[XLA] Propagate the layout of layout constrained custom calls with hi…

…gher priority because they have no ability to accept another layout. PiperOrigin-RevId: 674755625
Intel-tensorflow · Sep 15, 2024 · 14fbade · 14fbade
1 parent 84a4740
commit 14fbade
Show file tree

Hide file tree

Showing 2 changed files with 63 additions and 21 deletions.
diff --git a/third_party/xla/xla/service/layout_assignment.cc b/third_party/xla/xla/service/layout_assignment.cc
@@ -741,27 +741,6 @@ absl::Status LayoutAssignment::AddMandatoryConstraints(
           }
         }
       }
-    } else if (IsLayoutConstrainedCustomCall(instruction)) {
-      const HloCustomCallInstruction* custom_call =
-          DynCast<HloCustomCallInstruction>(instruction);
-
-      TF_RETURN_IF_ERROR(SetInstructionLayout(custom_call->shape(), custom_call,
-                                              /*mandatory=*/true, /*dfs=*/true,
-                                              /*allow_alias=*/true));
-      if (custom_call->IsCustomCall("LayoutConstraint")) {
-        TF_RETURN_IF_ERROR(
-            SetOperandLayout(custom_call->shape(), custom_call, 0));
-      } else {
-        for (int64_t i = 0; i < custom_call->operand_count(); ++i) {
-          if (AnyOperandBufferForwarded(custom_call, i)) {
-            TF_RET_CHECK(AllOperandBuffersForwarded(custom_call, i))
-                << "Partial alias of an operand is not supported";
-          } else {
-            TF_RETURN_IF_ERROR(SetOperandLayout(
-                custom_call->operand_shapes_with_layout()[i], custom_call, i));
-          }
-        }
-      }
     } else if (IsLayoutConstrainedCollective(instruction)) {
       TF_RETURN_IF_ERROR(
           SetInstructionLayout(instruction->shape(), instruction));
@@ -2476,6 +2455,33 @@ absl::Status LayoutAssignment::RunOnComputation(
   // Add any backend-specific constraints.
   TF_RETURN_IF_ERROR(AddBackendConstraints(constraints));
 
+  for (HloInstruction* instruction :
+       constraints->computation()->MakeInstructionPostOrder()) {
+    if (!IsLayoutConstrainedCustomCall(instruction)) {
+      continue;
+    }
+    const HloCustomCallInstruction* custom_call =
+        DynCast<HloCustomCallInstruction>(instruction);
+
+    TF_RETURN_IF_ERROR(SetInstructionLayout(custom_call->shape(), custom_call,
+                                            /*mandatory=*/true, /*dfs=*/true,
+                                            /*allow_alias=*/true));
+    if (custom_call->IsCustomCall("LayoutConstraint")) {
+      TF_RETURN_IF_ERROR(
+          SetOperandLayout(custom_call->shape(), custom_call, 0));
+    } else {
+      for (int64_t i = 0; i < custom_call->operand_count(); ++i) {
+        if (AnyOperandBufferForwarded(custom_call, i)) {
+          TF_RET_CHECK(AllOperandBuffersForwarded(custom_call, i))
+              << "Partial alias of an operand is not supported";
+        } else {
+          TF_RETURN_IF_ERROR(SetOperandLayout(
+              custom_call->operand_shapes_with_layout()[i], custom_call, i));
+        }
+      }
+    }
+  }
+
   // Propagates layouts from mandatory and backend constraints.
   TF_RETURN_IF_ERROR(PropagateConstraints(constraints));
 

diff --git a/third_party/xla/xla/service/layout_assignment_test.cc b/third_party/xla/xla/service/layout_assignment_test.cc
@@ -1250,6 +1250,42 @@ ENTRY %CustomCallWithLayoutConstraints (p0: f32[4,4], p1: f32[2,3]) -> f32[1,2,3
   ExpectLayoutIs(custom_call->operand(1)->shape(), {1, 0});
 }
 
+TEST_F(LayoutAssignmentTest, CustomCallLayoutConstrainedAndElementwise) {
+  const char* module_str = R"(
+HloModule CustomCallLayoutConstrained
+
+ENTRY %CustomCallWithLayoutConstraints (p0: f32[4,4], p1: f32[2,3]) -> f32[1,2,3,4] {
+  p0 = f32[4,4] parameter(0)
+  p1 = f32[2,3] parameter(1)
+  cc = f32[1,2,3,4]{3,2,0,1} custom-call(f32[4,4] %p0, f32[2,3] %p1), custom_call_target="baz", operand_layout_constraints={f32[4,4]{0,1}, f32[2,3]{1,0}}
+  ROOT e = f32[1,2,3,4] exponential(cc)
+}
+)";
+  TF_ASSERT_OK_AND_ASSIGN(
+      std::unique_ptr<VerifiedHloModule> m,
+      ParseAndReturnVerifiedModule(module_str, GetModuleConfigForTest()));
+  ComputationLayout computation_layout = m->entry_computation_layout();
+  *computation_layout.mutable_parameter_layout(0) =
+      ShapeLayout(ShapeUtil::MakeShapeWithDenseLayout(F32, {4, 4}, {1, 0}));
+  *computation_layout.mutable_parameter_layout(1) =
+      ShapeLayout(ShapeUtil::MakeShapeWithDenseLayout(F32, {2, 3}, {1, 0}));
+  *computation_layout.mutable_result_layout() = ShapeLayout(
+      ShapeUtil::MakeShapeWithDenseLayout(F32, {1, 2, 3, 4}, {2, 1, 0, 3}));
+  AssignLayouts(m.get(), &computation_layout);
+
+  // The custom call should be partially encapsulated in kCopy instructions
+  // because of the layout mismatches.
+  ASSERT_THAT(
+      m->entry_computation()->root_instruction(),
+      GmockMatch(m::Copy(m::Exp(m::CustomCall(m::Copy(), m::Parameter())))));
+
+  const HloInstruction* custom_call =
+      m->entry_computation()->root_instruction()->operand(0)->operand(0);
+  ExpectLayoutIs(custom_call->shape(), {3, 2, 0, 1});
+  ExpectLayoutIs(custom_call->operand(0)->shape(), {0, 1});
+  ExpectLayoutIs(custom_call->operand(1)->shape(), {1, 0});
+}
+
 TEST_F(LayoutAssignmentTest, CustomCallLayoutConstrainedAliasedOutput) {
   const char* module_str = R"(
 HloModule customcall.4