}

// -----

// CHECK-LABEL: listPushBack
func.func @listPushBack(%arg0: tensor<!tf_type.variant<tensor<?x1xf32>>>, %arg1: tensor<16x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>  {
  %0 = "tf.TensorListPushBack"(%arg0, %arg1) : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<16x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>

  %44 = "tf.AddV2"(%43, %37) {device = ""} : (tensor<2x16x1xf32>, tensor<2x16x1xf32>) -> tensor<2x16x1xf32>
  %45 = "tf.Mul"(%42, %35) {device = ""} : (tensor<2x16x1xf32>, tensor<2x16x1xf32>) -> tensor<2x16x1xf32>
  %46 = "tf.AddV2"(%45, %34) {device = ""} : (tensor<2x16x1xf32>, tensor<2x16x1xf32>) -> tensor<2x16x1xf32>

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

  // CHECK-DAG: %[[LOG_SOFTMAX_LOG:.*]] = "tf.Log"(%[[LOG_SOFTMAX_SUM]]) : (tensor<2x1xf32>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[LOG_SOFTMAX:.*]] = "tf.Sub"(%[[LOG_SOFTMAX_SHIFTED]], %[[LOG_SOFTMAX_LOG]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>

  %2 = "tf.Transpose"(%1, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %2 : tensor<2x1xf32>

// CHECK: %cst = arith.constant
// CHECK: %[[trans:.*]] = "tf.Transpose"
// CHECK-SAME: -> tensor<2x1xf32>
// CHECK: %[[q:.*]] = "tfl.quantize"(%[[trans]]) <{qtype = tensor<2x1x!quant.uniform<u8:f32, 1.000000e+00>>}>

// execute the production code path.
func.func @main(%arg0: tensor<2x1xf32>, %arg1: tensor<2x3xf32>) -> (tensor<2x4xf32>) {
  %0 = "tfl.quantize"(%arg0) {qtype = tensor<2x1x!quant.uniform<i16:f32, 1.0>>} : (tensor<2x1xf32>) -> tensor<2x1x!quant.uniform<i16:f32, 1.0>>
  %1 = "tfl.dequantize"(%0) : (tensor<2x1x!quant.uniform<i16:f32, 1.0>>) -> (tensor<2x1xf32>)

  func.return %3 : tensor<2x1xf32>
}
}

// CHECK:   func @simpleTest(%[[VAL_0:.*]]: tensor<1xf32>, %[[VAL_1:.*]]: tensor<1xf32>, %[[VAL_2:.*]]: tensor<1xf32>, %[[VAL_3:.*]]: tensor<1xf32>) -> tensor<2x1xf32> {

  %cst_0 = arith.constant dense<[[1.5, -2.5]]> : tensor<1x2xf32>
  %cst_1 = arith.constant dense<[[-3.], [4.]]> : tensor<2x1xf32>

  %0 = "tfl.add"(%cst_0, %cst_1) {fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<2x1xf32>) -> tensor<2x2xf32>

  func.return %0 : tensor<2x2xf32>

  %2 = "tfl.pseudo_const"() {value = dense<[[1.8328168392181396], [-1.897219181060791]]> : tensor<2x1xf32>} : () -> tensor<2x1xf32>
  %3 = "tfl.pseudo_const"() {value = dense<[1.4014043807983398, -1.0950859785079956]> : tensor<2xf32>} : () -> tensor<2xf32>

  %1 = "tf.BatchMatMulV2"(%arg0, %0) : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>
  func.return %1 : tensor<4x128x1xf32>

  // CHECK-LABEL: batchmatmul2fullyconnected
  // CHECK-DAG:  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32> 
  // CHECK:  %0 = "tfl.transpose"(%cst, %cst_0) : (tensor<2x1xf32>, tensor<2xi32>) -> tensor<1x2xf32>