Skip to content

Set default other = 0 for masked load #356

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Set default other = 0 for masked load #356

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8f27555
Set default other = 0 for masked load
ziliangzl Sep 18, 2025
9c6a575
update test
ziliangzl Sep 29, 2025
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: 6 additions & 3 deletions lib/Conversion/StructuredToMemref/StructuredToMemref.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -807,10 +807,13 @@ struct LoadConverter : public OpConversionPattern<tts::LoadOp> {
SmallVector<OpFoldResult> mixedDims = op.getMixedMaskDims();

// Fill load destination with other value
if (Value other = op.getOther()) {
fillWithValue(loc, alloc, other, tensorType.getShape(),
op.getMixedMaskDims(), op.getStaticMaskDims(), rewriter);
Value other = op.getOther();
if (!other) {
other = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(elemType));
}
fillWithValue(loc, alloc, other, tensorType.getShape(),
op.getMixedMaskDims(), op.getStaticMaskDims(), rewriter);

auto ptrDefiningOp = ptr.getDefiningOp();
if (ptrDefiningOp->hasAttr(WRAP_SIDE_BY_SIDE) ||
Expand Down
2 changes: 0 additions & 2 deletions python/examples/test_mm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,12 +62,10 @@ def mm_kernel(
a = tl.load(
A + (ram[:, None] * stride_am + rk[None, :] * stride_ak),
mask=mask_k[None, :],
other=0.0
)
b = tl.load(
B + (rk[:, None] * stride_bk + rbn[None, :] * stride_bn),
mask=mask_k[:, None],
other=0.0
)
if a.dtype != b.dtype:
a = a.to(C.dtype.element_ty)
Expand Down
10 changes: 8 additions & 2 deletions test/Conversion/StructuredToMemref/kernel-05-layer-norm-fwd.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,13 +225,20 @@ module {
// CHECK: [[VAR_24_2_:%.+]] = arith.maxsi [[VAR_23_2_]], [[VAR_20_5_]] : index
// CHECK-DAG: [[VAR_25_2_:%.+]] = arith.subi [[VAR_24_2_]], [[VAR_20_5_]] : index
// CHECK-DAG: [[RES_2_:%.+]] = memref.alloc() : memref<256xf32>
// CHECK: [[CMPI_:%.+]] = arith.cmpi slt, [[VAR_25_2_]], [[CST_256_1_]] : index
// CHECK: scf.if [[CMPI_]] {
// CHECK: linalg.fill ins([[CST_0_dot_000000_]] : f32) outs([[RES_2_]] : memref<256xf32>)
// CHECK: }
// CHECK-NOT: separator of consecutive DAGs
// CHECK-DAG: [[VAR_subview_2_:%.+]] = memref.subview [[VAR_reinterpret_cast_5_2_]][0] {{.}}[[VAR_25_2_]]{{.}} [1] : memref<256xf32, strided<[1], offset: ?>> to memref<?xf32, strided<[1], offset: ?>>
// CHECK-DAG: [[VAR_subview_6_2_:%.+]] = memref.subview [[RES_2_]][0] {{.}}[[VAR_25_2_]]{{.}} [1] : memref<256xf32> to memref<?xf32, strided<[1]>>
// CHECK: memref.copy [[VAR_subview_2_]], [[VAR_subview_6_2_]] : memref<?xf32, strided<[1], offset: ?>> to memref<?xf32, strided<[1]>>
// CHECK-DAG: [[VAR_26_2_:%.+]] = bufferization.to_tensor [[RES_2_]] restrict writable : memref<256xf32>
// CHECK-DAG: [[VAR_reinterpret_cast_7_:%.+]] = memref.reinterpret_cast [[PARAM_3_]] to offset: {{.}}[[VAR_20_5_]]{{.}}, sizes: [256], strides: [1] : memref<*xf32> to memref<256xf32, strided<[1], offset: ?>>
// CHECK-DAG: [[RES_3_:%.+]] = memref.alloc() : memref<256xf32>
// CHECK: scf.if [[CMPI_]] {
// CHECK: linalg.fill ins([[CST_0_dot_000000_]] : f32) outs([[RES_3_]] : memref<256xf32>)
// CHECK: }
// CHECK-NOT: separator of consecutive DAGs
// CHECK-DAG: [[VAR_subview_9_:%.+]] = memref.subview [[VAR_reinterpret_cast_7_]][0] {{.}}[[VAR_25_2_]]{{.}} [1] : memref<256xf32, strided<[1], offset: ?>> to memref<?xf32, strided<[1], offset: ?>>
// CHECK-DAG: [[VAR_subview_10_:%.+]] = memref.subview [[RES_3_]][0] {{.}}[[VAR_25_2_]]{{.}} [1] : memref<256xf32> to memref<?xf32, strided<[1]>>
Expand All @@ -241,8 +248,7 @@ module {
// CHECK-NOT: separator of consecutive DAGs
// CHECK-DAG: [[VAR_reinterpret_cast_11_:%.+]] = memref.reinterpret_cast [[PARAM_0_]] to offset: {{.}}[[VAR_28_2_]]{{.}}, sizes: [256], strides: [1] : memref<*xf32> to memref<256xf32, strided<[1], offset: ?>>
// CHECK-DAG: [[RES_4_:%.+]] = memref.alloc() : memref<256xf32>
// CHECK-DAG: [[VAR_29_2_:%.+]] = arith.cmpi slt, [[VAR_25_2_]], [[CST_256_1_]] : index
// CHECK: scf.if [[VAR_29_2_]] {
// CHECK: scf.if [[CMPI_]] {
// CHECK: linalg.fill ins([[CST_0_dot_000000_]] : f32) outs([[RES_4_]] : memref<256xf32>)
// CHECK: }
// CHECK-DAG: [[VAR_subview_13_:%.+]] = memref.subview [[VAR_reinterpret_cast_11_]][0] {{.}}[[VAR_25_2_]]{{.}} [1] : memref<256xf32, strided<[1], offset: ?>> to memref<?xf32, strided<[1], offset: ?>>
Expand Down