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

Count the sharding as compatible if the factors that need replication are unsharded across all operands and results. #313

Merged
merged 1 commit into from
Jan 21, 2025
+72 −11
Merged

Count the sharding as compatible if the factors that need replication are unsharded across all operands and results. #313

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
36 changes: 25 additions & 11 deletions shardy/dialect/sdy/transforms/export/insert_explicit_reshards.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,11 +62,18 @@ bool hasOverflowAxes(const ShardingProjection& projection) {

// Checks if factor sharding is compatible, that is, it satisfies:
// 1. Factors are sharded the same way across operands and results.
// 2. Factors that need replication are unsharded.
//
// Assumes factor shardings do not have overflow axes.
// TODO(enver): Handle the case when some factor shardings have overflow axes.
bool hasCompatibleFactorShardings(const ShardingProjection& projection) {
bool hasCompatibleFactorShardings(const ShardingProjection& projection,
OpShardingRuleAttr shardingRule) {
FactorIndexToSharding factorIndexToCommonSharding;
// Factors that need replication should be unsharded across all operands and
// results in order for it to have a compatible sharding.
for (int64_t factorIndex : shardingRule.getNeedReplicationFactors()) {
factorIndexToCommonSharding[factorIndex] = FactorSharding{};
}
for (const TensorFactorShardings& tensorFactorSharding :
llvm::concat<const TensorFactorShardings>(projection.getOperands(),
projection.getResults())) {
Expand Down Expand Up @@ -535,15 +542,12 @@ struct InsertExplicitReshardsPass
// Return without inserting reshards for operations with special
// dimensions.
// TODO(enver): Insert explicit reshards if special dimensions are
// unsharded, or all speical dimensions need replication and annotated as
// such on the sharding rule.
if (isa<stablehlo::CholeskyOp, stablehlo::ReverseOp,
stablehlo::BitcastConvertOp, stablehlo::BroadcastInDimOp,
stablehlo::ConcatenateOp, stablehlo::DynamicSliceOp,
stablehlo::DynamicUpdateSliceOp, stablehlo::PadOp,
stablehlo::SliceOp, stablehlo::SortOp, stablehlo::TransposeOp,
stablehlo::TriangularSolveOp, stablehlo::FftOp,
stablehlo::ReduceWindowOp, stablehlo::ScatterOp,
// unsharded.
// TODO(enver): Add need replication factors to fft.
if (isa<stablehlo::ReverseOp, stablehlo::BroadcastInDimOp,
stablehlo::DynamicSliceOp, stablehlo::DynamicUpdateSliceOp,
stablehlo::PadOp, stablehlo::SliceOp, stablehlo::TransposeOp,
stablehlo::FftOp, stablehlo::ReduceWindowOp, stablehlo::ScatterOp,
stablehlo::SelectAndScatterOp, stablehlo::GatherOp,
stablehlo::ReshapeOp, stablehlo::ConvolutionOp,
stablehlo::CustomCallOp, stablehlo::ReduceOp,
Expand All @@ -554,7 +558,17 @@ struct InsertExplicitReshardsPass
}

// Checks if factors are sharded the same way across operands and results.
if (hasCompatibleFactorShardings(shardingProjection)) {
if (hasCompatibleFactorShardings(shardingProjection, shardingRule)) {
return;
}

// Return without inserting reshards for operations with factors that need
// replication.
// TODO(enver): Insert explicit reshards also for the case that the
// factors that need replication are sharded.
if (isa<stablehlo::CholeskyOp, stablehlo::BitcastConvertOp,
stablehlo::ConcatenateOp, stablehlo::SortOp,
stablehlo::TriangularSolveOp>(op)) {
return;
}

Expand Down
47 changes: 47 additions & 0 deletions shardy/dialect/sdy/transforms/export/test/insert_explicit_reshards.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -542,6 +542,53 @@ func.func @sort_all_other_dims_size_one(%arg0: tensor<1x4x1xi32> {sdy.sharding =
return %0 : tensor<1x4x1xi32>
}

