Skip to content

[BugFix] Fix some bugs in lowering ParallelOp and VectorizeLoop #1607

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
LeiWang1999 merged 3 commits into from
Jan 6, 2026
Merged

[BugFix] Fix some bugs in lowering ParallelOp and VectorizeLoop #1607

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f2fe56c
fix
SiriusNEO Jan 4, 2026
5395503
[BugFix] Fix some bugs in lowering ParallelOp and VectorizeLoop
SiriusNEO Jan 5, 2026
11fe7a1
remove debug output
SiriusNEO Jan 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 7 additions & 2 deletions src/op/utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,8 +47,13 @@ inline bool IsGlobalBuffer(const Buffer &buffer) {
return buffer.defined() && buffer.scope() == "global";
}

inline bool IsLocalBuffer(const Buffer &buffer) {
return buffer.defined() && buffer.scope() == "local";
inline bool IsLocalBuffer(const Buffer &buffer, bool allow_var = false) {
if (allow_var) {
return buffer.defined() &&
(buffer.scope() == "local" || buffer.scope() == "local.var");
} else {
return buffer.defined() && buffer.scope() == "local";
}
}

inline bool IsLocalVarBuffer(const Buffer &buffer) {
Expand Down
17 changes: 13 additions & 4 deletions src/transform/loop_vectorize.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,15 +129,15 @@ class VectorizePlanner : public arith::IRMutatorWithAnalyzer {
return arith::IRMutatorWithAnalyzer::VisitExpr_(node);
}
}
UpdateVectorSize(node->indices, node->buffer);
UpdateVectorSize(node->indices, node->buffer, false);
return arith::IRMutatorWithAnalyzer::VisitExpr_(node);
}

Stmt VisitStmt_(const BufferStoreNode *node) final {
if (node->buffer.scope() == "shared" || node->buffer.scope() == "global" ||
node->buffer.scope() == "shared.dyn")
has_nonlocal_memory_access_ = true;
UpdateVectorSize(node->indices, node->buffer);
UpdateVectorSize(node->indices, node->buffer, true);
return arith::IRMutatorWithAnalyzer::VisitStmt_(node);
}

Expand Down Expand Up @@ -170,7 +170,8 @@ class VectorizePlanner : public arith::IRMutatorWithAnalyzer {
return arith::IRMutatorWithAnalyzer::VisitExpr_(node);
}

void UpdateVectorSize(const Array<PrimExpr> indices, const Buffer &buffer) {
void UpdateVectorSize(const Array<PrimExpr> indices, const Buffer &buffer,
bool is_store) {
if (!inner_for_)
return;
// 1. Compute raw element offset
Expand All @@ -187,8 +188,12 @@ class VectorizePlanner : public arith::IRMutatorWithAnalyzer {
for (int i = 0; i < indices.size(); ++i) {
elem_offset += indices[i] * strides[i];
}
// 2. If element offset is independent with loop_var, ignore it
// 2. If element offset is independent with loop_var, ignore it.
if (CanProveIndependent(elem_offset, inner_for_->loop_var, analyzer_)) {
// Specially, if it's a BufferStore, we should not vectorize it.
if (is_store) {
vector_size_ = 1;
}
return;
}
// 3. Check if current vector_size_ works with invariant boundary check
Expand All @@ -198,6 +203,10 @@ class VectorizePlanner : public arith::IRMutatorWithAnalyzer {
vector_size_ = arith::ZeroAwareGCD(
vector_size_, vector_load_bits_max_ /
(buffer->dtype.bits() * buffer->dtype.lanes()));
} else if (is_store) {
// If the indices is invariant for BufferStore, we should also not
// vectorize it.
vector_size_ = 1;
}
// 4. Try to vectorize buffer load
while (!IndiceCanVectorize(elem_offset, inner_for_->loop_var,
Expand Down
12 changes: 7 additions & 5 deletions src/transform/lower_tile_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -735,7 +735,7 @@ class LowerTileOpPass : arith::IRMutatorWithAnalyzer {
bool store_into_local = false;
PostOrderVisit(root, [&](const ObjectRef &obj) {
if (const auto *store = obj.as<BufferStoreNode>()) {
if (IsLocalBuffer(store->buffer)) {
if (IsLocalBuffer(store->buffer, true)) {
store_into_local = true;
}
}
Expand All @@ -748,11 +748,11 @@ class LowerTileOpPass : arith::IRMutatorWithAnalyzer {
bool local_register_only = true;
PostOrderVisit(root, [&](const ObjectRef &obj) {
if (const auto *store = obj.as<BufferStoreNode>()) {
if (!IsLocalBuffer(store->buffer)) {
if (!IsLocalBuffer(store->buffer, true)) {
local_register_only = false;
}
} else if (const auto *load = obj.as<BufferLoadNode>()) {
if (!IsLocalBuffer(load->buffer)) {
if (!IsLocalBuffer(load->buffer, true)) {
local_register_only = false;
}
}
Expand All @@ -766,11 +766,13 @@ class LowerTileOpPass : arith::IRMutatorWithAnalyzer {
bool has_non_local = false;
PostOrderVisit(for_node->body, [&](const ObjectRef &obj) {
if (const auto *load = obj.as<BufferLoadNode>()) {
if (!IsLocalBuffer(load->buffer) && !IsFragmentBuffer(load->buffer)) {
if (!IsLocalBuffer(load->buffer, true) &&
!IsFragmentBuffer(load->buffer)) {
has_non_local = true;
}
} else if (const auto *store = obj.as<BufferStoreNode>()) {
if (!IsLocalBuffer(store->buffer) && !IsFragmentBuffer(store->buffer)) {
if (!IsLocalBuffer(store->buffer, true) &&
!IsFragmentBuffer(store->buffer)) {
has_non_local = true;
}
}
Expand Down
34 changes: 34 additions & 0 deletions testing/python/issue/test_tilelang_issue_1549.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,34 @@
import tilelang as tl
import tilelang.testing
import tilelang.language as T
import torch


def test_issue_1549_strange_var_vectorization():
@tl.jit
def get_wrong_kernel(M: int = 4096):
dtype = "int32"
num_threads = 64

@T.prim_func
def main(
Data: T.Tensor((M,), dtype),
):
with T.Kernel(1, threads=num_threads) as _:
# Pre-allocated scalar variables (causes issue in 0.1.7.post1)
idx = T.alloc_var(T.int32)
for i in T.Parallel(M):
idx = i
Data[i] = idx

return main

kernel = get_wrong_kernel()
M = 2048
kernel = get_wrong_kernel(M)
data = torch.randint(0, 100, (M,), dtype=torch.int32, device="cuda")
kernel(data)


if __name__ == "__main__":
tilelang.testing.main()
Loading