When checking for the shapes, we should take care of the dynamic shapes.

PiperOrigin-RevId: 678026956
Intel-tensorflow · Sep 24, 2024 · 8dc568e · 8dc568e
1 parent e1b495d
commit 8dc568e
Show file tree

Hide file tree

Showing 2 changed files with 68 additions and 3 deletions.
diff --git a/tensorflow/compiler/mlir/tensorflow/tests/tpu_sharding_identification.mlir b/tensorflow/compiler/mlir/tensorflow/tests/tpu_sharding_identification.mlir
@@ -957,3 +957,47 @@ func.func @func(%arg0: tensor<f32>) -> tensor<4xf32> {
   func.return %1 : tensor<4xf32>
 }
 
+
+// -----
+// CHECK-LABEL: func @check_AddV2_variant_shape_with_input_sharding_propagation
+func.func @check_AddV2_variant_shape_with_input_sharding_propagation(%arg0: tensor<?x12x384xbf16>, %arg1: tensor<12x384xbf16>) {
+  // CHECK:      tf_device.cluster_func
+  // CHECK-SAME: input_sharding_configuration = ["sharding_info_1", "sharding_info_1"]
+  // CHECK-SAME: output_sharding_configuration = ["sharding_info_1"]
+  "tf_device.cluster_func"(%arg0, %arg1) {
+      func = @func,
+      use_spmd_for_xla_partitioning = true, num_cores_per_replica = 1 : i64
+  } : (tensor<?x12x384xbf16>, tensor<12x384xbf16>) -> tensor<?x12x384xbf16>
+  func.return
+}
+
+// CHECK-LABEL: func @func
+// CHECK: {{.*}}mhlo.sharding = "sharding_info_1"{{.*}}mhlo.sharding = "sharding_info_1"{{.*}}->{{.*}}mhlo.sharding = "sharding_info_1"
+func.func @func(%arg0: tensor<?x12x384xbf16>, %arg1: tensor<12x384xbf16>) -> tensor<?x12x384xbf16> {
+  %add = "tf.AddV2"(%arg0, %arg1) : (tensor<?x12x384xbf16>, tensor<12x384xbf16>) -> tensor<?x12x384xbf16>
+  %0 = "tf.XlaSharding"(%add) { _XlaSharding = "sharding_info_1"} : (tensor<?x12x384xbf16>) -> tensor<?x12x384xbf16>
+  func.return %0 : tensor<?x12x384xbf16>
+}
+
+
+
+// -----
+// CHECK-LABEL: func @check_BatchMatMul_variant_shape_without_input_sharding_propagation
+func.func @check_BatchMatMul_variant_shape_without_input_sharding_propagation(%arg0: tensor<?x12x256xbf16>, %arg1: tensor<256x384xbf16>) {
+  // CHECK:      tf_device.cluster_func
+  // CHECK-SAME: input_sharding_configuration = ["", ""]
+  // CHECK-SAME: output_sharding_configuration = ["sharding_info_1"]
+  "tf_device.cluster_func"(%arg0, %arg1) {
+      func = @func,
+      use_spmd_for_xla_partitioning = true, num_cores_per_replica = 1 : i64
+  } : (tensor<?x12x256xbf16>, tensor<256x384xbf16>) -> tensor<?x12x384xbf16>
+  func.return
+}
+
+// CHECK-LABEL: func @func
+// CHECK: {{.*}}mhlo.sharding = ""{{.*}}mhlo.sharding = ""{{.*}}->{{.*}}mhlo.sharding = "sharding_info_1"
+func.func @func(%arg0: tensor<?x12x256xbf16>, %arg1: tensor<256x384xbf16>) -> tensor<?x12x384xbf16> {
+  %mul = "tf.BatchMatMul"(%arg0, %arg1) : (tensor<?x12x256xbf16>, tensor<256x384xbf16>) -> tensor<?x12x384xbf16>
+  %0 = "tf.XlaSharding"(%mul) { _XlaSharding = "sharding_info_1"} : (tensor<?x12x384xbf16>) -> tensor<?x12x384xbf16>
+  func.return %0 : tensor<?x12x384xbf16>
+}
diff --git a/tensorflow/compiler/mlir/tf2xla/internal/passes/tpu_sharding_identification_pass.cc b/tensorflow/compiler/mlir/tf2xla/internal/passes/tpu_sharding_identification_pass.cc
@@ -129,13 +129,33 @@ bool BinaryOpHasTraitsForSharding(Operation* op) {
   return false;
 }
 
-bool DoTypesHaveSameShape(Value value_0, Value value_1) {
+bool DoTypesHavePartialSameShape(Value value_0, Value value_1) {
   auto shape_0 =
       mlir::dyn_cast_or_null<mlir::RankedTensorType>(value_0.getType());
   auto shape_1 =
       mlir::dyn_cast_or_null<mlir::RankedTensorType>(value_1.getType());
   if (shape_0 && shape_1) {
-    return shape_0.getShape() == shape_1.getShape();
+    if (shape_0.hasStaticShape() && shape_1.hasStaticShape())
+      return shape_0.getShape() == shape_1.getShape();
+    int i = 0, j = 0;
+    while (i < shape_0.getShape().size() && j < shape_1.getShape().size()) {
+      if (shape_0.getShape()[i] != shape_1.getShape()[j] &&
+          !shape_0.isDynamicDim(i) && !shape_1.isDynamicDim(j)) {
+        return false;
+      }
+      if (shape_0.getShape()[i] == shape_1.getShape()[j]) {
+        i++;
+        j++;
+      } else {
+        if (shape_0.isDynamicDim(i)) {
+          i++;
+        }
+        if (shape_1.isDynamicDim(j)) {
+          j++;
+        }
+      }
+    }
+    return i == shape_0.getShape().size() && j == shape_1.getShape().size();
   }
   return false;
 }
@@ -337,7 +357,8 @@ std::optional<llvm::StringRef> GetXlaShardingFromArg(
         }
 
         if (BinaryOpHasTraitsForSharding(owner)) {
-          if (DoTypesHaveSameShape(value_to_visit, owner->getResult(0))) {
+          if (DoTypesHavePartialSameShape(value_to_visit,
+                                          owner->getResult(0))) {
             next_values_to_visit.push_back(use.getOwner()->getResult(0));
             continue;
           }