Intel® Hyperflex™ Architecture High-Performance Design Handbook

ID 683353
Date 10/04/2021
Public

A newer version of this document is available. Customers should click here to go to the newest version.

Document Table of Contents

8.2.2.1. Example for Broadcast Control Signals

Broadcast control signals that fan-out to many destinations limit retiming. Asynchronous clears can limit retiming due to device support of specific register control signals. However, even synchronous signals, such as synchronous clear and clock enable, can limit retiming when part of a short path or long path critical chain. The use of a synchronous control signal is not a limiting reason by itself; rather the structure and placement of the circuit causes the limit.

A register must be available on all of the node’s inputs to forward retime a register over a node. To retime register A over register B in the following diagram, the Compiler must pull a register from all inputs, including register C on the clock enable input. Additionally, if the Compiler retimes a register down one side of a branch point, the Compiler must retime a copy of the register down all sides of a branch point. This requirement is the same for conventional retiming and Hyper-Retiming.

Figure 141. Retiming through a Clock Enable


There is a branch point at the clock enable input of register B. The branch point consists of additional fan-out to other destinations besides the clock enable. To retime register A over register B, the operation is the same as the previous diagram. However, the presence of the branch point means that a copy of register C must retime along the other side of the branch point, to register C.

Figure 142. Retiming through a Clock Enable with a Branch Point


Retiming Example

The following diagrams combine the previous two steps to illustrate the process of a forward Hyper-Retiming push in the presence of a broadcast clock enable signal or a branch point.

Figure 143. Retiming Example Starting PointHyper-Retiming can move a retimed register into the Hyper-Registers.

Each register’s clock enable has one Hyper-Register location at its input. Because of the placement and routing, the register-to-register path includes three Hyper-Register locations. A different compilation can result in more or fewer Hyper-Register locations. Additionally, there are registers on the data and clock enable inputs to this chain that Hyper-Retiming can retime. These registers exist in the RTL, or you can define them with options that the Pipeline Stages section describes.

One stage of the input registers retime into a Hyper-Register location between the two registers. Figure 144 shows one part of the Hyper-Retiming forward push. One of the registers on the clock enable input retimes over the branch point, with a copy going to a Hyper-Register location at each clock enable input.

Figure 144. Retiming Example Intermediate Point


Figure 145 shows the positions of the registers in the circuit after Hyper-Retiming completes the forward push. The two registers at the inputs of the left register retime to a Hyper-Register location. This diagram is functionally equivalent to the two previous diagrams. The one Hyper-Register location at the clock enable input of the second register remains occupied. There are no other Hyper-Register locations on the clock enable path to the second register, yet there is still one register at the inputs that the Compiler can retime.

Figure 145. Retiming Example Ending Point

Figure 146 shows the register positions Hyper-Retiming uses if a short path/long path critical chain do not limit the path. However, because no Hyper-Registers are available on the right-hand clock enable path, Hyper-Retiming cannot retime the circuit as shown in the diagram.

Figure 146. Retiming Example Limiting condition

Because the clock enable path to the second register has no more Hyper-Register locations available, the Compiler reports this as the short path. Because the register-to-register path is too long to operate at the performance requirement, although having more available Hyper-Register locations for the retimed registers, the Compiler reports this as the long path.

The example is intentionally simple to show the structure of a short path/long path critical chain. In reality, a two-fan-out load is not the critical chain in a circuit. However, broadcast control signals can become the limiting critical chains with higher fan-out. Avoid or rewrite such structures to improve performance.