Make sure IterDomain::resize generates non-symbolic IDs

[naoyam]

Just realized in some cases replaying resize resulted in generating symbolic IDs, even when it's guaranteed to be non-symbolic. Didn't see any particular error, likely just because it gets masked somehow, but we should not generate symbolic IDs unnecessarily.

[naoyam]

!test

[github-actions]

Review updated until commit 809b583

Description

• Added iter_type parameter to resize methods

• Ensured non-symbolic IDs in IterDomain::resize

• Added error check for symbolic IDs in ReplayForwardTransformOnLoopDomain

---

Changes walkthrough 📝

Relevant files

Enhancement

nodes.cpp Add iter_type to TensorDomain::resize                                        csrc/ir/nodes.cpp Added iter_type parameter to TensorDomain::resize method Passed iter_type to IterDomain::resize call +8/-2      loop_domain_scheduler.cpp Handle symbolic IDs in ReplayForwardTransformOnLoopDomain csrc/scheduler/tools/loop_domain_scheduler.cpp Added error check for symbolic IDs in ReplayForwardTransformOnLoopDomain::handle Passed iter_type to tv_->resize call +7/-1      tensor_view.cpp Add iter_type to TensorView::resize                                            csrc/tensor_view.cpp Added iter_type parameter to TensorView::resize method Passed iter_type to domain()->resize call +3/-2      interface_nodes.h Add iter_type to TensorView::resize signature                        csrc/ir/interface_nodes.h Added iter_type parameter to TensorView::resize method +5/-1      internal_base_nodes.h Add iter_type to TensorDomain::resize signature                    csrc/ir/internal_base_nodes.h Added iter_type parameter to TensorDomain::resize method +5/-1

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Possible Issue The new iter_type parameter is added, but it is not clear if all call sites of TensorDomain::resize have been updated to provide this parameter. This could lead to unexpected behavior if some call sites still pass only two arguments. void TensorDomain::resize( int64_t axis, Val* left_expansion, Val* right_expansion, std::optional<IterType> iter_type) { NVF_ERROR(nDims() > 0, "Tried to do resize on a 0-dim domain"); axis = wrapDim(axis); IterDomain* id = this->axis(axis); auto resized_id = IterDomain::resize( id, left_expansion, right_expansion, /*mark_as_rfactor=*/false, iter_type); loop_domain_.at(axis) = resized_id; Code Clarity The comment // Pass the iter type explicitly to avoid generating a symblic ID contains a typo (symblic instead of symbolic). This should be corrected for clarity. // Pass the iter type explicitly to avoid generating a symblic ID tv_->resize( getLoopIdPosition(input_loop_ids_.at(0)), resize->leftExpand(), resize->rightExpand(), resize->out()->getIterType()); } Code Clarity The comment // Pass the iter type explicitly to avoid generating a symblic ID contains a typo (symblic instead of symbolic). This should be corrected for clarity. domain()->resize(axis, left_expansion, right_expansion, iter_type); return this;

[jacobhinkle]

LGTM

[jacobhinkle]

Nit:

Suggested change

	IterDomain::resize(id, left_expansion, right_expansion, false, iter_type);
	IterDomain::resize(id, left_expansion, right_expansion, /*mark_as_rfactor=*/false, iter_type);

[naoyam]

!build

[naoyam]

!build