Multiple Clocks and Timing Exceptions

Transcription

1 10 Multiple Clocks and Timing Exceptions Learning Objectives This lab is intended to give you a better understanding of how static timing analysis works and how timing exceptions are properly applied. After completing this lab, you should be able to: Fully constrain and analyze a design using a minimum number of timing exceptions (set_max_delay, set_multicycle_path, set_false_path) Lab Duration: 60 minutes Multiple Clocks and Timing Exceptions Lab 10-1 Synopsys S38

2 Background The design you will be working with, called test, contains parallel paths shown in the figure below (resets not shown). Between inputs adr_i and coeff, and output dout there is a purely combinatorial path as well as a sequential path. The combinatorial and sequential paths actually have different constraints, and meet all setup and hold timing constraints. When you perform timing analysis, however, you will discover that there are initially many violations all due to incomplete or incorrect constraints. This is not a trivial matter to correct. adr_i[15:0] coeff[15:0] Combinatorial Logic dout[31:0] sel_combo Design test mul_result_reg This lab will step you through all the issues associated with constraining parallel sequential and combinatorial paths. You will also be shown an example of what does not work and why (set_max_delay). Task 1. Read Mapped Design 1. Make sure your current working directory is risc_design and invoke dc_shell-t. 2. Read the previously compiled design test into memory. read_db mapped/test.db current_design test link 3. Check the design s timing constraints. report_timing Question 1. Is there anything unusual? Lab 10-2 Multiple Clocks and Timing Exceptions

3 Task 2. Constrain the Design for Timing There are no constraints set on the design; you need to apply a constraint script. The timing specifications for the design are as follows: Clock period: 200 MHz Input delays on data inputs (adr_i[*] and coeff[*]): 2 ns relative to the external system clock Input delay on sel_combo: 0.35 ns relative to external system clock Output delay: 4 ns relative to the external system clock 1. A constraint script with the above specifications has been created for your use. Apply the constraint file lab10_constraints.tcl and check the constraints. source echo verbose lab10_constraints.tcl report_clock check_timing report_timing Question 2. How large is the worst negative slack? Question 3. What path exhibits this violation? Multiple Clocks and Timing Exceptions Lab 10-3

4 Task 3. Using a Max-Delay Constraint The violation you just discovered is very large. There are two possibilities, either the violations are real or the constraints specified were incorrect or insufficient. From the timing report analyze the constraints placed on critical path in more detail. Question 4. When does the data arrive at the input port? Question 5. When is the data expected at the output port? Question 6. What is the maximum path delay constraint for this critical path? It is pretty clear that this path is over constrained - there is no way it can meet a 1 ns maximum delay!! The designer of this block did not give you the proper information, go back to the designer and explain the situation: You find out that the maximum propagation delay through the combinational path should be constrained to 10 ns. This means that the constraints have to be adjusted. During lecture you learned that set_multicycle_path and set_false_path, are considered to be exceptions. If you were to read through the Synopsys documentation you would find two other exception commands, set_max_delay and set_min_delay. Look at the man-pages for the set_max_delay command for further details. Read the command description; it looks like you found a useful command to constrain your combinatorial path! Try it out (You will in fact discover in Task 4 that this command exhibits unexpected behavior, you will need to find a better solution). Lab 10-4 Multiple Clocks and Timing Exceptions

5 1. Apply the appropriate max delay constraint on the combinational path. set_max_delay 10 from [get_ports "coeff* adr_i*"] \ -to [all_outputs] Since set_max_delay is an exception, you should always make sure that the exception was actually applied correctly and that no exceptions were ignored. check_timing report_timing_requirements ignored You will notice a large number of exceptions that are being reported as ignored (non-existent paths). If you look closely, you will see that Design Compiler applied exceptions to paths that do not exist: e.g. Bit 1 of an input port to bits 2 and higher on the output port. Input bit 1 only connects to output bit 1. The command applies the exception to all the permutations of the first argument with the second argument, whether the paths exist or not. If the buses were wider, thereby creating many more ignored exceptions, this would start to impact optimization run time. You would need to write the exceptions differently so that they contain only the actual paths. This is a small design and run time is not a concern here; you can ignore the non-existent paths. 2. Generate another timing report. report_timing It looks like this has solved the combinatorial delay problem. The critical path is no longer the combinatorial path Question 7. Which path is the worst violator now? (Hint: Register reference names always end with _reg[bit_#] ) Multiple Clocks and Timing Exceptions Lab 10-5

