// -----

// CHECK-LABEL: testDotToDotGeneralMatrixVector
func.func @testDotToDotGeneralMatrixVector(%arg0: tensor<2x3072xf32>, %arg1: tensor<3072xf32>) -> tensor<2xf32> {
  %0 = "mhlo.dot"(%arg0, %arg1) : (tensor<2x3072xf32>, tensor<3072xf32>) -> tensor<2xf32>
  func.return %0 : tensor<2xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

  func.return %0#0, %1#0 : tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>
}
// CHECK: func @customAggregator
// CHECK-NEXT: %[[stats:.*]] = "quantfork.stats"(%arg0) <{layerStats = dense<[-1.000000e-01, 2.000000e-01]> : tensor<2xf32>}> : (tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xf32>
// CHECK-NEXT: return %[[stats]], %arg0

func.func @doNotHandleNoMinMaxCases(%arg0: tensor<8x8x8x8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>) {

  // CHECK: return %[[VAL4]]
  %0 = "tfl.pseudo_const"() { value = dense<[1, 32, 32, 16]> : tensor<4xi32> } : () -> tensor<4xi32>
  %1 = "tfl.pseudo_qconst"() {qtype = tensor<16x!quant.uniform<i32:f32, 1.0>>, value = dense<1> : tensor<16xi32>} : () -> tensor<16x1x1x8x!quant.uniform<i32:f32, 1.0>>

//     : (tensor<1x2xf32>) -> tensor<2xf32>
//   %items1:4 = "tf.Unpack"(%[[INP1]]) {axis = 0 : i64}
//     : (tensor<4x2xf32>) -> (tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
//     tensor<2xf32>)
//   %axis = "tf.Const"() {value = dense<0> : tensor<i64>}
//   %0 = "tf.Pack"(items1#3, items1#2, items1#1, items1#0, %items0, %axis)
//     : (tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,

    %s0 = "tf.Const"() {value = dense<[501, 1, 32, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>
    %s1 = "tf.Const"() {value = dense<[  1, 1,  1, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>

    // CHECK: "tf.BroadcastGradientArgs"
    // CHECK-NOT: _xla_outside_compilation
    %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<4xi32>, tensor<4xi32>) -> (tensor<1xi32>, tensor<3xi32>)
    tf_device.return

  %1 = "tfl.pad"(%arg1, %arg2) : (tensor<2x1x3xf32>, tensor<3x2xi32>) -> tensor<?xf32>
  %2 = "tfl.pad"(%0, %arg2) : (tensor<2x1x3xf32>, tensor<3x2xi32>) -> tensor<?xf32>
  func.return %1, %2 : tensor<?xf32>, tensor<?xf32>

// CHECK: %[[dq:.*]] = "tfl.dequantize"(%arg0)
// CHECK: %[[pad1:.*]] = "tfl.pad"(%arg1, %arg2)
// CHECK: %[[pad2:.*]] = "tfl.pad"(%[[dq]], %arg2)

}

// -----

// 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:.*]]

// CHECK-LABEL: QuantizeCustomOp
// CustomOp-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w_1 = arith.constant dense<127.0> : tensor<4096x1x1x1xf32>

func.func @no_qdq_for_non_weight_constant(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {tf._original_func_name = "main_0"} {
  %cst = "tf.Const"() {value = dense<4.000000e-02> : tensor<3xf32>} : () -> tensor<3xf32>
  %0 = "tf.XlaCallModule"(%arg0, %arg1, %cst) {
    Sout = [#tf_type.shape<1x3>], _entry_function = @composite_dot_general_with_bias_fn,
    _original_entry_function = "composite_dot_general_with_bias_fn",

    tensor type, the shape shouldn't be changed during the conversion.

    Example:

    ```mlir
    %1 = tfr.constant_tensor(%0) : f32 -> tensor<f32>
    %3 = tfr.constant_tensor(%2) : vector<1xf32> -> tensor<1xf32>
    ```
  }];

  let arguments = (ins TFR_AllAttrTypes:$arg);

  let results = (outs TFR_singleTensorType:$out);

  let hasCanonicalizer = 1;

  let hasVerifier = 1;
}