// specs.
absl::flat_hash_set<int> GetQuantizableOperands(Operation* op) {
  absl::flat_hash_set<int> quantizable_operands;
  if (isa<TF::DepthwiseConv2dNativeOp, TF::Conv2DOp, TF::Conv3DOp, TF::MatMulOp,
          TF::BatchMatMulOp>(op)) {
    quantizable_operands.insert(1);
  } else if (isa<TF::GatherOp>(op)) {
    quantizable_operands.insert(0);
  } else if (auto einsum_op = dyn_cast<TF::EinsumOp>(op)) {

  RewritePatternSet patterns(&getContext());
  auto *ctx = &getContext();
  auto func = getOperation();
  patterns.add<DequantizeConverter<TFL::Conv2DOp>,
               DequantizeConverter<TFL::Conv3DOp>,
               DequantizeConverter<TFL::DepthwiseConv2DOp>,
               DequantizeConverter<TFL::FullyConnectedOp>,
               DequantizeConverter<TFL::TransposeConvOp>>(ctx);

// func.func func_name_${key1}_fn (...) {
//   ...${key2}...
// }
// ```
// The above template with generate two functions by substituting `key1` and
// `key2` with given values.

module {

  for main_op in ["Conv2D", "DepthwiseConv2D", "MatMul"] {
    parameters[
      {"quantized_ops": ["${main_op}", "BiasAdd"], "act_func": "internal_requantize_no_activation_fn", "output_type": "!tf_type.qint8"},

    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  // away ops that can't get constant folded after PrepareTF pass. For example,
  // tf.Conv2D is split into tf.Transpose and tfl.Conv2D.
  pass_manager->addNestedPass<mlir::func::FuncOp>(
      mlir::createCanonicalizerPass());
  pass_manager->addNestedPass<mlir::func::FuncOp>(mlir::createCSEPass());
  // This pass does dead code elimination based on symbol visibility.

                      weight_new_shape, &rewriter);

  // Replace the fc with conv.
  // The output would be [1, 1, width, output].
  auto conv_output_type = RankedTensorType::get({1, 1, width, output_size},
                                                output_type.getElementType());
  auto conv = rewriter.create<TFL::Conv2DOp>(
      fc_op.getLoc(), conv_output_type, reshaped_input, reshaped_weight,

 *
 * The details object should effectively provide the identifying context for the operation.
 * This is information that is known before the operation is executed.
 *
 * The result object should convey the outcome.
 * This is information that is known after the operation is executed.
 *
 * These details and result types need to maintain backwards binary compatibility.

  %2 = "tfl.pseudo_qconst"() {qtype = tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>, value = dense<0> : tensor<32xi32>} : () -> tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>

  TargetHardwareOpRegistration<CpuHardware, Op> Op##_CpuHardware_hardware( \
      Create);

// Operation costs on CPU

// Currently used for these ops:
// tfl.conv_2d / tfl.depthwise_conv_2d / tfl.fully_connected
class CpuConvOp : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override {
    float cost = 0.0;
    int64_t arithmetic_count;

	n.X = walkExpr(n.X, &init)
	call := walkExpr(mkcall1(chanfn("chanrecv1", 2, n.X.Type()), nil, &init, n.X, typecheck.NodNil()), &init)
	return ir.InitExpr(init, call)
}

func convas(n *ir.AssignStmt, init *ir.Nodes) *ir.AssignStmt {
	if n.Op() != ir.OAS {
		base.Fatalf("convas: not OAS %v", n.Op())
	}
	n.SetTypecheck(1)

	if n.X == nil || n.Y == nil {
		return n
	}

	lt := n.X.Type()
	rt := n.Y.Type()