%2059 = "tfl.transpose"(%2058, %cst_6) : (tensor<256x16xf32>, tensor<2xi32>) -> tensor<16x256xf32>
  return %2059: tensor<16x256xf32>
  // CHECK-DAG: %cst = arith.constant dense<[16, 256]> : tensor<2xi32>
  // CHECK: %0 = "tfl.reshape"(%arg0, %cst) : (tensor<1x16x256xf32>, tensor<2xi32>) -> tensor<16x256xf32>
  // CHECK: return %0
}

// CHECK-LABEL: @FoldDoubleTranspose

  %2 = "tfl.batch_matmul"(%1, %arg2) {adj_x = true, adj_y = false, asymmetric_quantize_inputs = false} : (tensor<4x8xf32>, tensor<4x256xf32>) -> tensor<8x256xf32>
  func.return %2 : tensor<8x256xf32>
  // CHECK: return %[[RES1]] : tensor<8x256xf32>
}

// CHECK-LABEL: Batchmatmul2Fullyconnected
// CHECK-NOT: "tfl.batch_matmul"

// CHECK:         return %1 : tensor<1x?x256xf32>
// CHECK:         }
func.func @dynamic_update_slice_operand_has_dynamic_dim(%arg0: tensor<1x?x256xf32>, %arg1: tensor<1x1x256xf32>, %arg2: tensor<i32>, %arg3: tensor<i32>, %arg4: tensor<i32>) -> tensor<1x?x256xf32> {

    %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>

  // CHECK: %[[X_CENTER:.+]] = mhlo.subtract %[[X]], %[[MEAN_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[X_SCALED:.+]] = mhlo.multiply %[[X_CENTER]], %[[SCALE_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[X_NORMED:.+]] = mhlo.divide %[[X_SCALED]], %[[STDDEV_BCAST]] : tensor<4x256xf32>
  // CHECK: %[[RESULT:.+]] = mhlo.add %[[X_NORMED]], %[[OFFSET_BCAST]] : tensor<4x256xf32>

  func.return %0#2, %0#3, %0#4 : tensor<i32>, tensor<256x256xf32>, tensor<?x256x256xf32>
}

      "tf.Yield"(%0) : (tensor<512x256xf32>) -> ()
    }, {
      "tf.Yield"(%arg2) : (tensor<512x256xf32>) -> ()
    }) { is_stateless = true}: (tensor<i1>) -> tensor<512x256xf32>
    "tf.Yield"(%1) : (tensor<512x256xf32>) -> ()
  }, {
    "tf.Yield"(%arg2) : (tensor<512x256xf32>) -> ()
  }) { is_stateless = true}: (tensor<i1>) -> tensor<512x256xf32>
  func.return %2 : tensor<512x256xf32>

%2, %3, %3, %3, %3, %3, %3, %3, %5, %5, %4, %4) {_tflite_input_indices = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 18, 19], device = ""} : (tensor<28x1x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1x16xf32>, tensor<1x16xf32>)...

  // CHECK-DAG: %[[X_NORMED:.+]] = mhlo.multiply %[[X]], %[[MULTIPLIER_BCAST]] : tensor<4x256xf32>
  // CHECK-DAG: %[[RHS_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[RHS]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>
  // CHECK-DAG: %[[RESULT:.+]] = mhlo.add %[[X_NORMED]], %[[RHS_BCAST]] : tensor<4x256xf32>

  // Mean should compute reduction over NHWC spatial dimensions.
  // CHECK: %[[MEAN_DIMS:.*]] = "tf.Const"() <{value = dense<[1, 2]> : tensor<2xi32>}>
  // CHECK: %[[MEAN:[0-9]*]] = "tf.Mean"(%[[RELU]], %[[MEAN_DIMS]])
  // CHECK-SAME: (tensor<?x56x56x256xf32>, tensor<2xi32>) -> tensor<?x256xf32>
  // CHECK: return %[[MEAN]] : tensor<?x256xf32>

  func.return %16 : tensor<?x256xf32>