%1 = "quantfork.stats"(%0) {bitsNum = 8 : i64, layerStats = dense<[-2.0, 2.0]> : tensor<2xf32>, narrowRange = false} : (tensor<2x3xf32>) -> tensor<2x3xf32>
  %2 = stablehlo.convert %1 : (tensor<2x3xf32>) -> (tensor<2x3xf32>)
  func.return %2 : tensor<2x3xf32>
}

// -----

// CHECK-LABEL: func @dot_redundant_stats

  func.return %emb : tensor<2x2xf32>

// CHECK-NEXT: tfl.pseudo_qconst
// CHECK-NEXT: tfl.dequantize
// CHECK-NEXT: tfl.gather

func.func @main(tensor<4x5xbf16>) -> tensor<4x5xbf16> {
^bb0(%arg0: tensor<4x5xbf16>):
  // CHECK-LABEL: @main
  // CHECK:  (tensor<4x5xbf16>) -> tensor<4x5xf32>
  // CHECK-NEXT:  (tensor<4x5xf32>) -> tensor<4x5xbf16>
  %0 = "tfl.cast" (%arg0) : (tensor<4x5xbf16>) -> tensor<4x5xf32> loc("cast1")
  %1 = "tfl.cast" (%0) : (tensor<4x5xf32>) -> tensor<4x5xbf16> loc("cast2")
  func.return %1 : tensor<4x5xbf16>

# CHECK-SAME:             -> (tensor<5x5xf32> {{{.*}}})
# CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["key"]

# CHECK-NEXT: [[R0:%.*]] = "tf.ReadVariableOp"([[ARG0]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<5x3xf32>>>) -> tensor<5x3xf32>
# CHECK-NEXT: [[R1:%.*]] = "tf.ReadVariableOp"([[ARG1]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<3x5xf32>>>) -> tensor<3x5xf32>

// Confirm we can extract type info from reshape

func.func @main() -> tensor<2x2xf32> {
  // CHECK: %[[cst:.*]] = "tfl.pseudo_const"() <{value = dense<2> : tensor<2xi32>}> : () -> tensor<2xi32>
  // CHECK: %{{.*}} = "tfl.reshape"(%{{.*}}, %[[cst]]) : (tensor<4xf32>, tensor<2xi32>) -> tensor<2x2xf32>
  %cst = arith.constant dense<[2, 2]> : tensor<2xi32>

  // CHECK-DAG: return %[[RESULT]]
  func.return %0 : tensor<2x3xf32>
}

// CHECK-LABEL: dequantize_quint8
func.func @dequantize_quint8(%arg0: tensor<2x3x!tf_type.quint8>, %min_range: tensor<f32>, %max_range: tensor<f32>) -> tensor<2x3xf32> {
  // CHECK-NEXT: %[[C255:.*]] = "tf.Const"() <{value = dense<2.550000e+02> : tensor<f32>}>

    } : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[0.000000e+00, 6.000000e+00]> : tensor<2xf32>} : (tensor<2x2xf32>) -> tensor<2x2xf32>
    return %2 : tensor<2x2xf32>
  }

  // CHECK-LABEL: composite_dot_general
  func.func private @composite_dot_general(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> tensor<2x2xf32> {

  %arg1: tensor<1x5xf32>) -> (tensor<1x10x!quant.uniform<i8:f32, 0.2:42>>, tensor<1x5xf32>) {
  %0 = "tf.MyRequantize"(%arg0) : (tensor<1x10x!quant.uniform<i8:f32, 0.1:-128>>) -> tensor<1x10x!quant.uniform<i8:f32, 0.2:42>>
  %1 = "tf.Intermediate"(%arg1) : (tensor<1x5xf32>) -> tensor<1x5xf32>
  func.return %0, %1 : tensor<1x10x!quant.uniform<i8:f32, 0.2:42>>, tensor<1x5xf32>

          %outputs_5, %control_6 = tf_executor.island(%control_4) wraps "tf._XlaHostComputeMlir"() {host_mlir_module = "module {\0A func.func @host_func() -> tensor<1x2xf32> {\0A %0 = \22tf.Const\22() {value = dense<0.1> : tensor<1x2xf32>} : () -> tensor<1x2xf32> \0A return %0 : tensor<1x2xf32>}}", manual_sharding = true, recv_key = "host_compute_channel_1_retvals", send_key = "host_c...

    %3 = "tfl.reshape"(%2, %cst_0) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
    %4 = "tfl.reshape"(%3, %cst_1) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
    return %4 : tensor<2x1xf32>
  }
```

#### Compute Costs Pass
In the compute cost pass, we will essentially compute the cost of each op within