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

#sdy. Fix sharding rule for SliceOp. #267

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+37 −24
Open

#sdy. Fix sharding rule for SliceOp. #267

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
47 changes: 30 additions & 17 deletions shardy/dialect/sdy/transforms/propagation/op_sharding_rule_registry.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -863,23 +863,36 @@ OpShardingRuleAttr createOpShardingRule(Operation* op,
.addPointwise(getTensorShape(select.getResult()))
.build();
})
.Case<stablehlo::SliceOp>(
[conservativePropagation](stablehlo::SliceOp slice) {
// If `conservativePropagation` is false, we propagate through
// sliced dimensions, even though that would require communication.
//
// This is different from `DynamicSliceOp`, where we don't
// propagate through sliced dimensions regardless of
// `conservativePropagation`, and the reason is that for `SliceOp`
// the start indices are static, so we know how to shift the data
// to keep the sliced dimension sharded.
return OpShardingRuleBuilder(slice)
.addPointwiseIfDimSizesMatch(
getTensorShape(slice.getOperand()),
getTensorShape(slice.getResult()),
/*alwaysAddFactor=*/!conservativePropagation)
.build();
})
.Case<stablehlo::SliceOp>([conservativePropagation](
stablehlo::SliceOp slice) {
// If `conservativePropagation` is false, we propagate through
// sliced dimensions, even though that would require communication.
//
// There is an exception. If the input dimension size is larger than 1
// and the output dimension size is 1, we do not propagate through this
// sliced dimension.
//
// This is different from `DynamicSliceOp`, where we don't
// propagate through sliced dimensions regardless of
// `conservativePropagation`, and the reason is that for `SliceOp`
// the start indices are static, so we know how to shift the data
// to keep the sliced dimension sharded.
ArrayRef<int64_t> inShape = getTensorShape(slice.getOperand());
ArrayRef<int64_t> outShape = getTensorShape(slice.getResult());
auto onMismatchFn = [&](int64_t dim, OpShardingRuleBuilder& builder) {
if (conservativePropagation) {
return;
}
if (inShape[dim] != 1 && outShape[dim] == 1) {
return;
}
builder.addFactor(dim, inShape[dim]);
};
return OpShardingRuleBuilder(slice)
.addPointwiseIfDimSizesMatch(
inShape, outShape, /*alwaysAddFactor=*/false, onMismatchFn)
.build();
})
.Case<stablehlo::SortOp>([](stablehlo::SortOp sort) {
// If the input is sharded along the sort dimension, and any of the
// non-sort dimensions has size >1, the partitioner will add an
Expand Down
6 changes: 3 additions & 3 deletions shardy/dialect/sdy/transforms/propagation/sharding_projection_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -360,9 +360,9 @@ TEST_F(ShardingProjectionBuildTest, FactorWithSmallerSizeThanDimOverflows) {
sdy.mesh @mesh = <["a"=2, "b"=4, "c"=2, "d"=4, "e"=2]>

func.func @main(%arg0: tensor<32x4x16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"a", ?}, {"c", ?}, {"b", "d":(2)2, "e"}]>})
-> tensor<32x1x2xf32> {
%0 = stablehlo.slice %arg0 [0:32, 1:2, 4:6] : (tensor<32x4x16xf32>) -> tensor<32x1x2xf32>
return %0 : tensor<32x1x2xf32>
-> tensor<32x2x2xf32> {
%0 = stablehlo.slice %arg0 [0:32, 0:2, 4:6] : (tensor<32x4x16xf32>) -> tensor<32x2x2xf32>
return %0 : tensor<32x2x2xf32>
})mlir";

OwningOpRef<ModuleOp> module =
Expand Down
8 changes: 4 additions & 4 deletions shardy/dialect/sdy/transforms/propagation/test/op_sharding_rule_registry.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -452,10 +452,10 @@ func.func @pad(%arg0: tensor<28x28x16xf32>, %arg1: tensor<f32>) -> tensor<30x26x
}

// CHECK-LABEL: func @slice
func.func @slice(%arg0: tensor<32x4x8xf32>) -> tensor<32x1x2xf32> {
// CHECK: sdy.sharding_rule = #sdy.op_sharding_rule<([i, j, k])->([i, j, k]) {i=32, j=4, k=8}>
%0 = stablehlo.slice %arg0 [0:32, 1:2, 4:8:2] : (tensor<32x4x8xf32>) -> tensor<32x1x2xf32>
return %0 : tensor<32x1x2xf32>
func.func @slice(%arg0: tensor<32x4x8x1xf32>) -> tensor<32x1x2x1xf32> {
// CHECK: sdy.sharding_rule = #sdy.op_sharding_rule<([i, l, j, k])->([i, m, j, k]) {i=32, j=8, k=1, l=1, m=1}>
%0 = stablehlo.slice %arg0 [0:32, 1:2, 4:8:2, 0:1] : (tensor<32x4x8x1xf32>) -> tensor<32x1x2x1xf32>
return %0 : tensor<32x1x2x1xf32>
}

// Sort is currently treated as a pointwise op, and we add a factor for the sort
Expand Down