func.return %mm_s : tensor<1x3x3x2xf32>

// CHECK: %[[w:.*]] = arith.constant dense<1.270000e+02> : tensor<512x2xf32>
// CHECK: %[[q_w:.*]] = "tfl.quantize"(%[[w]]) <{qtype = tensor<512x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>}>
// CHECK: %[[dq_w:.*]] = "tfl.dequantize"(%[[q_w]]) : (tensor<512x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>) -> tensor<512x2xf32>

    %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>

  %21 = "tf.Const"() {device = "", name = "Const_143", dtype = "tfdtype$DT_FLOAT", value = dense<0.24288677062973696> : tensor<1x1x6x2xf32>} : () -> tensor<1x1x6x2xf32>
  // CHECK-DAG: value = #tf_type<tensor_proto
  // CHECK-DAG: tf.Const{{.*}} dense<0.242886767> : tensor<1x1x6x2xf32>
  func.return %0, %21 : tensor<4xf32>, tensor<1x1x6x2xf32>
}

// CHECK-LABEL: func @testAdd() -> tensor<2x2xi32>
func.func @testAdd() -> tensor<2x2xi32> {
^bb0:

// %1 = stablehlo.constant dense<4>
// %2 = stablehlo.constant dense<2>
// %3 = stablehlo.convolution(%%arg0, %%arg1) :
//          (tensor<?x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<?x3x4x2xf32>
// %4 = stablehlo.get_dimension_size %3, dim = 0 :
//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :

// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_7]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>

  %0 = stablehlo.constant() {value = dense<3> : tensor<3x3x4x2xi8>} : () -> tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>
  %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>) -> tensor<1x3x3x2xf32>

  %result = "tf_device.launch"() ({
    %3 = "tf.Transpose"(%2, %1) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x8xf32>
    tf_device.return %3: tensor<1x4x4x8xf32>
  }) {device = "device"} : () -> tensor<1x4x4x8xf32>

  func.return %result : tensor<1x4x4x8xf32>

  // CHECK-DAG: %[[CONST1:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>

  %1 = "tf.Const"() {value = dense<[2, 16, 2]> : tensor<3xi32>} : () -> tensor<3xi32>
  %2 = "tf.Const"() {value = dense<[2, 4, 4, 1]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Sub"(%0, %arg1) {device = ""} : (tensor<1x4x4x2xf32>, tensor<2x4x4x2xf32>) -> tensor<2x4x4x2xf32>

    // CHECK-NEXT: %[[CMP:.*]] = mhlo.compare GT, %[[INP]], %[[ZERO]], NOTYPE : (tensor<1x4x4x3xf32>, tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xi1>
    // CHECK-NEXT: %[[RES:.*]] = mhlo.select %[[CMP]], %[[INP]], %[[LEAKY]] : tensor<1x4x4x3xi1>, tensor<1x4x4x3xf32>
    // CHECK-NEXT: return %[[RES]] : tensor<1x4x4x3xf32>
    %0 = "tf.LeakyRelu"(%arg0) {alpha = 2.000000e-01 : f32, device = ""} : (tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xf32>

}

// -----

// CHECK-LABEL: topk_2
func.func @topk_2(%arg0: tensor<3x4x8xf32>) -> (tensor<3x4x2xf32>, tensor<3x4x2xi32>) {
  %0 = arith.constant dense<2> : tensor<i32>
  %1:2 = "tfl.topk_v2"(%arg0, %0) : (tensor<3x4x8xf32>, tensor<i32>) -> (tensor<3x4x2xf32>, tensor<3x4x2xi32>)
  func.return %1#0, %1#1: tensor<3x4x2xf32>, tensor<3x4x2xi32>
}

// -----

// CHECK-LABEL: topk_d