Skip to content
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

Jump to bottom

[Mosaic] Fix mask creation for packed sublanes #24852

Merged
merged 1 commit into from
Nov 12, 2024
+18 −7
Merged

[Mosaic] Fix mask creation for packed sublanes #24852

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
25 changes: 18 additions & 7 deletions jaxlib/mosaic/dialect/tpu/transforms/apply_vector_layout.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2664,7 +2664,6 @@ LogicalResult tpu_concatenate_rule(RewriteContext &ctx, Operation &op,

const auto bitwidth = res_ty.getElementTypeBitWidth();
const int packing = res_layout->packing();

SmallVector<int64_t> out_idx;
vreg.Each([&](absl::Span<const int64_t> idx, Value *v) {
out_idx.assign(idx.begin(), idx.end());
Expand All @@ -2674,17 +2673,29 @@ LogicalResult tpu_concatenate_rule(RewriteContext &ctx, Operation &op,
const VectorType vmask_ty = getNativeVregOrVmaskType(
builder.getI1Type(), bitwidth, ctx.target_shape);
if (tiling_dim.value() == 0) { // sublane
mask = builder.create<tpu::CreateMaskOp>(
op.getLoc(), vmask_ty,
ArrayRef<Value>{boundIdxConst(0), boundIdxConst(0)},
ArrayRef<Value>{boundIdxConst(operand_offset * packing),
boundIdxConst(layout->tiling()[1])});
if (operand_offset % packing != 0) {
// Packed case, degenerate where we have a half or quarter
// sublane.
// TODO(mvoz): We can probably always use the
// CreateSubelementMaskOp if (1) optimize it on TPUv4 and (2) Add
// support for unpacked types in some of the invariants in
// lower_to_llo.
mask = builder.create<tpu::CreateSubelementMaskOp>(
op.getLoc(), vmask_ty, 0, operand_offset, packing);
} else {
auto sublane_offset = operand_offset / packing;
mask = builder.create<tpu::CreateMaskOp>(
op.getLoc(), vmask_ty,
ArrayRef<Value>{boundIdxConst(0), boundIdxConst(0)},
ArrayRef<Value>{boundIdxConst(sublane_offset),
boundIdxConst(layout->tiling()[1])});
}
} else { // lane
mask = builder.create<tpu::CreateMaskOp>(
op.getLoc(), vmask_ty,
ArrayRef<Value>{boundIdxConst(0), boundIdxConst(0)},
ArrayRef<Value>{boundIdxConst(layout->tiling()[0]),
boundIdxConst(operand_offset * packing)});
boundIdxConst(operand_offset)});
}
// Blend the current value with the existing value in the output.
*v = builder.create<arith::SelectOp>(op.getLoc(), mask,
Expand Down
Loading