// MixedPrecision-NEXT: %[[q:.*]] = "tfl.quantize"(%arg0)
// MixedPrecision-NEXT: %[[dq:.*]] = "tfl.dequantize"(%[[q]])
// MixedPrecision-NEXT: %[[q_0:.*]] = "tfl.quantize"(%arg1)
// MixedPrecision-NEXT: %[[dq_0:.*]] = "tfl.dequantize"(%[[q_0]])
// MixedPrecision-NEXT: %[[c:.*]] = "tfl.concatenation"(%[[dq]], %[[dq_0]])
// MixedPrecision-NEXT: %[[q_1:.*]] = "tfl.quantize"(%[[c]])

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true" -cse | FileCheck %s
// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true legacy-float-scale=true" -cse| FileCheck --check-prefix=Legacy %s

// CHECK-LABEL: QuantizeLstmCellInput
func.func @QuantizeLstmCellInput(%arg0: tensor<1x28x28xf32>) -> tensor<1x28x20xf32> {
    %cst_2 = "tfl.no_value"() {value = unit} : () -> none

^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)
  %2 = "tf.Transpose"(%1, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>

    if (!combined_scale_constant_op) {
      LLVM_DEBUG(llvm::dbgs()
                 << "Failed to match combined_scale_constant_op.\n");
      return failure();
    }

    // Quantize -> Dequantize following r3.
    auto output_uniform_quantize_call_op = dyn_cast_or_null<func::CallOp>(
        *combined_scale_multiply_op.getResult().user_begin());
    if (!output_uniform_quantize_call_op->hasOneUse()) {

        call_op, result_types, args,
        FlatSymbolRefAttr::get(new_quant_func_name));

    return success();
  }

  // For composite functions followed by Dequantize ops, merges the Dequantize
  // op into the functions by creating quantized functions with float output.
  LogicalResult mergeDequantizeOpFollowingQuantizedFunction(
      TF::PartitionedCallOp call_op, const SmallVector<Value, 4>& args,

  %2 = "tfl.dequantize"(%1) : (tensor<1x384x384x!quant.uniform<i8:f32, 0.003:-128>>) -> tensor<1x384x384xf32>
  %3 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<1x1x384xf32>} : () -> tensor<1x384x384xf32>

  // The original model reshape->custom->custom->squeeze.
  ASSERT_THAT(*float_graph->operators(), SizeIs(4));
  // The resulting model should be:
  // reshape->dequantize->custom->custom->quantize->squeeze.
  ASSERT_THAT(subgraph->operators, SizeIs(6));
  const std::vector<BuiltinOperator> op_codes = {
      BuiltinOperator_RESHAPE,  BuiltinOperator_DEQUANTIZE,

  // before converting TF_Conv to TFL_Conv
  (void)applyPatternsAndFoldGreedily(func, std::move(patterns));

  // Remove the wrapper of the tf.FakeQuant* ops and also insert the
  // tfl.quantize and tfl.dequantize to preserve the quantization parameters.
  // This is done after the first round of optimization to make sure all the
  // min/max operands of the tf.FakeQuant* are constants to be matched. The

foreach BinaryOp = [TFL_DivOp, TFL_MulOp]<Op> in
  defm : FuseMulOrDivWithConv2dOrDepthwiseConv2d<BinaryOp>;


// This pattern applies when the same quantize/dequantize have been used twice
// with the same scale. We want to remove the redundancy.
// TODO(fengliuai): move this to the sanity check of pre-quantize pass.
def eliminate_dq_q_pairs : Pat<
  (TFL_QuantizeOp (TFL_DequantizeOp $in), $qt),
  (replaceWithValue $in),

      "DataFormatDimMap",
      "DataFormatVecPermute",
      "DepthToSpace",
      "DepthwiseConv2dNative",
      "DepthwiseConv2dNativeBackpropFilter",
      "DepthwiseConv2dNativeBackpropInput",
      "Dequantize",
      "Diag",
      "DynamicInfeedEnqueueTupleOp",
      "DynamicInfeedDequeueTupleOp",
      "DynamicStitch",
      "DynamicPartition",
      "Einsum",
      "EmptyTensorList",