[TOPI] update depthconv padding api; fix shared memory overflow by Huyuwei · Pull Request #365 · apache/tvm

Huyuwei · 2017-08-21T09:57:45Z

No description provided.

The below compilation warning is fixed. In file included from /mnt/D_DRIVE/work/nnvm_8_feb/dmlc-core/include/dmlc/any.h:16:0, from include/nnvm/./base.h:11, from include/nnvm/graph.h:15, from src/compiler/fold_scale_axis.cc:6: src/compiler/fold_scale_axis.cc: In function ‘nnvm::Graph nnvm::compiler::FoldScaleAxis(nnvm::Graph)’: src/compiler/fold_scale_axis.cc:155:39: warning: enum constant in boolean context [-Wint-in-bool-context] CHECK(kind == kPassTroughFirst || kMulConsumer); ^ /mnt/D_DRIVE/work/nnvm_8_feb/dmlc-core/include/dmlc/./logging.h:110:9: note: in definition of macro ‘CHECK’ if (!(x))

* Update RunCodegen pass * update TensorRT codegen for the new interface * typo fix * fix unused var warning * leave todo comment on dep on RemoveUnusedfunctions * introduce src/tranform/utils.h * also move GetExtSymbol to transform/utils.h * build fix * add missing file

Add compiled structural visit, mutate, and maybe-in-place-mutate hooks for every Expr and Stmt node family using the current tvm-ffi structural-hook ABI. The branch pins tvm-ffi at `9d784c4da74ff7360d76c79a89fe60a63516f880`, including the callback-owned `StructuralMutate` API from tvm-ffi apache#751 and the visit-hook success-tail macro from apache#753. Checked Map-valued hook fields use the current Expected-returning type-error path. `StmtNode` and `TypeNode` declare `_type_s_eq_hash_subclass_kind_fixed = true`. The current hierarchy census finds 18 final registered children under Stmt and 14 under Type, with no FreeVar or DAG kind override in either hierarchy. This lets typed structural-map callbacks over Stmt and Type fold out the identity-remap branch. `Expr`, `PrimExpr`, and `BaseFunc` retain the default because their descendants include FreeVar or DAG kinds. Definition-region annotations are aligned with reflected fallback behavior: direct definitions use non-recursive regions, container-owned definitions are established at their Var sites, and dynamic Var type descent remains a use. The hooks preserve the established visitor/mutator field set and same-object short circuits on unchanged mutations. Call hooks keep ordinary descent for args and nonempty type arguments, while skipping primitive result types, interned Op operators, and empty type-argument containers. Dynamic result types, function-valued operators, and populated type arguments retain ordinary descent. `PostOrderVisit` and `Substitute` remain unchanged. Direct `StructuralMap` roots in this PR are confined to focused tests: the intentionally shared lvalue exercises copy-on-write, while owned roots move into the by-value API to preserve eligible in-place reuse. The six core hook source files each use one anonymous namespace and non-static file-local hook functions. Rebuilds box moved typed copies directly, unchanged paths box typed `self`, and constant leaves explicitly extract typed `self`. Every visit hook uses direct early-return macro calls per field and terminates with `TVM_FFI_S_VISIT_RETURN_NONE()`. `MaybeInplaceMutateSeqStmtRaw` separately verifies unique ownership of its `seq` field and borrows each element from raw array storage so the element remains eligible for in-place mutation. Nested SeqStmt results contribute zero, one, or many statements through one cursor: the existing array is reshaped when capacity permits, while overflow evaluates the remaining suffix exactly once and allocates one exactly sized array. Differential tests cover first and last positions, multiple splices, growth, size preservation, shrink, an empty result, array reuse, overflow replacement, and callback non-replay. Owner measurements in the apache#373 harness on a 256-element SeqStmt record the accepted timing guardrail. Two replacements change from 23,511 to 24,444 ns (+4.0%), 16 from 24,430 to 24,768 ns (+1.4%), 64 from 25,464 to 25,132 ns (-1.3%), and 128 from 27,367 to 25,245 ns (-7.8%). Rebuilds fall from 3-5 to zero at every density, with crossover near one quarter. This repair lands for correct element-level in-place semantics and reduced splice allocation, not as a uniform speedup. Earlier general performance measurements for this work in apache#364/apache#365 were collected at the previous `556514c76225cb35a55e0649098b07be3fba2991` engine pin and predate this bump; they should not be interpreted as measurements of the newer engine. The SeqStmt-specific figures above are the later apache#373 measurement of the accepted repair tradeoff.

