Skip to content

Fix nested stableHLO composite regions #9385

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.

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 5 commits into
base: master
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
16 changes: 8 additions & 8 deletions test/stablehlo/test_composite.py
Original file line number Diff line number Diff line change
Expand Up @@ -147,10 +147,10 @@ def forward(self, x, y):
stablehlo = self.run_func_get_stablehlo(M(), input_args)
self.assertEqual(stablehlo.count("stablehlo.composite \"test.sdpa\""), 2)
self.assertTrue(
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl_0}'
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl}'
in stablehlo)
self.assertTrue(
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl}'
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl_0}'
in stablehlo)

def test_composite_builder_sdpa_pattern(self):
Expand All @@ -175,10 +175,10 @@ def forward(self, x, y):
stablehlo = self.run_func_get_stablehlo(M(), input_args)
self.assertEqual(stablehlo.count("stablehlo.composite \"test.sdpa\""), 2)
self.assertTrue(
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl_0}'
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl}'
in stablehlo)
self.assertTrue(
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl}'
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl_0}'
in stablehlo)

def test_composite_builder_export_sdpa_pattern(self):
Expand Down Expand Up @@ -208,10 +208,10 @@ def forward(self, x, y):
stablehlo = stablehlo_gm.get_stablehlo_text()
self.assertEqual(stablehlo.count("stablehlo.composite \"test.sdpa\""), 2)
self.assertTrue(
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl_0}'
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl}'
in stablehlo)
self.assertTrue(
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl}'
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl_0}'
in stablehlo)
if has_tf_package():
self.assertTrue(os.path.exists(os.path.join(tmp_path, 'saved_model.pb')))
Expand Down Expand Up @@ -240,10 +240,10 @@ def forward(self, x, y):
stablehlo = stablehlo_gm.get_stablehlo_text()
self.assertEqual(stablehlo.count("stablehlo.composite \"test.sdpa\""), 2)
self.assertTrue(
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl_0}'
'{composite_attributes = {scale = 2.500000e-01 : f32}, decomposition = @test.sdpa.impl}'
in stablehlo)
self.assertTrue(
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl}'
'{composite_attributes = {scale = 2 : i64}, decomposition = @test.sdpa.impl_0}'
in stablehlo)
if has_tf_package():
self.assertTrue(os.path.exists(os.path.join(tmp_path, 'saved_model.pb')))
Expand Down
42 changes: 40 additions & 2 deletions torch_xla/csrc/runtime/stablehlo_composite_helper.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -120,8 +120,30 @@ class BuildStableHLOCompositePass : public mlir::OperationPass<mlir::ModuleOp> {
std::unordered_map<std::string, llvm::SmallVector<mlir::Operation*>>
boundary_output_ops_map = BuildBoundaryOutputOpsMap(func_op);

for (const auto& [unused, ops] : boundary_output_ops_map) {
if (mlir::failed(BuildStableHLOComposite(ops, op_order_map))) {
struct BoundaryGroup {
std::string key;
llvm::SmallVector<mlir::Operation*> ops;
size_t last_order;
};

llvm::SmallVector<BoundaryGroup> groups;
groups.reserve(boundary_output_ops_map.size());

for (auto& kv : boundary_output_ops_map) {
size_t last_ord = 0;
for (mlir::Operation* op : kv.second) {
if (op != nullptr) last_ord = std::max(last_ord, op_order_map.at(op));
}
groups.push_back({kv.first, kv.second, last_ord});
}

llvm::sort(groups, [](const BoundaryGroup& a, const BoundaryGroup& b) {
return a.last_order < b.last_order;
});

for (auto& grp : groups) {
op_order_map = BuildOpOrderMap(func_op);
if (mlir::failed(BuildStableHLOComposite(grp.ops, op_order_map))) {
func_op.emitError() << "failed to build composite.";
return signalPassFailure();
}
Expand Down Expand Up @@ -321,6 +343,22 @@ class BuildStableHLOCompositePass : public mlir::OperationPass<mlir::ModuleOp> {
}
}

llvm::DenseSet<mlir::Operation*> wrapper_set(output_ops.begin(),
output_ops.end());

for (mlir::Operation* mark : output_ops)
if (mark->use_empty()) mark->erase();

for (mlir::Operation* op : llvm::reverse(impl_ops)) {
if (wrapper_set.contains(op) || !op->use_empty()) continue;

bool pure_or_composite = mlir::wouldOpBeTriviallyDead(op) ||
llvm::isa<mlir::stablehlo::CompositeOp>(op) ||
llvm::isa<mlir::stablehlo::CustomCallOp>(op);

if (pure_or_composite) op->erase();
}

if (!mlir::sortTopologically(composite_op->getBlock())) {
composite_op->emitError()
<< "The graph is not acyclic after BuildStableHLOCompositePass pass.";
Expand Down