// RUN: tf-tfrt-opt -tf-to-tfrt %s | FileCheck %s --dump-input=fail

// _output_shapes and f.* attributes are removed during tf-to-tfrt lowering.
// CHECK-LABEL: func @remove_unused_attr
func.func @remove_unused_attr() {
  // CHECK: %out_op_chain = tfrt_fallback_async.executeop.seq(%arg0) key(0) cost({{.*}}) device("/device:CPU:0") "tf.SomeOp2"()

  // CHECK: device = "xpu"
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  // CHECK: device = "cpu"
  %2 = "tf.Relu"(%1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "cpu"} : (tensor<3x3xf32>) -> tensor<3x3xf32>

    // CHECK: tf_executor.island wraps "tf.VarHandleOp"() <{container = "a", shared_name = "x"}> {_mlir_name = "x", _output_shapes = [#tf_type.shape<>], allowed_devices = [], device = "/device:CPU:0", dtype = i64, shape = #tf_type.shape<>} : () -> tensor<!tf_type.resource<tensor<i64>>>

# Verify importing with _output_shapes enabled works as expected.

node {
  name: "_Arg"
  op: "_Arg"
  attr {
    key: "T"
    value {
      type: DT_STRING
    }
  }
  attr {
    key: "index"
    value {
      i: 0
    }
  }
  attr {
    key: "_output_shapes"
    value {
      list {
        shape {
          dim {
            size: 1

// CHECK-NEXT:   name: "input0"
// CHECK-NEXT:   op: "_Arg"
// CHECK:          key: "T"
// CHECK-NEXT:     value {
// CHECK-NEXT:       type: DT_INT32
// CHECK-NEXT:     }
// CHECK:          key: "_output_shapes"
// CHECK-NEXT:     value {
// CHECK-NEXT:       list {
// CHECK-NEXT:         shape {
// CHECK-NEXT:           dim {
// CHECK-NEXT:             size: 10
// CHECK:        name: "output0"

  %0 = "tf.ReadVariableOp"(%handle) {_output_shapes = ["tfshape$dim { size: 3 }"], device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<3xf32>>>) -> tensor<3xf32>

  std::array<int64_t, 4> output_shape;
  // NHWC <- NCHW
  output_shape[0] = composite_result_shape[0];
  output_shape[1] = composite_result_shape[2];
  output_shape[2] = composite_result_shape[3];
  output_shape[3] = composite_result_shape[1];

  auto input_type = mlir::cast<ShapedType>(old_op->getOperand(0).getType());

  return RankedTensorType::get(output_shape, input_type.getElementType());
}

  // CHECK: device = "gpu"
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  // CHECK: device = "cpu"
  %2 = "tf.Relu"(%1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "cpu"} : (tensor<3x3xf32>) -> tensor<3x3xf32>

// Transpose conv supports hybrid computation with quantized weights.
def FoldQuantWeightsIntoTposeConv : Pat<
  (TFL_TransposeConvOp
    $output_shape,
    (TFL_DequantizeOp $quant_weights),
    $quant_input,
    $bias, $padding, $stride_h, $stride_w, $faf),
  (TFL_TransposeConvOp $output_shape, $quant_weights,
    $quant_input, $bias, $padding, $stride_h, $stride_w, $faf), 

# RUN: tf-mlir-translate -graphdef-to-mlir -tf-enable-shape-inference-on-import=false %s -o - | FileCheck %s

node {
  name: "unnamed"
  op: "foo"
  attr {
    key: "_output_shapes"
    value {
      list {
        shape {
        }
      }
    }
  }
  attr {
    key: "_disable_call_shape_inference"
    value {
      b: true
    }
  }
  experimental_debug_info {
  }
}
node {
  name: "unnamed1"