Add compiled structural visit, mutate, and maybe-in-place-mutate hooks for every Expr and Stmt node family using the current tvm-ffi structural-hook ABI. The branch pins tvm-ffi at `9d784c4da74ff7360d76c79a89fe60a63516f880`, including the callback-owned `StructuralMutate` API from tvm-ffi apache#751 and the visit-hook success-tail macro from apache#753. Checked Map-valued hook fields use the current Expected-returning type-error path. `StmtNode` and `TypeNode` declare `_type_s_eq_hash_subclass_kind_fixed = true`. The current hierarchy census finds 18 final registered children under Stmt and 14 under Type, with no FreeVar or DAG kind override in either hierarchy. This lets typed structural-map callbacks over Stmt and Type fold out the identity-remap branch. `Expr`, `PrimExpr`, and `BaseFunc` retain the default because their descendants include FreeVar or DAG kinds. Definition-region annotations are aligned with reflected fallback behavior: direct definitions use non-recursive regions, container-owned definitions are established at their Var sites, and dynamic Var type descent remains a use. The hooks preserve the established visitor/mutator field set and same-object short circuits on unchanged mutations. Call hooks keep ordinary descent for args and nonempty type arguments, while skipping primitive result types, interned Op operators, and empty type-argument containers. Dynamic result types, function-valued operators, and populated type arguments retain ordinary descent. `PostOrderVisit` and `Substitute` remain unchanged. Direct `StructuralMap` roots in this PR are confined to focused tests: the intentionally shared lvalue exercises copy-on-write, while owned roots move into the by-value API to preserve eligible in-place reuse. All eight hook source files use one anonymous namespace and non-static file-local hook functions. Rebuilds box moved typed copies directly, unchanged paths box typed `self`, and constant leaves explicitly extract typed `self`. Every visit hook uses direct early-return macro calls per field and terminates with `TVM_FFI_S_VISIT_RETURN_NONE()`. `MaybeInplaceMutateSeqStmtRaw` separately verifies unique ownership of its `seq` field and borrows each element from raw array storage so the element remains eligible for in-place mutation. Mapped `Evaluate(0)` entries are omitted just as they are by `SeqStmt::Flatten`; zero or one remaining statement normalizes to `Evaluate(0)` or the sole statement without constructing a prohibited SeqStmt. This does not change the SeqStmt constructor or contract. Both SeqStmt mutation hooks scan their unchanged prefix with only one same-object test per element, then tail-return to changed-path helpers at the first divergence. Ordinary mutation allocates and initializes its unchanged prefix only there. In-place mutation begins its `total` cursor there: drops and splices compact into the existing array while writes remain behind unread input; a splice that would cross that boundary transfers once to an `InplaceSplice` output and continues the remaining callbacks exactly once. Differential tests cover first and last positions, multiple splices, growth, size preservation, `Evaluate(0)` shrink, zero/one-result normalization, array reuse, overflow replacement, and callback non-replay without constructing raw empty or size-one SeqStmt nodes. Owner measurements in the apache#373 harness on a 256-element SeqStmt record the accepted timing guardrail. Two replacements change from 23,511 to 24,444 ns (+4.0%), 16 from 24,430 to 24,768 ns (+1.4%), 64 from 25,464 to 25,132 ns (-1.3%), and 128 from 27,367 to 25,245 ns (-7.8%). Rebuilds fall from 3-5 to zero at every density, with crossover near one quarter. A follow-up L=8 allocation check finds that repaired shrink reuses the node array instead of allocating once, while growth remains one allocation with or without spare capacity. This repair lands for correct element-level in-place semantics and reduced splice allocation, not as a uniform speedup. Earlier general performance measurements for this work in apache#364/apache#365 were collected at the previous `556514c76225cb35a55e0649098b07be3fba2991` engine pin and predate this bump; they should not be interpreted as measurements of the newer engine. The SeqStmt-specific figures above are the later apache#373 measurement of the accepted repair tradeoff.

Huyuwei added 2 commits August 21, 2017 15:56

modify depthconv padding

6d223fe

fix shared memory overflow in depthconv schedule

71aa4fd

tqchen approved these changes Aug 21, 2017

View reviewed changes

tqchen merged commit 310f56f into apache:master Aug 21, 2017

tlopex pushed a commit to tlopex/tvm that referenced this pull request Mar 9, 2026

[Megakernel] Add sglang and flashinfer into MOE unittest (apache#365)

bddcd7d

tqchen mentioned this pull request Sep 6, 2026

[REFACTOR][IR] Add structural hooks for Expr and Stmt #20275

Merged

Navigation Menu

Uh oh!

[TOPI] update depthconv padding api; fix shared memory overflow - #365

[TOPI] update depthconv padding api; fix shared memory overflow#365
tqchen merged 2 commits into
apache:masterfrom
Huyuwei:topi

Huyuwei commented Aug 21, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

Huyuwei commented Aug 21, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants