module {
    func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

  %0 = "tf.MatMul"(%input, %filter_new) {
  } : (tensor<8x9xf32>, tensor<9x10xf32>) -> tensor<8x10xf32>
  return %0 : tensor<8x10xf32>
})mlir";
  constexpr absl::string_view kProgram = R"mlir(
func.func @main(%input: tensor<8x9xf32>, %filter: tensor<9x10xi8>) -> tensor<8x10xf32> {
  %filter_scale = "tf.Const"() { value = dense<0.0235> : tensor<f32> } : ()
  -> tensor<f32>

// CHECK-SAME:                      %[[VAL_0:.*]]: tensor<8x128xf32>,
// CHECK-SAME:                      %[[VAL_1:.*]]: tensor<f32>) -> tensor<7x128xf32> {
// CHECK:           %[[VAL_2:.*]] = arith.constant dense<{{\[\[}}0, 0], [0, 1]]> : tensor<2x2xi64>
// CHECK:           %[[VAL_3:.*]] = "tf.PadV2"(%[[VAL_0]], %[[VAL_2]], %[[VAL_1]]) : (tensor<8x128xf32>, tensor<2x2xi64>, tensor<f32>) -> tensor<8x129xf32>

    func.return %0 : tensor<*xf32>
  }

  // CHECK-LABEL: func.func private @main_00(%arg0: tensor<?x1024xf32>) -> tensor<?x3xf32>
  func.func private @main_00(%arg0: tensor<?x1024xf32>) -> tensor<*xf32> attributes {tf._original_func_name = "main_0"} {
    %cst = "tf.Const"() <{value = dense<1.000000e+00> : tensor<1024x3xf32>}> : () -> tensor<1024x3xf32>

// CHECK-NEXT:    %[[RESHAPED_1:.*]] = mhlo.reshape %arg1
// CHECK-NEXT:    %[[BMM_0:.*]] = "tfl.batch_matmul"(%[[RESHAPED_0]], %[[RESHAPED_1]]) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : {{.*}} -> tensor<1x1024xf32>
// CHECK-NEXT:    %[[RESHAPED_BMM:.*]] = mhlo.reshape %[[BMM_0]]
// CHECK-NEXT:    return %[[RESHAPED_BMM]] : tensor<1x1x1024xf32>
}

      tf_device.return %5 : tensor<512xi32>
    }) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> (tensor<512xi32>, tensor<1024xi32>, tensor<1024xi32>, tensor<1024xf32>)
    // CHECK: %[[COMPILE_OUTPUT:[0-9]*]]:4 = "tf_device.launch"
    // CHECK: "tf._TPUCompileMlir"()
    // CHECK: is_bounded_dynamic_dim: true