Skip to content

[FLYDSL]: Bug fixes for algebra not being the simplest #170

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
coderfeli merged 1 commit into from
Mar 3, 2026
+145 −11
Merged

[FLYDSL]: Bug fixes for algebra not being the simplest #170

Show file tree
Hide file tree
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
5 changes: 5 additions & 0 deletions include/flydsl/Dialect/Fly/Utils/IntTupleUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -845,6 +845,11 @@ std::pair<IntTuple, IntTuple> intTupleZip2ByImpl(const IntTupleBuilder<IntTuple>
assert(t.rank() == 2 && "intTupleZip2By expects rank-2 tuple at terminal");
return {builder.at(t, 0), builder.at(t, 1)};
}
// Canonicalize singleton guide wrappers so 1D profiles behave as leaf guides.
// This keeps zip2By robust after singleton unwrapping in product/divide type canonicalization.
if (guide.rank() == 1) {
return intTupleZip2ByImpl(builder, t, guide.at(0));
}
Collector firsts;
Collector seconds;

Expand Down
17 changes: 14 additions & 3 deletions include/flydsl/Dialect/Fly/Utils/LayoutUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,11 @@ std::pair<IntTuple, IntTuple> canonicalizeStridePair(const IntTupleBuilder<IntTu
}
return {shape, stride};
}
// Canonicalize singleton tuple wrappers so rank-1 trees print as leaf modes.
// Example: ((4), 4):((1), 4) -> (4, 4):(1, 4).
if (shape.rank() == 1) {
return canonicalizeStridePair(builder, builder.at(shape, 0), builder.at(stride, 0));
}
typename IntTupleBuilder<IntTuple>::ElemCollector shapeElems;
typename IntTupleBuilder<IntTuple>::ElemCollector strideElems;
for (int i = 0; i < shape.rank(); ++i) {
Expand Down Expand Up @@ -769,7 +774,9 @@ Layout layoutComposition(LayoutBuilder<Layout> &builder, Layout outerLayout, Lay
auto [retShape, retStride] =
detail::compositionImpl(builder, coalShape, coalStride, builder.getShape(innerLayout),
builder.getStride(innerLayout));
return builder.makeLayout(retShape, retStride);
auto [canonicalShape, canonicalStride] =
detail::canonicalizeStridePair(builder, retShape, retStride);
return builder.makeLayout(canonicalShape, canonicalStride);
}
template <class Layout>
Layout layoutComposition(LayoutBuilder<Layout> &builder, Layout outerLayout,
Expand Down Expand Up @@ -815,7 +822,9 @@ Layout layoutComposition(LayoutBuilder<Layout> &builder, Layout outerLayout,
retStride.push_back(builder.at(lhsStride, i));
}
}
return builder.makeLayout(builder.makeTuple(retShape), builder.makeTuple(retStride));
auto [canonicalShape, canonicalStride] =
detail::canonicalizeStridePair(builder, builder.makeTuple(retShape), builder.makeTuple(retStride));
return builder.makeLayout(canonicalShape, canonicalStride);
}

template <class Layout>
Expand Down Expand Up @@ -1214,7 +1223,9 @@ Layout layoutLogicalProduct(LayoutBuilder<Layout> &builder, Layout blockLayout,
retStrideElems.push_back(builder.getStride(blockLayout));
retStrideElems.push_back(builder.getStride(composed));

return builder.makeLayout(builder.makeTuple(retShapeElems), builder.makeTuple(retStrideElems));
auto [canonicalShape, canonicalStride] = detail::canonicalizeStridePair(
builder, builder.makeTuple(retShapeElems), builder.makeTuple(retStrideElems));
return builder.makeLayout(canonicalShape, canonicalStride);
}

template <class Layout>
Expand Down
55 changes: 51 additions & 4 deletions test/LayoutAlgebra/divide.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,14 +50,61 @@ func.func @test_flat_divide(%layout: !fly.layout<(4, 8) : (1, 4)>,
}

