcOp); }); |
Collaborator
There was a problem hiding this comment.
The same fix was not applied to PTOPlanMemoryModern.cpp.
--plan-memory-impl defaults to legacy (tools/ptoas/ptoas.cpp:416), so this pass is the default path and the fix lands where it matters. But lib/PTO/Transforms/PTOPlanMemoryModern.cpp:1653-1666 still carries exactly the block that is deleted here:
moduleOp.walk([&](ModuleOp childModule) {
if (childModule == moduleOp)
return;
bool hasTileOpHelper = llvm::any_of(
childModule.getOps<func::FuncOp>(), [](func::FuncOp funcOp) {
return funcOp->hasAttr("pto.tileop.helper");
});
if (!hasTileOpHelper)
return; // <-- the new child shape is skipped here
for (func::FuncOp funcOp : childModule.getOps<func::FuncOp>())
funcs.push_back(funcOp);
});With _DEFAULT_KERNEL_KIND now None, a pto.backend child module commonly carries neither pto.kernel_kind nor any pto.tileop.helper function. Whenever the driver does not promote it to a top-level compile unit (isUserVisibleIROutputRequested() modes, or direct pto-test-opt invocation), --plan-memory-impl=modern hits the if (!hasTileOpHelper) return; early exit and the child's functions are never planned - silently, with no diagnostic. That is the same failure mode this hunk fixes for the legacy planner.
The new regression test test/lit/pto/nested_resolve_reserved_buffers.pto runs --pto-plan-memory, i.e. the legacy planner only, so the modern planner has no coverage for the nested shape either.
Is the modern planner intentionally kept on the old behaviour? If so a short comment there would help. Otherwise the same moduleOp.walk(...) change plus an extra --plan-memory-impl=modern RUN line on the new lit test would close the gap.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Summary
pto.mioperations.pto.section("cube")andpto.section("vector")as explicit section hints.Motivation
Mixed Cube/Vector kernels require each operation region to be assigned to a physical section. PTOAS can usually infer the section from the operation kind, pipeline, or buffer address space.
However, the current inference cannot determine the section of
set_flag/wait_flagoperations betweenPIPE_MTE2andPIPE_S. Both pipelines are available on the Cube and Vector cores, so the pipe pair alone does not provide enough information to identify the intended physical section.Silently choosing either section could result in incorrect placement. Therefore, when such an ambiguous region cannot be classified from its surrounding operations, PTOAS now reports an explicit error and asks the user to provide a section hint:
Explicit section hints are intended as an escape hatch for ambiguous regions. Regions that can be classified unambiguously continue to use automatic inference.
Changes
pto.mioperations.pto.section("cube")andpto.section("vector")PTODSL context manager.PIPE_MTE2↔PIPE_Ssynchronization.