%19 = stablehlo.convert %18 : (tensor<8x16x4xi32>) -> tensor<8x16x4xf32>
    %20 = stablehlo.broadcast_in_dim %9, dims = [0, 1, 2] : (tensor<1x1x1xf32>) -> tensor<8x16x4xf32>
    %21 = stablehlo.multiply %19, %20 : tensor<8x16x4xf32>  // * s1 s2
    %22 = call @uniform_quantize_1(%21, %7, %8) : (tensor<8x16x4xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<8x16x4xi8>

  // CHECK: %[[ADD:.*]] = tfl.add %[[VAR_VAL]], %arg0 {fused_activation_function = "NONE"} : tensor<1x10xf32>
  // CHECK: "tfl.assign_variable"(%[[RESOURCE]], %[[ADD]]) : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()
  // CHECK: %[[RESULT:.*]] = "tfl.read_variable"(%[[RESOURCE]]) : (tensor<!tf_type.resource<tensor<1x10xf32>>>) -> tensor<1x10xf32>

    %4:4 = "tf.CustomAggregator"(%3) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<3x11xf32>) -> (tensor<3x11xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    %5 = stablehlo.add %4#0, %1 : tensor<3x11xf32>
    %6 = stablehlo.multiply %5, %1 : tensor<3x11xf32>

// CHECK-NEXT:  %[[cst:.*]] = arith.constant dense<1> : tensor<i32>
  %cst = arith.constant dense<1> : tensor<i32>
// CHECK-NEXT:  %[[softmax:.*]] = "tfl.softmax"(%arg0) <{beta = 1.000000e+00 : f32}> : (tensor<128x16xf32>) -> tensor<128x16xf32>
  %0 = "tfl.softmax"(%arg0) {beta = 1.000000e+00 : f32} : (tensor<128x16xf32>) -> tensor<128x16xf32>

  %2 = "tfl.pseudo_const"() {value = dense<[[0.2]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %3 = "tfl.pseudo_const"() {value = dense<[[0.3]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %4 = "tfl.pseudo_const"() {value = dense<[[0.4]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %5 = "tfl.pseudo_const"() {value = dense<[[0.5]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

func.func @QuantizedCatsAddRequantsTest(%arg0: tensor<1x1xf32>, %arg1: tensor<1x1xf32>, %arg2: tensor<1x1xf32>, %arg3: tensor<1x1xf32>) -> (tensor<1x4xf32>, tensor<1x3xf32>) {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[-0.440728068, 0.189515018]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %1 = "quantfork.stats"(%arg1) {layerStats = dense<[-0.154693216, 0.26483655]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>

  %1 = "tf.MatMul"(%arg0, %0) {T = f32, device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  func.return %1 : tensor<3x3xf32>
}

// CHECK-LABEL: func @gpu_device
func.func @gpu_device(%arg0: tensor<3x1xf32>, %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<3x3xf32> {
  // CHECK: {{%.*}} = corert.get_op_handler %arg0 "/device:GPU:0"

  %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-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>

  "tfl.custom"(%0, %2) {custom_code = "FlexAssignVariableOp", custom_option = #tfl<const_bytes : "0x1041737369676E5661726961626C654F70003B121041737369676E5661726961626C654F701A001A002A0B0A056474797065120230012A140A0E76616C69646174655F736861706512022800320000024F3E1414042801">} : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()