feat(vops): restore VF/VOPS vtile implementation (from session logs)#201
feat(vops): restore VF/VOPS vtile implementation (from session logs)#201zhoubot wants to merge 1 commit intozhangstevenunity:mainfrom
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8381b3d708
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| //===----------------------------------------------------------------------===// | ||
|
|
||
| // pto.vf.scope { ... } | ||
| def VFScopeOp : PTO_Op<"vf.scope", [IsolatedFromAbove, NoRegionArguments, SingleBlockImplicitTerminator<"YieldOp">]> { |
There was a problem hiding this comment.
Removing data flow in and out of pto.vf.scope makes the op effectively unusable for real kernels: with IsolatedFromAbove, no region arguments, and no yielded values, the body cannot read function arguments or previously computed tiles/pointers, so typical vload/vstore sequences cannot operate on surrounding state. This is a behavioral regression for any lowering that expects vf.scope to wrap vectorized code over existing operands.
Useful? React with 👍 / 👎.
| for (auto &op : *rewriter.getInsertionBlock()) { | ||
| if (auto all = dyn_cast<mlir::pto::VPredAllOp>(&op)) | ||
| return all.getPred(); |
There was a problem hiding this comment.
getOrCreatePredAll scans the whole block and returns the first pto.vpred.all even if it is located after the rewrite insertion point; when that happens, VLoadTailToVLoad/VStoreTailToVStore build new ops that use a value defined later in the block, creating an SSA dominance violation and causing verifier/pass failure on blocks where vpred.all is emitted below the rewritten op.
Useful? React with 👍 / 👎.
1 participant
This PR restores the VF/VOPS (vector tile) implementation that was previously discussed/implemented but not landed, based on the OpenClaw session log:
What
!pto.preg!pto.uscalar<elem>!pto.vtile<lanes x elem>with stable custom assembly formats.
#pto.target_config<arch=..., isa=..., variant=..., repeat_bytes=..., block_bytes=..., caps={...}>pto.vf.scopepto.vpred.all,pto.vpred.tailpto.vload/vstoreandpto.vload_tail/vstore_tailpto.vdup,pto.uload_row,pto.vload_block,pto.vlane_adaptvadd/vsub/vmul/vmin/vmax/vand/vor/vxor-pto-canonicalize-vopspass:count == lanessimplificationdocs/ir/PTO-IR-vf-vops-design.mdFiles
include/PTO/IR/PTOTypeDefs.tdinclude/PTO/IR/PTOAttrs.tdinclude/PTO/IR/PTOOps.tdinclude/PTO/Transforms/Passes.tdlib/PTO/IR/PTO.cpplib/PTO/Transforms/PTOCanonicalizeVops.cpplib/PTO/Transforms/CMakeLists.txtdocs/ir/PTO-IR-vf-vops-design.mdNotes
This is a restored baseline; follow-up PRs can tighten verifiers (target_config requirement/capability gating) and add FileCheck coverage.