// CHECK-SAME: value = dense<1> : tensor<i32>

// Check that the constants that are input to XlaConvV2's 3rd and 4th
// arguments are not duplicated.
// CHECK-NOT: "tf.Const"()

// Check that the constants used for tf.XlaConvV2 and tf.AddV2s are different.

// CHECK: %[[round:.*]] = "tf.Round"(%[[minimum]]
// CHECK: %[[quant:.*]] = "tf.Cast"(%[[round]]) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xi8>
// CHECK: %[[pad:.*]] = "tf.PadV2"(%[[quant]]
// CHECK: %[[xlaconv:.*]] = "tf.XlaConvV2"(%[[pad]]
// CHECK: %[[sub:.*]] = "tf.Sub"(%[[xlaconv]]
// CHECK: %[[cast:.*]] = "tf.Cast"(%[[sub]]) <{Truncate = false}> : (tensor<1x3x2x2xi32>) -> tensor<1x3x2x2xf32>
// CHECK: %[[dequant1:.*]] = "tf.Mul"(%[[cast]]

                              .Device(DEVICE),                                 \
                          XlaCompileOnDemandOp);                               \
  REGISTER_KERNEL_BUILDER(Name("XlaConvV2")                                    \
                              .HostMemory("window_strides")                    \
                              .HostMemory("padding")                           \

def IsEinsumOpSupported : Constraint<
  CPred<"IsEinsumOpSupported($0, $1, $2)">,
  "Check if the given einsum op could be converted into a XlaDotV2 op.">;

// Converts inlined Conv2D pattern to TF XlaConvV2 op. This pattern doesn't
// support non-constant weights.
def ConvertTFConv2DToXLAConvOp : Pat<
  (TF_Conv2DOp:$conv
    (TF_SubOp (TF_CastOp $input, $truncate), $input_zp),

    %strides = "tf.Const"() {value = dense<[3, 1, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
    // expected-error @below {{'tf.XlaConvV2' op feature_group_count must be a positive number, got -1}}

// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<-237772800> : tensor<1x1x1x512xi32>}> : () -> tensor<1x1x1x512xi32>
// CHECK: %[[PADV2_0:.*]] = "tf.PadV2"
// CHECK: %[[XLACONVV2_0:.*]] = "tf.XlaConvV2"(%[[PADV2_0]]
// CHECK: %[[SUB_0:.*]] = "tf.Sub"(%[[XLACONVV2_0]], %[[CONST]])

        self._output_saved_model_path_2
    )
    graphdef = loader.get_meta_graph_def_from_tags(tags).graph_def
    if target_opset == quant_opts_pb2.XLA:
      self.assertTrue(
          self._contains_op(graphdef, 'XlaConvV2', node_name='sample/conv2d.*')
      )

    new_outputs = converted_model.signatures[signature_key](
        input=ops.convert_to_tensor(input_data)
    )

// CHECK: %[[PADV2_0:.*]] = "tf.PadV2"({{.*}}, %[[CONST_4]], %[[CONST_5]]) : (tensor<1x3x4x3xi8>, tensor<4x2xi32>, tensor<i8>) -> tensor<1x4x5x3xi8>
// CHECK: %[[XLACONVV2_0:.*]] = "tf.XlaConvV2"(%[[PADV2_0]], %[[CONST_6]], %[[CONST_0]], %[[CONST_3]], %[[CONST_1]], %[[CONST_1]], %[[CONST_2]])

  %strides = "tf.Const"() {value = dense<[3, 1, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
  // / expected-error@+1 {{'tf.XlaConvV2' op expects feature_group_count to be a scalar}}

    new absl::flat_hash_set<std::string>{"XlaBroadcastHelper",
                                         "XlaCallModule",
                                         "XlaConv",
                                         "XlaConvV2",
                                         "XlaDequantize",
                                         "XlaDot",
                                         "XlaDotV2",
                                         "XlaDynamicSlice",