Clock gating is one of the most frequently used techniques in RTL to reduce dynamic power consumption without affecting the functionality of the design. One method involves inserting gating conditions in the RTL, which the synthesis tool translates to clock gating cells in the clock-path of a register bank. This helps to reduce the switching activity on the clock network, thereby reducing dynamic power consumption in the design. Since the translation done by the synthesis tool is purely combinational, it is referred to as combinational clock gating. This transformation does not alter the behavior of the register being gated.
In the example shown in Figure 1, the netlist generated by the RTL synthesis tool from the Verilog code snippet is shown at the top right. Register Q loads new data when the EN signal is HIGH; otherwise, it holds the data. Opportunities to insert combinational clock gating can be found by looking for conditional assignments in the code. Clock gating logic is substituted when code like if (cond) out <= in is present. Power aware logic synthesis tools identify RTL coding patterns and make the appropriate substitution. This is shown in the circuit at the bottom right of Figure 1.
Combinational clock-gating, though a potent power saving optimization, does not cover various redundant activities in the design that cause wasted computations and power. In some scenarios writes done to a register are either unobservable down-stream or have the same value in consecutive cycles; that is, they are redundant. Identifying such redundant writes requires multi-cycle analysis of the design. The gating of the first type of redundant write is called observability-based clock gating, and the second is called stability-based clock gating.Since such gating conditions alter the behavior of the registers, they are also referred to as sequential clock gating.
Figure 2 is an example of the first type, where under some conditions, writes to a register will never be observed at the design output. If the signal vld_2 is low in a particular cycle, the register dout retains its older value. This means that writes that would have happened in register d_2, one cycle back and in d_1, two cycles back are redundant. Observability-based clock gating identifies this redundant write and adds a suitable gating condition based on the vld signal.
Figure 3 shows an example of the second type of redundant writes. This is where the same value gets written to registers across consecutive clock cycles. When the signals vld_1 and vld_2 are low, registers f_1 and g_1 retain their previously held values. Consequently, values written into registers f_2 and g_2 in the next clock cycle are identical to what was written into them in the previous cycle. Stability-based clock gating identifies the redundant write and adds a suitable gating condition using a one-cycle delayed version of vld_1 for f_2 (and similarly, a one-cycle delayed version of vld_2 for g_2).
Consider an example of a sequential clock gating transformation. In the RTL code snippet shown in the top left corner of Figure 4, registers q0 and q1 are latching a new data value every cycle. Hence, when taken through low-power RTL synthesis tools, they would not have clock gating. If we observe carefully, we notice that if en is HIGH, then the data latched into q1 in the previous cycle is not used. Thus, we can hold the previous data on q1 during that cycle. By performing this sequential reasoning, we can identify ~SEL as the new enable condition for register q1. Similar sequential analysis will identify SEL as the new enable condition for register q0 and essentially translate to the RTL code snippet shown in the bottom left. When this modified RTL is taken through a low-power synthesis tool, it will insert appropriate clock gating logic for registers q0 and q1.
Sequential clock gating provides significant power savings because it not only switches off the clock going to registers but also the datapath logic in the fanout of gated registers. However, it requires some changes in the way clock gating optimizations have been traditionally handled in the design flow. In the subsequent sections, we will highlight some aspects of the design flow that require changes. RTL designers will benefit from knowing these upfront.
Identifying Sequential Clock Gating Opportunities
Combinational clock gating has been part of RTL synthesis tools for several years and has become dependable for optimizing for power. Very rarely do synthesis tools miss a combinational clock gating opportunity. Yet, in certain cases the self-assignment to a register uses complex logic (either spanning multiple design hierarchies or written as separate functions) that may cause difficulties for RTL synthesis tools in identifying combinational clock gating. In such cases, designers simplify the complex logic and rewrite it as an “if” statement on the register assignment, thereby simplifying the job of RTL synthesis tools. However, identifying sequential clock gating opportunities is beyond the scope of RTL synthesis tools. Power conscious designers try to analyze the registers for redundant accesses (unobservable or stable writes) and look for conditions under which such accesses can be shut off. There is no single known method of achieving this, and designers mostly develop this expertise over time. Even so, the process can get very tedious and error prone without suitable assistance.
Sequential clock-gating tools introduced to the market either aid the manual exploration of sequential clock gating opportunities or do it automatically. Designers pick the desired methodology (manual versus automated) based on their schedule and comfort with letting a tool make the decisions for them. The tool must fulfill some requirements to make the job of an RTL designer easier, regardless of whether it is used in an automated or manual exploration mode:
- Early feedback about possible total power savings through the elimination of redundant writes to registers
- Clock gating expressions and directions about how to put them into the RTL and additionally, provide a way to change the expressions so that users can configure them to suit their requirements (ease of patching, size of expression, etc.)