// CHECK-LABEL: func @sort_single_input_output
func.func @sort_single_input_output(%arg0: tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh, [{"x"}, {"y"}, {}]>}) -> (tensor<4x32x8xi32>) {
// CHECK-NOT: sdy.reshard
%0 = "stablehlo.sort"(%arg0) ({
^bb0(%arg2: tensor<i32>, %arg3: tensor<i32>):
%1 = stablehlo.compare GT, %arg2, %arg3 : (tensor<i32>, tensor<i32>) -> tensor<i1>
stablehlo.return %1 : tensor<i1>
}) {dimension = 0 : i64, is_stable = true} : (tensor<4x32x8xi32>) -> (tensor<4x32x8xi32>)
return %0 : tensor<4x32x8xi32>
}

// CHECK-LABEL: func @sort_compatible
func.func @sort_compatible(%arg0: tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh, [{}, {"y"}, {}]>}) -> (tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh, [{}, {"y"}, {}]>}) {
// CHECK-NOT: sdy.reshard
%0 = "stablehlo.sort"(%arg0) <{dimension = 0 : i64, is_stable = true}> ({
^bb0(%arg2: tensor<i32>, %arg3: tensor<i32>):
%1 = stablehlo.compare GT, %arg2, %arg3 : (tensor<i32>, tensor<i32>) -> tensor<i1>
stablehlo.return %1 : tensor<i1>
}) {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{}, {"y"}, {}]>]>} : (tensor<4x32x8xi32>) -> (tensor<4x32x8xi32>)
return %0 : tensor<4x32x8xi32>
}


// CHECK-LABEL: func @sort_input_and_output_shardings_are_same_on_sorting_dimension
func.func @sort_input_and_output_shardings_are_same_on_sorting_dimension(%arg0: tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh, [{"x"}, {"y"}, {}]>}) -> (tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh, [{"x"}, {"y"}, {}]>}) {
// CHECK-NOT: sdy.reshard
// TODO(enver): Support cases that factors need replication and sharded in the same way, which still requires resharding since the sorting dimension is fully replicated.
%0 = "stablehlo.sort"(%arg0) <{dimension = 0 : i64, is_stable = true}> ({
^bb0(%arg2: tensor<i32>, %arg3: tensor<i32>):
%1 = stablehlo.compare GT, %arg2, %arg3 : (tensor<i32>, tensor<i32>) -> tensor<i1>
stablehlo.return %1 : tensor<i1>
}) {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"x"}, {"y"}, {}]>]>} : (tensor<4x32x8xi32>) -> (tensor<4x32x8xi32>)
return %0 : tensor<4x32x8xi32>
}


// CHECK-LABEL: func @sort_input_and_output_shardings_are_different_on_sorting_dimension
func.func @sort_input_and_output_shardings_are_different_on_sorting_dimension(%arg0: tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh_xyz, [{"x"}, {"z"}, {}]>}) -> (tensor<4x32x8xi32> {sdy.sharding = #sdy.sharding<@mesh_xyz, [{"y"}, {"z"}, {}]>}) {
// CHECK-NOT: sdy.reshard
%0 = "stablehlo.sort"(%arg0) <{dimension = 0 : i64, is_stable = true}> ({
^bb0(%arg2: tensor<i32>, %arg3: tensor<i32>):
%1 = stablehlo.compare GT, %arg2, %arg3 : (tensor<i32>, tensor<i32>) -> tensor<i1>
stablehlo.return %1 : tensor<i1>
}) {sdy.sharding = #sdy.sharding_per_value<[<@mesh_xyz, [{"y"}, {"z"}, {}]>]>} : (tensor<4x32x8xi32>) -> (tensor<4x32x8xi32>)
return %0 : tensor<4x32x8xi32>
}

// CHECK-LABEL: func @transpose
func.func @transpose(%arg0: tensor<256x32x64x100xf32> {sdy.sharding = #sdy.sharding<@mesh, [{}, {"x"}, {}, {}]>}) -> tensor<100x32x256x64xf32> {
// CHECK-NOT: sdy.reshard
Expand Down