%0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>

  func.func private @quantize_conv_fn(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"} {
    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.XlaCallModule"(%0, %cst) {

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

// RUN: tf-quant-opt %s -quant-convert-fake-quant-to-qdq -quant-lift-quantizable-spots-as-functions -quant-insert-quantized-functions -quant-quantize-composite-functions -symbol-dce | FileCheck %s

func.func @fake_quant_conv(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

    // CHECK-NEXT:  %[[SELU_VAL:.*]] = "tf.Mul"(%[[ELU_VAL]], %[[SCALED_ALPHA]]) : (tensor<1x4x4x3xf32>, tensor<f32>) -> tensor<1x4x4x3xf32>
    // CHECK-NEXT:  %[[RES:.*]] = "tf.SelectV2"(%[[PRED]], %[[SCALED_FEATURES]], %[[SELU_VAL]]) : (tensor<1x4x4x3xi1>, tensor<1x4x4x3xf32>, tensor<1x4x4x3xf32>) -> tensor<1x4x4x3xf32>
    // CHECK-NEXT:  return %[[RES]] : tensor<1x4x4x3xf32>

    %1 = "tf.PartitionedCall"(%arg0, %cst, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_with_bias_and_relu6_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>, tensor<2xf32>) -> tensor<*xf32> loc(callsit...

    rhs_quantization_max_val = 127 : i64
  } : (tensor<1x6x6x3xf32>, tensor<2x3x3x2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1x4x1x2xf32>
  func.return %0 : tensor<1x4x1x2xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantized_convolution_hybrid_same
func.func @uniform_quantized_convolution_hybrid_same(%input: tensor<1x2x2x3xf32>) -> tensor<1x2x1x2xf32> {

  func.return %mm : tensor<1x3x3x3xf32>

// CHECK: %[[mm:.*]] = "tfl.batch_matmul"(%arg0, %arg0) <{adj_x = false, adj_y = true}>
// CHECK: return %[[mm:.*]]

// PerTensor: %[[mm:.*]] = "tfl.batch_matmul"(%arg0, %arg0) <{adj_x = false, adj_y = true}>
// PerTensor: return %[[mm:.*]]
}

// CHECK-LABEL: NotQuantizeConst
// Float16-LABEL: NotQuantizeConst

  // CHECK: return %[[RESULT]] : tensor<2x3x4x6xf32>
}

// -----

func.func @batchMatMulV2Broadcast(%arg0: tensor<2x1x4x5xf32>, %arg1: tensor<1x3x5x6xf32>) -> tensor<2x3x4x6xf32> {
  %0 = "tf.BatchMatMulV2"(%arg0, %arg1) : (tensor<2x1x4x5xf32>, tensor<1x3x5x6xf32>) -> tensor<2x3x4x6xf32>
  func.return %0 : tensor<2x3x4x6xf32>

  // CHECK-LABEL: batchMatMulV2Broadcast

func.func @testGatherV2(%arg0: tensor<16x2x3xf32>, %arg1: tensor<16x5xi32>) -> tensor<16x2x5x3xf32> {
  %0 = "tf.Const"() { value = dense<[-1]> : tensor<1xi32> } : () -> tensor<1xi32>
  %1 = "tf.GatherV2"(%arg0, %arg1, %0) {batch_dims = -1 : i64} : (tensor<16x2x3xf32>, tensor<16x5xi32>, tensor<1xi32>) -> tensor<16x2x5x3xf32>
  func.return %1 : tensor<16x2x5x3xf32>
}

// -----