- Accurate estimates of power savings and area cost for the clock gating expressions
- If the tool misses creating some clock gating expressions, then feedback must be provided to the user about the potential power savings and the changes (RTL, tool constraints, etc.) that the user can make to help realize those expressions
Such a flow will inform the designer very early on whether a significant potential for power optimization exists in the design. If not, the designer can focus solely on objectives like area and performance. The level of scope will also indicate how much effort should be put towards power optimization. Since the designer understands the power impact of each potential change (putting in a clock gating expression, an RTL or constraints change, the registers with most power saving potential, etc.), the designer can prioritize changes based on delivery schedule. On completion of the sequential clock gating tool’s run, the designer will know how much of the total potential power savings has been realized for the design.
Impact of Sequential Clock Gating on Power
Estimating the power saving potential of a combinational clock gating expression is relatively easy. The change in power of a clock-gated register is computed after reducing the switching activity of the clock net driving the register. The signal that disables the clock net is present in the design, and it is easy to compute its impact on the switching of the clock net by using combinational switching activity propagation techniques. However, sequential clock gating alters the sequential nature of the design. For example, to perform observability-based clock gating for the circuit shown in Figure 2, signal vld_1 will be used to gate the clock going to register d_2. This will not only change the clock toggles but also the toggle activity at the output of register d_2, which in turn will impact switching activity of the entire fanout of register d_2. Figure 5 shows the region (shaded) where switching activity would change as a result of sequentially clock gating register d_2. If only the impact on register power is reported, then a potentially good clock gating expression (since it could have saved a large amount of datapath power) might get rejected.
Hence, to compute accurate power, switching activities for all affected signals and new signals need to be computed. The tool must take the actual power saving impact of a sequential clock gating change into account while making optimization decisions or when presenting the opportunity to the user. An automated tool has to evaluate hundreds and thousands of such gating expressions, so the power computation must be very quick, and at the same time the accuracy of computation should not be compromised.
Verification Impact
The act of combinational clock gating does not change the functionality of the registers in the design, and hence, traditional logical equivalence checkers (LEC) can be used to verify the correctness of such clock gating transformations. Sequential clock gating, on the other hand, changes the sequential nature of the design. For example, if register d_2 in Figure 2 is gated using signal vld_1, then its output is no longer identical to the original design. However, the design output remains unchanged. Traditional LEC tools cannot verify the equivalence of such changes since they require that the register behavior of the two designs being verified is identical. Sequential logic equivalence checkers (SLEC) on the market can help verify the correctness of sequential changes made to the design (like pipelining, retiming, rescheduling, clock gating, etc.). Such tools can be deployed to verify the correctness of sequential clock gating changes. Users must be aware of the following requirements for verifying the correctness of clock gating changes:
- The optimization tool itself cannot make claims about the correctness of the changes it makes
- The verification tool should be able to independently verify the correctness of sequential changes made to the design (either by automation or manually)
- Verification should be fast and formal
- Simulation vector-based verification cannot guarantee the correctness of sequential changes since the bug could be several cycles deep and could remain undetected
Design Flow Support
Combinational clock gating is already present as an internal optimization step in all RTL synthesis tools. RTL synthesis tools themselves support various design flow requirements: bottom-up, engineering change orders (ECO), etc. However, sequential clock gating is still an intermediate stage in the design flow. Either the designer is doing it manually or an automated tool is making such changes to the RTL. For an automated tool, the requirements for writing changed RTL include: complete support for all HDLs (SystemVerilog, Verilog, VHDL, etc.) and a mix thereof;minimal changes to the RTL (preserve all user comments, formatting, and pre-processor directives like macros, defines, includes, etc.); comprehensive configurability of the written-out RTL so as to support Lint rule requirements specific to individual design houses; and clock gating changes written in such a way that they are understood and honored by RTL synthesis tools.
If the sequential clock gating tool is sitting in an automated design flow, then it needs to provide the complete support for ECOs. An ECO change in the RTL could potentially make some of the optimizations done by the automated tool invalid. The requirements for an automated tool for ECO support are:
- Only drop those optimizations which are invalid, and no more
- The output RTL should be identical to the RTL output from the previous run (other than the changes required to make some optimizations invalid)
- Clear directives on how the invalidation of some optimizations would be achieved in the gate-level netlist
- Work in conjunction with RTL synthesis tools to achieve the desired ECO at the gate level
Several other considerations are important. For example, optimizations must be timing aware, and there must be a way to factor in feedback from the specific RTL synthesis tool used.
Optimizations should not introduce any clock domain crossing violations. Finally, automation is necessary to cover various aspects of the optimizations such as discovering the clock-gating expressions, evaluating their power savings and automatically inserting them into the RTL, as well as handling other stages in the design flow. The latter include verification, clock-domain checks, and integration of the sequential clock-gating tool with mainstream RTL synthesis tools.
Conclusion
Sequential clock gating provides significant power savings because it not only switches off the clock going to registers but also the datapath logic in the fanout of gated registers. However, RTL synthesis tools are not capable of identifying sequential clock gating opportunities. Fortunately, recently introduced sequential clock-gating tools either automate this process or facilitate manual exploration of sequential clock gating opportunities, helping RTL designers achieve lower power designs. Sequential clock-gating brings unique challenges of its own for power estimation, verification, and design flow support, all of which an automatic tool must solve to provide the best power savings in the shortest time.