%cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", device = "", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

  %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_1 = "tf.Const"() {value = dense<0.200000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_2 = "tf.Const"() {value = dense<0.300000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>

    %cst_0 = "tf.Const"() {value = dense<[-2.000000e+00, 3.000000e+00]> : tensor<2xf32>} : () -> tensor<2xf32>
    %0 = "quantfork.qcast"(%cst) : (tensor<2x2x3x2xf32>) -> tensor<2x2x3x2x!quant.uniform<i8<-127:127>:f32:3, {4.000000e-03,5.000000e-03}>>

  module {
    func.func @main(%arg0: tensor<2x2x2xf32>, %arg1: tensor<2x2x2xf32>) -> tensor<2x2x2xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1,
        batching_dims = [0] x [0],
        contracting_dims = [2] x [1],
        precision = [DEFAULT, DEFAULT]
      : (tensor<2x2x2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
        return %0 : tensor<2x2x2xf32>
    }
  }
)mlir";

  %w = arith.constant dense<[[[[-1.0, 1.0]]], [[[1.0, 2.0]]], [[[-2.0, 1.0]]]]> : tensor<3x1x1x2xf32>
  %b = arith.constant dense<[1.0e-2, 2.1473647e1, -2.1473647e2]> : tensor<3xf32>
  %transpose_conv = "tfl.transpose_conv"(%arg1, %w, %0, %b) {
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, fused_activation_function = "NONE"
  } : (tensor<4xi32>, tensor<3x1x1x2xf32>, tensor<1x5x5x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>

            index_vector_dim = 2>,
          slice_sizes = array<i64: 1, 2, 2>,
          indices_are_sorted = false
        } : (tensor<3x4x2xf32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32>
        func.return %2 : tensor<2x3x2x2xf32>
      }
    }
  )mlir";
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(kModuleGather);
  ASSERT_TRUE(module_op);

  func::FuncOp main_fn = FindMainFuncOp(*module_op);

  return %2 : tensor<?x2x2x1xi8>
}

// -----

// CHECK-LABEL: func @convolution_add_add_f32
func.func @convolution_add_add_f32(
    %lhs: tensor<?x3x2x1xf32>, %rhs: tensor<2x1x1x1xf32>,
    %zp_offset: tensor<?x2x2x1xf32>, %bias: tensor<1xf32>
  ) -> tensor<?x2x2x1xf32> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution
  // CHECK-DAG: %[[combined:.*]] = chlo.broadcast_add %[[conv:.*]], %[[zp_offset:.*]]

  } : (tensor<1x6x6x3xf32>, tensor<2x3x3x2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1x4x1x2xf32>
  func.return %0 : tensor<1x4x1x2xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantized_convolution_hybrid_same
func.func @uniform_quantized_convolution_hybrid_same(%input: tensor<1x2x2x3xf32>) -> tensor<1x2x1x2xf32> {

  return %0 : tensor<1x1x1x2xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_5(%arg0: tensor<1x1x1x7xf32>) -> tensor<1x1x1x2xf32>

// CHECK-LABEL: avg_pool2d_6
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>

// CHECK:           %[[VAL_7:.*]] = "tfl.concatenation"(%[[VAL_5]], %[[VAL_6]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>
// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_7]], %[[VAL_3]]) : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>
// CHECK:           %[[VAL_9:.*]] = "tfl.reshape"(%[[VAL_8]], %[[VAL_4]]) : (tensor<2xf32>, tensor<2xi32>) -> tensor<2x1xf32>