Skip to content

[Fix issue #1023]: T.tile.atomic_add with slice syntax + vector lane variables #1034

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
fuhouyu-hw merged 3 commits into from
May 19, 2026
+140 −34
Merged

[Fix issue #1023]: T.tile.atomic_add with slice syntax + vector lane variables #1034

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9f950c6
[fix] Fix T.tile.atomic_add with slice syntax + vector lane variables
May 16, 2026
56a0589
[Fix] fix the issue #1023. Make sure linear attention bwd pass.
May 18, 2026
a21abf2
[chore] clang-format for src/op/ascend.cc
May 18, 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
10 changes: 6 additions & 4 deletions src/op/ascend.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,6 +330,9 @@ Stmt AscendCopy::Lower(const LowerArgs &T, arith::Analyzer *analyzer) const {
auto compute_valid_extent = [](PrimExpr min_val, PrimExpr extent,
PrimExpr shape) -> PrimExpr {
PrimExpr remaining = shape - min_val;
if (remaining.dtype().lanes() > 1) {
return extent;
}
return Select(remaining >= extent, extent,
Select(remaining > 0, remaining, 0));
};
Expand Down Expand Up @@ -570,6 +573,9 @@ Stmt AscendAtomicAdd::Lower(const LowerArgs &T,
auto compute_valid_extent = [](PrimExpr min_val, PrimExpr extent,
PrimExpr shape) -> PrimExpr {
PrimExpr remaining = shape - min_val;
if (remaining.dtype().lanes() > 1) {
return extent;
}
Comment on lines +576 to +578
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot May 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The logic for compute_valid_extent is duplicated here from AscendCopy::Lower (line 333). To improve maintainability and reduce code duplication, consider refactoring this into a shared helper function or a static method within the file.

return Select(remaining >= extent, extent,
Select(remaining > 0, remaining, 0));
};
Expand Down Expand Up @@ -612,10 +618,6 @@ Stmt AscendAtomicAdd::Lower(const LowerArgs &T,
ICHECK_EQ(dst_extents.size(), dst->shape.size())
<< "tl.ascend_atomic_add destination region rank must match "
"destination buffer rank";
ICHECK_EQ(src_extents.size(), dst_extents.size())
<< "tl.ascend_atomic_add requires src and dst regions to have the same "
"rank, got src "
<< src_extents.size() << " and dst " << dst_extents.size();

Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot May 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The rank equality check between src_extents and dst_extents has been removed to support cases where the source and destination regions have different ranks (e.g., 2D to 4D). While this is necessary for the described fix, it is important to ensure that the total number of elements in both regions still matches. Consider adding a check to verify that src_len == dst_len after they are computed (around line 664) to prevent potential out-of-bounds access or mismatched data transfers.

std::stringstream ss;
if (src.scope() == "shared") {
Expand Down
15 changes: 13 additions & 2 deletions src/target/codegen_ascend.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2252,8 +2252,19 @@ void CodeGenTileLangAscend::CopyCodegen(const CallNode *op) {
dst_var_id = dst_var->name_hint;
}

auto src_offset = PrintExpr(op->args[1].as<CallNode>()->args[2]);
auto dst_offset = PrintExpr(op->args[2].as<CallNode>()->args[2]);
auto src_offset_expr = op->args[1].as<CallNode>()->args[2];
auto dst_offset_expr = op->args[2].as<CallNode>()->args[2];
// AscendLowerParallelToVector may produce Ramp offsets (e.g. vid * stride).
// AscendC GlobalTensor/LocalTensor operator[] expects a scalar start offset;
// per-element strides within a tile are handled by DataCopyExtParams / DMA.
if (const auto *ramp = src_offset_expr.as<RampNode>()) {
src_offset_expr = ramp->base;
}
if (const auto *ramp = dst_offset_expr.as<RampNode>()) {
dst_offset_expr = ramp->base;
}
auto src_offset = PrintExpr(src_offset_expr);
auto dst_offset = PrintExpr(dst_offset_expr);

auto src_type = GetAccessPtrDtype(op->args[1].as<CallNode>());
auto dst_type = GetAccessPtrDtype(op->args[2].as<CallNode>());
Expand Down
10 changes: 7 additions & 3 deletions src/transform/ascend_storage_rewrite.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1460,9 +1460,13 @@ class VectorTypeAccessChecker : public StmtExprVisitor {
// vectorized pointer types (e.g. float16x4*). Once they do, this if
// statement should instead be replaced by the below ICHECK_EQ.
if (index_lanes * var_info.element_dtype.lanes() != value_dtype.lanes()) {
ICHECK_EQ(index_lanes, value_dtype.lanes());
lanes_used = 1;
var_info.element_dtype = var_info.element_dtype.with_lanes(1);
if (index_lanes > 1 && value_dtype.lanes() == 1) {
lanes_used = 1;
} else {
ICHECK_EQ(index_lanes, value_dtype.lanes());
lanes_used = 1;
var_info.element_dtype = var_info.element_dtype.with_lanes(1);
}
}

// TODO(Lunderberg): Uncomment this check once it can be applied.
Expand Down
Loading