{"BN",
           {"FusedBatchNorm", "FusedBatchNormV2", "FusedBatchNormV3",
            "_FusedBatchNormEx", "FusedBatchNormGrad", "FusedBatchNormGradV2",
            "FusedBatchNormGradV3"}},
          {"Conv", {"_FusedConv2D"}},
          {"SORT", {"TopKV2"}},  // XLA version much faster then TF version.
          {"MISC",
           // clang-format off
     {"ApproxTopK", "BroadcastTo", "ExpandDims", "Fill", "NoOp",

    // Replace filter uses with uniform quantized filter.
    rewriter.replaceAllUsesWith(filter_op->getResult(0),
                                quantized_filter_constant_op.getResult());

    // Replace conv op with a new convolution op that has quantized output type.
    // Quantize -> Dequantize following r3.
    auto output_uniform_quantize_call_op = cast<func::CallOp>(
        *combined_scale_multiply_op.getResult().user_begin());

    return ...
  }
  func @_func(%input: tensor<2x112x112x12xf32>, %filter: tensor<7x7x3x64xf32>) {
    %filter_transform = "tf.Pad/tf.Transpose/tf.Reshape"(%filter): tensor<7x7x3x64xf32>) -> tensor<4x4x12x64xf32>
    %conv = "tf.Conv2D"(%input, %filter_transfrom) {strides = [1, 1, 1, 1]}: (tensor<2x112x112x12xf32>, tensor<4x4x12x64xf32>) -> tensor<2x112x112x64xf32>
  }
}
```