func.return %1 : tensor<16x128xf32>
  // CHECK: return %0 : tensor<16x128xf32>
}

// CHECK-LABEL: FuseTransposeFCLhsToBatchMatmul
func.func @FuseTransposeFCLhsToBatchMatmul(%arg0: tensor<1024x4xf32>, %arg1: tensor<8x1024xf32>, %arg2: tensor<4x256xf32>) -> tensor<8x256xf32> {
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>

  %w = arith.constant dense<127.0> : tensor<4x12xf32>
  %b = arith.constant dense<0.0> : tensor<4xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<4x12xf32>, tensor<4xf32>) -> tensor<1x112x112x4xf32>
  func.return %fc : tensor<1x112x112x4xf32>

// CHECK:           %[[VAL_14:.*]] = "tfl.reshape"(%[[VAL_13]], %[[VAL_8]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x384x128xf32>, tensor<2xi32>) -> tensor<384x128xf32>
// CHECK:           %[[VAL_15:.*]] = "tfl.reshape"(%[[VAL_14]], %[[VAL_5]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x128xf32>, tensor<3xi32>) -> tensor<1x384x128xf32>

  func.return %0 : tensor<8x16xf32>

// CHECK-LABEL:minimum
// CHECK:  "tfl.minimum"(%arg0, %arg1) : (tensor<8x16xf32>, tensor<8x16xf32>) -> tensor<8x16xf32>
}

func.func @realDiv(%arg0: tensor<8x16xf32>, %arg1: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.RealDiv"(%arg0, %arg1) : (tensor<8x16xf32>, tensor<8x16xf32>) -> tensor<8x16xf32>

}

func.func @matmulNoTransposeAOrB(%arg0: tensor<1x1280xf32>, %arg1: tensor<1280x1000xf32>) -> tensor<1x1000xf32> {
  %166 = "tf.MatMul"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", _output_shapes = ["tfshape$dim { size = 1} dim { size = 1000}"], device = "", name = "matmul", transpose_a = false, transpose_b = false} : (tensor<1x1280xf32>, tensor<1280x1000xf32>) -> tensor<1x1000xf32>
  func.return %166 : tensor<1x1000xf32>

// CHECK-SAME: (tensor<1x37xf32>, tensor<40x37xf32>, none) -> tensor<1x40xf32>
  func.return %1 : tensor<1x40xf32>
}

// CHECK-LABEL: RemoveLstmQuantZeroBias
func.func @RemoveLstmQuantZeroBias(
  %arg0: tensor<1x528xf32>,
  %arg1: tensor<2048x528xf32>,
  %arg2: tensor<2048x528xf32>,
  %arg3: tensor<2048x528xf32>,
  %arg4: tensor<2048x528xf32>,
  %arg5: tensor<2048x640xf32>,
  %arg6: tensor<2048x640xf32>,

    func.return %0 : tensor<4x128xf32>
  }
  func.func @_func(%arg0: tensor<4x128xf32>) -> tensor<4x128xf32> {
    func.return %arg0 : tensor<4x128xf32>
  }

  %split_dim = arith.constant dense<0> : tensor<i32>
  // expected-error @+1 {{'tfl.split' op output #0 should be 'tensor<8x4xf32>'}}
  %0, %1 = "tfl.split"(%split_dim, %arg0) {num_splits = 2 : i32} : (tensor<i32>, tensor<16x4xf32>) -> (tensor<8x2xf32>, tensor<8x2xf32>)
  func.return %0, %1 : tensor<8x2xf32>, tensor<8x2xf32>
}

// -----

  // CHECK:     [[ENTRY_SELECT:%.*]] = mhlo.select [[COMPARE_GT]], [[FEATURES]], [[ELSE_SELECT]]
  %0 = "tf.Softplus"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>

  // CHECK:     return [[ENTRY_SELECT]] : tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
}

// -----

// CHECK-LABEL: func @softplus_f64
// CHECK-SAME: ([[FEATURES:%.*]]: tensor<8x16xf64>)