6 Task 4. Constrain Multicycle Path Once again, it seems that the designer did not give you the whole story. The path to the mul_result_reg is actually a multicycle path which is allowed to take up to 3 clock cycles (instead of 1, the default). It is time for another exception. 1. Apply a multicycle path of 3 cycles to all paths that lead to mul_result_reg. set_multicycle_path 3 setup to mul_result_reg* Verify that the exceptions were applied correctly: check_timing report_timing_requirements ignored Ignoring the non-existent combinatorial paths again, the multicycle path has been accepted. 2. Add the above multicycle exception to your constraint file lab10_constraints.tcl. 3. Check the timing: report_timing Question 8. What path is violating now? What just happened? You are again looking at the combinatorial input-output path that you were hoping to address earlier. Because the register-register path had a larger violation, it replaced the combinatorial path with a worst negative slack of After applying the multicycle constraint for the register-register path, you are back to the combinatorial path with a WNS of From the timing report, analyze the constraints placed on the critical path in more detail: Lab 10-6 Multiple Clocks and Timing Exceptions

7 Question 9. Notice at the top the Input external delay of 2ns applied to the input coeff where does this come from? Question 10. Notice in the Data Required Time section the Output external delay of 4ns where does this come from? Question 11. What is the maximum path delay constraint for this combinatorial path? Design Compiler considers the input and output delays that were previously specified on the ports as part of the max-delay constraint; you can see that in the timing report. The max-delay constraint number (10ns) for the internal logic was reduced by the amount of external delay at the inputs (2ns) and outputs (4ns), constraining the combinatorial path to =4ns maximum delay. This is NOT what you wanted! What do you learn from this? You do not want to mix input/output delay constraints (relative to a clock) with asynchronous max-delay constraints for the same start- or endpoints! The set_max_delay and set_min_delay commands should be used to constrain only non-clocked paths, e.g. asynchronous reset paths. Conclusion: Since max-delay and input/output delays are not compatible, you have to come up with other solutions. Task 5. Use Virtual Clocks A max_delay constraint did not solve the problem. You could instead consider using a multicycle path for the combinatorial path. You will still need to handle the issue of having identical input and output delays for the combinatorial and sequential paths. Try a completely different approach: virtual clocks. Multiple Clocks and Timing Exceptions Lab 10-7

8 1. Remove the max_delay constraint you applied earlier. reset_path from [get_ports "coeff* adr_i*"] \ -to [all_outputs] report_timing_requirements ignored ;# should be empty 2. Apply a constraint to the combinatorial logic that is independent of the constraints that apply to the sequential logic paths. To accomplish this, constrain the combinatorial paths with respect to a different clock than the sequential paths; this way multiple constraints can be overlayed without interfering with each other. The following schematic illustrates this concept: 0 ns v 0 ns 2 ns Combinatorial Logic v 4 ns mul_result_reg sel_combo The diagram shows that two clocks, and v are now clocking the design. You will constrain the sequential path using, and the combinatorial path using the virtual clock v. The combinatorial path must have a maximum delay of 10 ns so v will have a period of 10 ns and the input and output delays will be zero. 3. Execute the following commands in DC: create_clock name v period 10 set_input_delay 0 clock v add_delay \ [get_ports "coeff* adr_i*"] set_output_delay 0 clock v add_delay [all_outputs] 4. Add the above commands to your constraint script lab10_constraints.tcl. (Hint: Make sure the input and output delay commands are placed AFTER those already in the script. Ordering matters when using the -add_delay option): Lab 10-8 Multiple Clocks and Timing Exceptions

9 5. Make sure that the constraints are complete and analyze for timing: check_timing report_timing You will now see two timing path reports: One for the worst violator in the path group and one for the worst violator of the v path group. Question 12. What is the WNS now? Concentrate on the second timing report first, (the report for the path group v.) It shows that the launch clock is and the capture clock is v. This is not what you wanted! The idea was to have two independent constraint paths: from v through the combinatorial logic to v, and from through the sequential logic to. 6. To solve the above problem you need to mask out the paths from to v and from v to. Execute the following in DC-shell. Do NOT add these commands to your script yet. set_false_path from to v set_false_path from v to 7. Generate a report_timing. That takes care of the path from to v. Notice that the original combinatorial constraint including the input and output delays are still shown as the critical constraint. The constraint with regard to is still constraining the combinatorial path. Question 13. What would happen if you masked out that path using set_false_path from to? Multiple Clocks and Timing Exceptions Lab 10-9

10 Task 6. Refine Virtual Clock Definitions The constraint problem is nearly solved. You still need to completely isolate the sequential from the combinatorial paths. After specifying the required input and output delays you will mask out the non-relevant paths. The following schematic illustrates this point. There are 3 clocks overall. Two virtual clocks will be used to constrain the input and output ports for the combinatorial and the sequential I/O paths. The third (non-virtual) clock will be used for the internal register-register paths. 0 ns v_com 0 ns 2 ns Combinatorial Logic v_com 4 ns v_seq v_seq mul_result_reg sel_combo 1. Indicate in the following table what clock combinations need to be marked as a false path: Launch Clock Capture Clock True / False v_com v_com v_com v_seq v_seq v_seq v_com v_seq v_com v_seq v_com v_seq If you are eager to see a solution, follow the directions under (a). If you want to complete your own script and add all the needed commands, follow (b). a. Reset the design and apply the solution script: reset_design source e v lab10_solution_a.tcl Lab Multiple Clocks and Timing Exceptions