// CHECK-LABEL: @test_logical_divide_1d
func.func @test_logical_divide_1d() -> !fly.layout<((4), 4) : ((1), 4)> {
// Divide a 1D contiguous layout: (16):(1) / (4):(1) -> ((4),4):((1),4)
func.func @test_logical_divide_1d() -> !fly.layout<(4, 4) : (1, 4)> {
// Divide a 1D contiguous layout: (16):(1) / (4):(1) -> (4,4):(1,4)
%s = fly.static {elems = [16 : i32]} : () -> !fly.int_tuple<(16)>
%d = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%layout = fly.make_layout(%s, %d) : (!fly.int_tuple<(16)>, !fly.int_tuple<(1)>) -> !fly.layout<(16) : (1)>
%ds = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%dd = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%divisor = fly.make_layout(%ds, %dd) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.logical_divide(%layout, %divisor) : (!fly.layout<(16) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<((4), 4) : ((1), 4)>
return %result : !fly.layout<((4), 4) : ((1), 4)>
%result = fly.logical_divide(%layout, %divisor) : (!fly.layout<(16) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(4, 4) : (1, 4)>
return %result : !fly.layout<(4, 4) : (1, 4)>
}

// CHECK-LABEL: @test_zipped_divide_1d
func.func @test_zipped_divide_1d() -> !fly.layout<(4, 4) : (1, 4)> {
%s = fly.static {elems = [16 : i32]} : () -> !fly.int_tuple<(16)>
%d = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%layout = fly.make_layout(%s, %d) : (!fly.int_tuple<(16)>, !fly.int_tuple<(1)>) -> !fly.layout<(16) : (1)>
%ds = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%dd = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%divisor = fly.make_layout(%ds, %dd) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.zipped_divide(%layout, %divisor) : (!fly.layout<(16) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(4, 4) : (1, 4)>
return %result : !fly.layout<(4, 4) : (1, 4)>
}

// CHECK-LABEL: @test_tiled_divide_1d
func.func @test_tiled_divide_1d() -> !fly.layout<(4, 4) : (1, 4)> {
%s = fly.static {elems = [16 : i32]} : () -> !fly.int_tuple<(16)>
%d = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%layout = fly.make_layout(%s, %d) : (!fly.int_tuple<(16)>, !fly.int_tuple<(1)>) -> !fly.layout<(16) : (1)>
%ds = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%dd = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%divisor = fly.make_layout(%ds, %dd) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.tiled_divide(%layout, %divisor) : (!fly.layout<(16) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(4, 4) : (1, 4)>
return %result : !fly.layout<(4, 4) : (1, 4)>
}

// CHECK-LABEL: @test_flat_divide_1d
func.func @test_flat_divide_1d() -> !fly.layout<(4, 4) : (1, 4)> {
%s = fly.static {elems = [16 : i32]} : () -> !fly.int_tuple<(16)>
%d = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%layout = fly.make_layout(%s, %d) : (!fly.int_tuple<(16)>, !fly.int_tuple<(1)>) -> !fly.layout<(16) : (1)>
%ds = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%dd = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%divisor = fly.make_layout(%ds, %dd) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.flat_divide(%layout, %divisor) : (!fly.layout<(16) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(4, 4) : (1, 4)>
return %result : !fly.layout<(4, 4) : (1, 4)>
}

// CHECK-LABEL: @test_logical_divide_wrapped_tuple_1d
func.func @test_logical_divide_wrapped_tuple_1d(
%layout: !fly.layout<((16, 1)) : ((1, 16))>,
%divisor: !fly.layout<((4, 1)) : ((1, 4))>) -> !fly.layout<((4, 1), 4) : ((1, 0), 4)> {
// Outer singleton wrappers are accepted and handled in inference.
%result = fly.logical_divide(%layout, %divisor)
: (!fly.layout<((16, 1)) : ((1, 16))>, !fly.layout<((4, 1)) : ((1, 4))>)
-> !fly.layout<((4, 1), 4) : ((1, 0), 4)>
return %result : !fly.layout<((4, 1), 4) : ((1, 0), 4)>
}
79 changes: 75 additions & 4 deletions test/LayoutAlgebra/product.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,14 +68,85 @@ func.func @test_raked_product(%base: !fly.layout<(4, 8) : (1, 4)>,
}

