%0 = "tfl.conv_2d"(%arg0, %arg1, %cst_0) {dilation_h_factor = 2 : i32, dilation_w_factor = 3 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 4 : i32, stride_w = 5 : i32} : (tensor<256x32x32x3xf16>, tensor<16x3x3x3xf16>, tensor<16xf16>) -> tensor<256x8x7x16xf16>

// Float16-DAG: %[[b:.*]] = arith.constant dense<0.000000e+00> : tensor<16xf16>
// Float16-DAG: %[[const:.*]] = "tfl.no_value"() <{value}> : () -> none
// Float16-DAG: %[[dq_w:.*]] = "tfl.dequantize"(%[[w]]) : (tensor<3x3x3x8x16xf16>) -> tensor<3x3x3x8x16xf32>
// Float16-DAG: %[[dq_b:.*]] = "tfl.dequantize"(%[[b]]) : (tensor<16xf16>) -> tensor<16xf32>

  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_5:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

    // CHECK: %[[CAST_1:.*]] = "tf.Cast"(%[[ARG_1]]) <{Truncate = false}> : (tensor<10xf32>) -> tensor<10xbf16>
    // CHECK: %[[RESULT:.*]]:2 = "tf.XlaCallModule"(%[[CAST_0]], %[[CAST_1]], %[[ARG_2]])
    // CHECK-SAME: (tensor<10xbf16>, tensor<10xbf16>, tensor<6xi32>) -> (tensor<10xbf16>, tensor<6xi32>)

	{TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256, 16, 32, 16, ecdheRSAKA, suiteECDHE | suiteTLS12, cipherAES, macSHA256, nil},
	{TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA, 16, 20, 16, ecdheRSAKA, suiteECDHE, cipherAES, macSHA1, nil},
	{TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256, 16, 32, 16, ecdheECDSAKA, suiteECDHE | suiteECSign | suiteTLS12, cipherAES, macSHA256, nil},

      (tensor<4x256xf16>, tensor<256xf16>, tensor<256xf16>, tensor<256xf16>,
        tensor<256xf16>) -> tensor<4x256xf16>
  func.return %0 : tensor<4x256xf16>
}

// Validate that epsilon is overflow
func.func @batchNormInference_f16_overflow(
    %x: tensor<4x256xf16>, %scale: tensor<256xf16>, %offset: tensor<256xf16>,
    %mean: tensor<256xf16>, %variance: tensor<256xf16>)
    -> (tensor<4x256xf16>) {

  %2 = func.call @fold_all_test(%arg0, %0, %1) : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %2 : tensor<256x30x30x16xf32>
}

// ALL-LABEL: @fold_all_test
func.func @fold_all_test(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>) -> tensor<256x30x30x16xf32> {

  pass_config.legalize_custom_tensor_list_ops =
      toco_flags.legalize_custom_tensor_list_ops();
  // Disable the unfolding of the 16x16 TF::BatchMatMulOp to avoid the
  // conversion to an unsupported 16x16 TFL::FullyConnectedOp.
  if (toco_flags.inference_type() == toco::IODataType::QUANTIZED_INT16) {
    pass_config.unfold_batch_matmul = false;
  }

// CHECK:  return
}

func.func @mul_i16(%arg0: tensor<14xi16>, %arg1: tensor<14xi16>) -> tensor<14xi16> {
  %0 = "tf.Mul"(%arg0, %arg1) : (tensor<14xi16>, tensor<14xi16>) -> tensor<14xi16>
  func.return %0: tensor<14xi16>

// CHECK-LABEL: mul_i16
// CHECK:  tfl.mul %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<14xi16>
// CHECK:  return
}

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>