// RUN: tf-mlir-translate -mlir-to-graph %s -o - | FileCheck %s

func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")
// CHECK: node {
// CHECK-NEXT: name: "const1"
// CHECK-NEXT: op: "Const"
// CHECK: dtype: DT_HALF
// CHECK: half_val: 15360

  %2 = "tf.Mul"(%1, %value) {T = "tfdtype$DT_FLOAT"} : (tensor<256x8x7x16xf32>, tensor<16xf32>) -> tensor<256x8x7x16xf32>
  func.return %2 : tensor<256x8x7x16xf32>

// CHECK-DAG: %[[cst:.*]] = "tf.Const{{.*}} dense<8.000000e+00> : tensor<3x3x3x16xf32>
// CHECK-DAG: %[[cst_0:.*]] = "tf.Const{{.*}} dense<1.200000e+01> : tensor<16xf32>

// CHECK: Placeholder
// CHECK: key: "type"
// CHECK: type: DT_INT8

func.func @main(%arg0 : tensor<16xf32>) {
  tf_executor.graph {
    %1:2 = tf_executor.island wraps "tf.MlirPassthroughOp"(%arg0) {extra_type_attr = [tensor<5xi32>, tensor<16xf32>], Tinputs = [tensor<16xf32>], Toutputs = [tensor<16xf32>], mlir_module = ""} : (tensor<16xf32>) -> tensor<16xf32>
    tf_executor.fetch
  }
  func.return
}

func.func @plain() {
  tf_executor.graph {

  %1 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16xf32>} : () -> tensor<16xf32>
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %2 = "tfl.conv_2d"(%arg0, %0, %1) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : 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 = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

func.func @main() -> (tensor<1x2xf16>, tensor<2xf16>) {
  %graph:2 = tf_executor.graph {
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<1.0> : tensor<1x2xf16>} : () -> tensor<1x2xf16> loc("const1")
    %1:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_HALF", value = dense<[1.0, 2.0]> : tensor<2xf16>} : () -> tensor<2xf16> loc("const2")

  %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
}

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DGPU(tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x32x32x16xf32> {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: tac.cost = 0x4C300000