// CHECK-LABEL: @test_logical_product_1d
func.func @test_logical_product_1d() -> !fly.layout<((8), (4)) : ((1), (8))> {
// 1D base with 1D tile preserves nesting structure
func.func @test_logical_product_1d() -> !fly.layout<(8, 4) : (1, 8)> {
// 1D base with 1D tile canonicalizes singleton tuple wrappers
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.logical_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<((8), (4)) : ((1), (8))>
return %result : !fly.layout<((8), (4)) : ((1), (8))>
%result = fly.logical_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(8, 4) : (1, 8)>
return %result : !fly.layout<(8, 4) : (1, 8)>
}

// CHECK-LABEL: @test_zipped_product_1d
func.func @test_zipped_product_1d() -> !fly.layout<(8, 4) : (1, 8)> {
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.zipped_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(8, 4) : (1, 8)>
return %result : !fly.layout<(8, 4) : (1, 8)>
}

// CHECK-LABEL: @test_tiled_product_1d
func.func @test_tiled_product_1d() -> !fly.layout<(8, 4) : (1, 8)> {
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.tiled_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(8, 4) : (1, 8)>
return %result : !fly.layout<(8, 4) : (1, 8)>
}

// CHECK-LABEL: @test_flat_product_1d
func.func @test_flat_product_1d() -> !fly.layout<(8, 4) : (1, 8)> {
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.flat_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(8, 4) : (1, 8)>
return %result : !fly.layout<(8, 4) : (1, 8)>
}

// CHECK-LABEL: @test_blocked_product_1d
func.func @test_blocked_product_1d() -> !fly.layout<(8, 4) : (1, 8)> {
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.blocked_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(8, 4) : (1, 8)>
return %result : !fly.layout<(8, 4) : (1, 8)>
}

// CHECK-LABEL: @test_raked_product_1d
func.func @test_raked_product_1d() -> !fly.layout<(4, 8) : (8, 1)> {
%s1 = fly.static {elems = [8 : i32]} : () -> !fly.int_tuple<(8)>
%d1 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%base = fly.make_layout(%s1, %d1) : (!fly.int_tuple<(8)>, !fly.int_tuple<(1)>) -> !fly.layout<(8) : (1)>
%s2 = fly.static {elems = [4 : i32]} : () -> !fly.int_tuple<(4)>
%d2 = fly.static {elems = [1 : i32]} : () -> !fly.int_tuple<(1)>
%tile = fly.make_layout(%s2, %d2) : (!fly.int_tuple<(4)>, !fly.int_tuple<(1)>) -> !fly.layout<(4) : (1)>
%result = fly.raked_product(%base, %tile) : (!fly.layout<(8) : (1)>, !fly.layout<(4) : (1)>) -> !fly.layout<(4, 8) : (8, 1)>
return %result : !fly.layout<(4, 8) : (8, 1)>
}

// CHECK-LABEL: @test_logical_product_wrapped_tuple_1d
func.func @test_logical_product_wrapped_tuple_1d(
%base: !fly.layout<((8, 1)) : ((1, 8))>,
%tile: !fly.layout<((4, 1)) : ((1, 4))>) -> !fly.layout<((8, 1), (4, 1)) : ((1, 0), (8, 0))> {
// Outer singleton wrappers are accepted and handled in inference.
%result = fly.logical_product(%base, %tile)
: (!fly.layout<((8, 1)) : ((1, 8))>, !fly.layout<((4, 1)) : ((1, 4))>)
-> !fly.layout<((8, 1), (4, 1)) : ((1, 0), (8, 0))>
return %result : !fly.layout<((8, 1), (4, 1)) : ((1, 0), (8, 0))>
}