Skip to content

[bugfix] fix smem alloc for single warp reduce #1643

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 2 commits into from
Jan 9, 2026
+2 −2
Merged

[bugfix] fix smem alloc for single warp reduce #1643

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
581d51f
[bugfix] fix smem alloc for single warp reduce
botbw Jan 8, 2026
b299cfc
[bugfix] fix doc string
botbw Jan 8, 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
4 changes: 2 additions & 2 deletions src/op/reduce.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,7 +172,7 @@ std::string ReduceOpNode::MakeCodegenReducer() const {
* - Detects parallel thread splitting from the normalized iterator sum and
* emits a call to a templated `tl::AllReduce<...>::run` (or `run_hopper`)
* via `builtin::call_extern`. For sufficiently large reducing thread counts
* (>= 32) a workspace is allocated via T.AddWorkspace and passed to the
* (> 32) a workspace is allocated via T.AddWorkspace and passed to the
* AllReduce call.
* - The final body is wrapped in parallel loops over the destination spatial
* dimensions and partitioned by the lowering thread variable. If a temporary
Expand Down Expand Up @@ -322,7 +322,7 @@ Stmt ReduceOpNode::Lower(const LowerArgs &T, arith::Analyzer *analyzer) const {
}
Array<PrimExpr> thread_reduce_args = {
StringImm(ss.str()), BufferLoad(clear_buffer, dst_indices)};
if (reducing_threads >= 32) {
if (reducing_threads > 32) {
PrimExpr workspace = T.AddWorkspace(
*as_const_int(T.thread_bounds->extent), clear_buffer->dtype);
thread_reduce_args.push_back(workspace);
Expand Down
Loading