11 b. Open your constraints file lab10_constraints.tcl and add the missing commands. Define the two virtual clocks v_com and v_seq. Constrain the inputs and outputs with regard to the correct clock. Apply the set_false_path exceptions according to the table from above. Source your script when the edits are completed. 2. Generate a timing report. report_timing Question 14. Do all paths meet setup timing now? 3. Check for all violations. report_constraint -all Question 15. Where did these hold violations suddenly come from? Task 7. Back to the Multicycle Path The violations shown in the timing report are all rooted in the multicycle path you applied to the design earlier. By setting a multicycle path for setup at 3 cycles, the hold check capture edge was moved along with the setup check edge. 1. Perform timing analysis for hold. report_timing delay min Question 16. What capture edge is the hold check performed at for? Multiple Clocks and Timing Exceptions Lab 10-11

12 2. Move the hold capture edge back by 2 clock cycles to time 0 and generate reports again. set_multicycle_path 2 hold to mul_result_reg* report_constraint all report_timing delay min Question 17. Are all violations gone now? 3. Quit the dc_shell-t program. Lab Multiple Clocks and Timing Exceptions

13 Answers / Solutions Question 1. Is there anything unusual? The design is not constrained for timing. The timing report only shows the longest path, but does not show any required arrival time information. Reading a db does not necessarily mean that you have a constrained design. Question 2. How large is the worst negative slack? ns. Question 3. What path exhibits this violation? The path with the largest violation is a purely combinatorial path from the input ports to the output ports. The path passes through a Designware multiplier and a mux. Question 4. When does the data arrive at the input port? The data arrives at 2ns due to the external input delay. Question 5. When is the data expected at the output? The output is captured at the 5ns clock edge, and is expected to reach the output port 4ns before the capture edge, or at 1ns (due to the external output delay). Question 6. What is the maximum path delay constraint for this path? -1 ns! (output expected time of 1ns input arrival time of 2ns) Question 7. Which path is the worst violator now? The worst violator now is an internal register-register path ending in mul_result_reg*. The worst negative slack is 5.65 ns. Question 8. What path is violating now? The combinatorial path! It is violating by 5.35ns. Question 9. Notice at the top the Input external delay of 2ns applied to the input coeff where does this come from? The input delay comes from the original set_input_delay constraint, which was applied to coeff and adr_i. Multiple Clocks and Timing Exceptions Lab 10-13

14 Question 10. Notice in the Data Required Time section the Output external delay of 4ns where does this come from? The delay comes from the original set_output_delay constraint, which was applied to dout. Question 11. What is the maximum path delay constraint for this combinatorial path? 4ns! (10ns 2ns 4ns) Question 12. What is the WNS now? ns for the path group. The worst negative slack for the v path group is Question 13. What would happen if you masked out that path using set_false_path from to? This is not a good idea. It would also cause all the internal registerregister timing paths to be ignored. Refine Virtual Clock Definitions Solution for step 1: Launch Clock Capture Clock True / False T v_com F v_seq T v_com F v_com v_com T v_com v_seq F v_seq T v_seq v_com F v_seq v_seq F Lab Multiple Clocks and Timing Exceptions

15 Solution for step 2a: set _period 5 set combo_delay 10 # Real existing clock create_clock -name -period $_period \ [get_ports ] # Define virtual dummy clock for regular in/out Constraining create_clock -name v_seq -period $_period # Define virtual dummy clock for combo-feedthru constraining create_clock -name v_com -period $combo_delay # Constraining for Combo-Feedthru I/O-Delay set_input_delay 0 -clock v_com \ [get_ports "adr_i* coeff*"] set_output_delay 0 -clock v_com [get_ports dout* # Constraining for Combo-Feedthru I/O-Delay set_input_delay 0 -clock v_com \ [get_ports "adr_i* coeff*"] set_output_delay 0 -clock v_com [get_ports dout*] # These are the remaining I/O delays relevant for the paths that # start at an input port and end at internal registers and paths # that start at internal registers and end at output ports! set_input_delay 2 -clock v_seq -add_delay \ [remove_from_collection [all_inputs] [get_ports ]] set_output_delay 4 -clock v_seq -add_delay \ [all_outputs] # Constrain the reg-reg multicycle path set_multicycle_path 3 -setup -to mul_result_reg* # Mask out the non-relevant paths set_false_path -from -to v_com set_false_path -from v_com -to set_false_path -from v_com -to v_seq set_false_path -from v_seq -to v_com set_false_path -from v_seq -to v_seq Multiple Clocks and Timing Exceptions Lab 10-15

16 Question 14. Do all paths meet setup timing now? Yes. All setup constraints are met. Question 15. Where do these hold violations suddenly come from? When the multicycle exception was applied to move the setup capture edge to the third clock edge, the hold capture-edge moved as well. See your student guide for details. Question 16. What capture edge is the hold check performed at for? 10 ns. Question 17. Are all violations gone now? Yes. Finally done. Lab Multiple Clocks and Timing Exceptions