// CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_4:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_5:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>

    %0 = "tfl.dequantize"(%arg0) : (tensor<1x224x224x3x!quant.uniform<u8:f32, 1.0>>) -> tensor<1x224x224x3xf32>
    %1 = "tfl.dequantize"(%arg1) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 1.0>>) -> tensor<32x3x3x3xf32>
    %2 = "tfl.dequantize"(%arg2) : (tensor<32x!quant.uniform<i32:f32, 1.0>>) -> tensor<32xf32>

    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    Value value = tf_op.getOutputs();
    auto quantize = rewriter.create<TFL::QuantizeOp>(
        tf_op.getLoc(), qtype.getValue(), value, qtype);
    auto dequantize = rewriter.create<TFL::DequantizeOp>(
        tf_op.getLoc(), res_type, quantize.getOutput());

  }
};

// Inserts a "tfl.quantize" and "tfl.dequantize" op pair (QDQs) after the
// "tf.FakeQuantWithMinMaxVarsOp" to be constant folded. Since the constant
// folding logic will use a "arith.constant" op to replace the
// "tf.FakeQuantWithMinMaxVarsOp", the "tfl.quantize" op is used to preserve
// the quantization parameters as a TypeAttr and "tfl.dequantize" op used to

// Dequantize ops will produce 3x larger tensors, so we want to move it after
// some passthrough ops to reduce the memory consumption.
struct PushDownDequantize : public OpRewritePattern<DequantizeOp> {
  explicit PushDownDequantize(MLIRContext* context)
      : OpRewritePattern<DequantizeOp>(context) {}

  LogicalResult matchAndRewrite(DequantizeOp dequantize_op,

  }
  TypeAttr type_attr = TypeAttr::get(new_type);
  auto quantize = builder.create<TFL::QuantizeOp>(value.getLoc(), new_type,
                                                  value, type_attr);
  auto dequantize = builder.create<TFL::DequantizeOp>(
      value.getLoc(), expressed_type, quantize.getOutput());
  value.replaceAllUsesWith(dequantize);

  // `quantize` is using `dequantize` now, so we should set its operand to
  // `value`.

// CHECK:   %{{.*}} = "tfl.dequantize"(%{{.*}}) : (tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>) -> tensor<1x401408xf32>

  %cst = arith.constant dense<[1, 401408]> : tensor<2xi32>

include "tensorflow/compiler/mlir/lite/ir/tfl_ops.td"

// Quantize attribute $0 by using quantization parameter from %1.
def QuantizeByQuantizedType : NativeCodeCall<"quant::Quantize($0, $1.getValue())">;
def F32ElementsAttr : ElementsAttrBase<
  CPred<"$_self.cast<ElementsAttr>().getShapedType().getElementType().isF32()">, "float constant tensor">;

// Squash tfl.dequantize and tfl.quantize pairs.

    auto op_before_dequantize = original_dequantize_op.getOperand(0);

    // Create a new dequantize op that is propagated.
    rewriter.setInsertionPointAfter(user_op);
    TF::PartitionedCallOp new_dequantize_op =
        cast<TF::PartitionedCallOp>(rewriter.clone(*original_dequantize_op));

    // Skip the original dequant op and connect the op before dequantize to the
    // user op.
    user_op->setOperand(user_idx, op_before_dequantize);

    // Finally, use the quantization parameter to create the quantize and
    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    auto quantize = rewriter.create<quantfork::QuantizeCastOp>(
        tf_op.getLoc(), qtype.getValue(), input);
    auto dequantize = rewriter.create<quantfork::DequantizeCastOp>(
        tf_op.getLoc(), res_type, quantize.getResult());