}

func.func @matmul_batchv2(%arg0: tensor<2x10x15xf32>, %arg1: tensor<15x17xf32>) -> tensor<2x10x17xf32> {
  %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", device = "/device:CPU:0", name = "MatMul", adj_x = false, adj_y = false} :
(tensor<2x10x15xf32>, tensor<15x17xf32>) -> tensor<2x10x17xf32>
  func.return %0 : tensor<2x10x17xf32>

  func.return %0: tensor<2x3xf32>
}

// CHECK-LABEL: testBatchMatMulToV2
func.func @testBatchMatMulToV2(%arg0: tensor<2x3x5xf32>, %arg1: tensor<2x5x7xf32>) -> tensor<2x3x7xf32> {
  // CHECK: "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false, grad_x = false, grad_y = false}> {device = "/job:localhost/replica:0/task:0/device:GPU:0"}

      "AssignVariableOp",
      "AssignVariableXlaConcatND",
      "AvgPool",
      "AvgPool3D",
      "AvgPool3DGrad",
      "AvgPoolGrad",
      "BatchMatMul",
      "BatchMatMulV2",
      "BatchMatMulV3",
      "BatchToSpace",
      "BatchToSpaceND",
      "BesselI0e",
      "BesselI1e",
      "Betainc",
      "BiasAddV1",
      "Bincount",
      "Bucketize",

  llvm::ArrayRef<int64_t> y_batches = y_shape.drop_back(2);

  // Check compatibility of batch dimensions if both input shapes are known.
  // BatchMatMul should have exactly the same batch dimensions and
  // BatchMatMulV2 should have broadcastable batch dimensions.
  //
  // The last two dimensions are non-batch dimensions that don't need to
  // participate in batch dimension compatibility check.
  if (std::is_same<OpT, BatchMatMulOp>()) {

    if target_opset == quant_opts_pb2.UNIFORM_QUANTIZED:
      self.assertFalse(self._contains_op(output_graphdef, 'XlaDotV2'))
      self.assertTrue(self._contains_op(output_graphdef, 'BatchMatMulV2'))
    else:
      self.assertFalse(self._contains_op(output_graphdef, 'XlaDotV2'))
      self.assertTrue(self._contains_op(output_graphdef, 'Einsum'))

  @parameterized.named_parameters(

  let summary = "Batch Matrix Multiply Operator";

  let description = [{
Performs a batched matrix multiplication on the inputs. Follows the
conventions of TensorFlow BatchMatMulV2, with support for unknown dimensions
in the batch dimensions and broadcasting.

    Inputs:
      `inputs[0]`: required: input LHS
      `inputs[1]`: required: input RHS

                                           Value *out_lhs, Value *out_rhs,
                                           PatternRewriter *rewriter) {
  // The dimension structure of the relevant operands to a tf.BatchMatMulV2 is:
  // - lhs: [LHSBATCHDIMS..., LHSROWS, LHSCOLS]
  // - rhs: [RHSBATCHDIMS..., RHSROWS, RHSCOLS]
  // - result: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., LHSROWS, RHSCOLS]

  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;

  let hasCanonicalizer = 1;

  let hasVerifier = 1;
}

def TF_BatchMatMulV2Op : TF_Op<"BatchMatMulV2", [Pure, TF_SameOperandsAndResultElementTypeResolveRef]> {
  let summary = "Multiplies slices of two tensors in batches.";

  let description = [{
Multiplies all slices of `Tensor` `x` and `y` (each slice can be

    *   Remove unused `StringViewVariantWrapper`.
    *   Delete unused `Fingerprint64Map` op registration
    *   Add broadcasting support to `tf.matmul`.
    *   Add C++ Gradient for `BatchMatMulV2`.
    *   Add `tf.math.cumulative_logsumexp` operation.
    *   Add ellipsis (...) support for `tf.einsum()`.
    *   Add expand_composites argument to all `nest.*` methods.
    *   Added `strings.byte_split`.