}

func.func @einsum_ellipsis(%arg0: tensor<1x512x128xf32>, %arg1: tensor<128x256xf32>) -> tensor<1x512x256xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "...x,xy->...y"} : (tensor<1x512x128xf32>, tensor<128x256xf32>) -> tensor<1x512x256xf32>
  func.return %0 : tensor<1x512x256xf32>
// CHECK-LABEL: einsum_ellipsis

    %0 = "tfl.batch_matmul"(%arg0, %cst_0) {adj_x = false, adj_y = false} : (tensor<1x256xf32>, tensor<256x256xi8>) -> tensor<1x256xf32>
    %1 = "tfl.batch_matmul"(%0, %cst_1) {adj_x = false, adj_y = false} : (tensor<1x256xf32>, tensor<256x256x!quant.uniform<i8:f32, 1.000000e+00>>) -> tensor<1x256xf32>
    %2:2 = "tfl.while"(%cst_2, %1) ({
    ^bb0(%arg1: tensor<i32>,  %arg2: tensor<1x256xf32>):
      %cst_3 = arith.constant dense<10> : tensor<i32>

    }) {dimensions = dense<1> : tensor<1xi64>} : (tensor<1x256xf32>, tensor<f32>) -> tensor<1xf32>
    %5  = mhlo.add %3, %arg6 : tensor<1xf32>
    "mhlo.return"(%1, %arg3, %arg4, %arg5, %5) : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>) -> ()
  }) : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>) -> (tensor<i32>, tensor<i32>, tensor<i32>, tensor<1x256xf32>, tensor<1xf32>)

  %w = arith.constant dense<127.0> : tensor<512x2xf32>
  %mm = "tfl.batch_matmul"(%0, %w) {adj_x = false, adj_y = false} : (tensor<1x3x3x512xf32>, tensor<512x2xf32>) -> tensor<1x3x3x2xf32>
  %mm_s = "quantfork.stats"(%mm) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<1x3x3x2xf32>) -> tensor<1x3x3x2xf32>

}

// CHECK-LABEL: @FuseFullyConnectedRelu
func.func @FuseFullyConnectedRelu(%arg0: tensor<1x256xf32>, %arg1: tensor<128x256xf32>, %arg2: tensor<128xf32>) -> tensor<1x128xf32> {
  %0 = "tfl.fully_connected" (%arg0, %arg1, %arg2) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x256xf32>, tensor<128x256xf32>, tensor<128xf32>) -> tensor<1x128xf32>
  %1 = "tfl.relu"(%0) : (tensor<1x128xf32>) -> tensor<1x128xf32>