// Note: Argument quantization sequence omitted.
// CHECK: stablehlo.dot_general %{{.+}}, %{{.+}}, contracting_dims = [1] x [0] : (tensor<1x1024xi8>, tensor<1024x3xi8>) -> tensor<1x3xi32>

// Note: Result dequantization sequence omitted.
// CHECK: return %{{.+}} : tensor<1x3xf32>

// -----

// Tests that a simple dot_general (lifted as a function) with CustomAggregators

        }) : () -> tensor<f32>
        "tf_device.cluster"() ({
          %group = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>
          %arg1_reduced = "tf.XlaAllReduce"(%arg1_id, %group) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
          "tf.OpA"(%arg1_reduced) : (tensor<f32>) -> ()
          tf_device.return
        }) : () -> ()

func.func @QuantizeSlice(tensor<2x3x5x!quant.uniform<u8:f32, 0.1>>, tensor<3xi32>, tensor<3xi32>) -> tensor<?x3x5xf32> {
^bb0(%arg0: tensor<2x3x5x!quant.uniform<u8:f32, 0.1>>, %arg1: tensor<3xi32>, %arg2: tensor<3xi32>):
  %0 = "tfl.dequantize"(%arg0) : (tensor<2x3x5x!quant.uniform<u8:f32, 0.1>>) -> tensor<2x3x5xf32>
  %1 = "tfl.slice"(%0, %arg1, %arg2) : (tensor<2x3x5xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<?x3x5xf32>
  func.return %1 : tensor<?x3x5xf32>

  %1 = "quantfork.stats"(%0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],
      [-8.0, 8.0],
      [-0.5, 0.5]
    ]> : tensor<3x2xf32>, axis = 2 : i64
  } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
  func.return %1 : tensor<8x4x3xf32>

// CHECK: %[[CONST_0:.+]] = stablehlo.constant() <{value = dense<{{.*}}> : tensor<2x3xi8>}> : () -> tensor<2x3x!quant.uniform<i8<-127:127>:f32:1, {{.*}}>
// CHECK: %[[CONST_1:.+]] = stablehlo.constant() <{value = dense<{{.*}}> : tensor<1x3xi32>}> : () -> tensor<1x3x!quant.uniform<i32:f32:1, {{.*}}>
// CHECK: %[[UNIFORM_QUANTIZE_0:.+]] = stablehlo.uniform_quantize %[[ARG_0]] : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, {{.*}}>>

}

// Returns a RankedTensorType which is similar to `input_type` but replaces the
// dimension size of `dim` with `dim_size`.  For example,
// `SubstituteRankedTensorTypeDimSize(tensor<3x4xi32>, 1, 2)` returns
// `tensor<3x2xi32>`.
static RankedTensorType SubstituteRankedTensorTypeDimSize(
    RankedTensorType input_type, int64_t dim, int64_t dim_size) {
  auto shape = input_type.getShape().vec();
  shape[dim] = dim_size;

  %19 = "tf.StringLength"(%arg0) {device = "", unit = "BYTE"} : (tensor<1x!tf_type.string>) -> tensor<1xi32>
  %20 = "tf.ExpandDims"(%19, %7) {device = ""} : (tensor<1xi32>, tensor<i32>) -> tensor<1x1xi32>
  %21 = "tf.Cast"(%20) {Truncate = false, device = ""} : (tensor<1x1xi32>) -> tensor<1x1xi64>
  %22 = "tf.Reshape"(%21, %12) {device = ""} : (tensor<1x1xi64>, tensor<1xi64>) -> tensor<1xi64>