// CHECK: %[[BITCAST_CONVERT_0:.*]] = stablehlo.bitcast_convert %[[UQ]] : (tensor<1x!quant.uniform<i8:f32, 1.000000e+00:3>>) -> tensor<1xi8>
// CHECK: %[[BITCAST_CONVERT_1:.*]] = stablehlo.bitcast_convert %[[BITCAST_CONVERT_0]] : (tensor<1xi8>) -> tensor<1x!quant.uniform<i8:f32, 1.000000e+00:3>>

// CHECK: %[[maxpool_i8:.*]] = "tf.MaxPool"(%[[sc1]])
// CHECK-SAME: (tensor<*xi8>) -> tensor<*xi8>
// CHECK: %[[reshape_i8:.*]] = "tf.Reshape"(%[[maxpool_i8]]
// CHECK-SAME: (tensor<*xi8>, tensor<2xi32>) -> tensor<*xi8>
// CHECK: %[[sc2:.*]] = "quantfork.scast"(%[[reshape_i8]])
// CHECK: %[[dq:.*]] = "quantfork.dcast"(%[[sc2]]) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>)

  func.return %11 : tensor<*xf32>
}

// CHECK: %[[maxpool_i8:.*]] = "tf.MaxPool"
// CHECK-SAME: (tensor<*xi8>) -> tensor<*xi8>
// CHECK: %[[reshape_i8:.*]] = "tf.Reshape"(%[[maxpool_i8]]
// CHECK-SAME: (tensor<*xi8>, tensor<2xi32>) -> tensor<*xi8>
// CHECK: %[[scast:.*]] = "quantfork.scast"(%[[reshape_i8]]
// CHECK: %[[matmul:.*]] = "tf.PartitionedCall"(%[[scast]]
// CHECK-SAME: f = @composite_matmul_fn_1

  // Argument has rank 2, but its dimensions are unknown.
  constexpr absl::string_view kArgumentWithUnknownDims = R"mlir(
    func.func @unknown_dims_arg(%arg: tensor<?x?xi8>) -> tensor<?x?xi8> {
      return %arg : tensor<?x?xi8>
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kArgumentWithUnknownDims);
  ASSERT_TRUE(module_op);

  %7 = "tf.opH"() : () -> tensor<*xf32>
  %8 = "tf.opI"() : () -> tensor<*xf32>
  %9 = "tf.opJ"() : () -> tensor<*xi8>
  %10 = "tf.opK"() : () -> tensor<*xi16>
  %11 = "tf.opL"() : () -> tensor<*xi64>

// CHECK-LABEL:   func.func @convert_argmax_bool
func.func @convert_argmax_bool(%arg0: tensor<2xi1>) -> tensor<i32> {
  %0 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<2xi32>
  %1 = mhlo.constant dense<false> : tensor<i1>
  %2 = mhlo.constant dense<0> : tensor<i32>
  %3:2 = mhlo.reduce(%arg0 init: %1), (%0 init: %2) across dimensions = [0] : (tensor<2xi1>, tensor<2xi32>, tensor<i1>, tensor<i32>) -> (tensor<i1>, tensor<i32>)

}

// -----

// CHECK-LABEL: func @convolution_add_add_i8
func.func @convolution_add_add_i8(
    %lhs: tensor<?x3x2x1xi8>, %rhs: tensor<2x1x1x1xi8>,
    %zp_offset: tensor<?x2x2x1xi8>, %bias: tensor<1xi8>
  ) -> tensor<?x2x2x1xi8> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution
  // CHECK-DAG: %[[combined:.*]] = chlo.broadcast_add %[[zp_offset:.*]], %[[bias:.*]]

// operate on quantized values.
//
// Examples from storage to quantized type:
//   i8 -> !quant<"uniform[i8:f32]{1.0}">
//   tensor<4xi8> -> tensor<4x!quant<"uniform[i8:f32]{1.0}">>
//   vector<4xi8> -> vector<4x!quant<"uniform[i8:f32]{1.0}">>
def Quantization_StorageCastOp : Quantization_Op<"scast", [Pure]> {
  let arguments = (ins quant_RealOrStorageValueType:$arg);

    %11 = stablehlo.reshape %2 : (tensor<1x1x1x1xi8>) -> tensor<1xi8>
    %12 = stablehlo.broadcast_in_dim %11, dims = [0] : (tensor<1xi8>) -> tensor<1x3x3x4xi8>
    %13 = stablehlo.convert %12 : (tensor<1x3x3x4xi8>) -> tensor<1x3x3x4xf32>
    %14 = stablehlo.convert %3 : (tensor<3x3x4x4xi8>) -> tensor<3x3x4x4xf32>

          )
      )
      module_str = self._extract_first_xla_call_module_op(
          self._output_saved_model_path
      )
      self.assertTrue(
          re.search('stablehlo.dot_general.*xi8>.*xi8>.*xi32>', module_str)
      )

    # Run quantization the first time, calibration is expected to be skipped.
    output_saved_model_path_2 = self.create_tempdir('output2').full_path