} : (tensor<9x10x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<9x10xf32>
  %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> } : ()

// CHECK-NEXT:  [[NEG_ONE_1:%.*]] = arith.constant dense<-1> : tensor<i32>
// CHECK-NEXT:  [[RESULT:%.*]] = "tf.Concat"([[ZERO]], [[INPUT]], [[EXTENDED_PART]]) : (tensor<i32>, tensor<3x10xf32>, tensor<?x10xf32>) -> tensor<?x10xf32>
// CHECK-NEXT:  return [[RESULT]] : tensor<?x10xf32>


// CHECK-LABEL:  func private @cond_false

    %6 = "tfl.reshape"(%5, %cst_1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<200xf32>, tensor<2xi32>) -> tensor<2x100xf32>
    %7 = "tfl.quantize"(%6) {qtype = tensor<2x100x!quant.uniform<i8:f32, 2.000000e-01:-3>>, tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2x100xf32>) -> tensor<2x100x!quant.uniform<i8:f32, 2.000000e-01:-3>>

    : ( tensor<1x28x28xf32>,
        tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>,
        tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20x20xf32>, none,
        tensor<1x20xf32>, tensor<1x20xf32>,

  // CHECK-LABEL: @get_next
  func.func private @get_next(%arg0: tensor<!tf_type.resource>) -> tensor<200x10xf32> {
    // CHECK: %0 = "tf.Iterator"
    // CHECK-SAME: shared_name = "foo_iterator"
    // CHECK: "tf.IteratorGetNext"(%0)
    %0 = "tf.IteratorGetNext"(%arg0) : (tensor<!tf_type.resource>) -> tensor<200x10xf32>
    return %0 : tensor<200x10xf32>
  }

  // CHECK-LABEL: @main
  func.func @main()

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant
// CHECK:  %[[FQ:.*]] = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2)
// CHECK:  %[[R1:.*]] = "tf.Reshape"(%[[FQ]], %cst)
// CHECK-SAME: tensor<2x1xf32>
// CHECK:  %[[R2:.*]] = "tf.Reshape"(%[[FQ]], %cst)

  func.func @matmul(%arg0: tensor<2x12xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<12x2xf32>} : () -> tensor<12x2xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant
// CHECK:  %[[FQ:.*]] = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2)
// CHECK:  %[[R1:.*]] = "tf.Reshape"(%[[FQ]], %cst)
// CHECK-SAME: tensor<2x1xf32>
// CHECK:  %[[R2:.*]] = "tf.Reshape"(%[[FQ]], %cst)