palaniappan-r's picture
Add markdown sources
d59bb1b verified

flipflop + inverter + rebuffer + output buffer

Tool: Clock Tree Synthesis

Subcategory: Unexpected optimization

Conversation

oharboe

Is the "rebuffer" buffer superfluous here?

The "rebuffer" buffer does not exist in placement, only after CTS.

make DESIGN_CONFIG=designs/asap7/mock-array/Element/config.mk gui_final:

image

make DESIGN_CONFIG=designs/asap7/mock-array/Element/config.mk gui_place:

image

maliberty

It is there for timing optimization as the path has negative slack which is addressed post-CTS. Is there a reason to think it superfluous?

oharboe

No extra buffers after removing set_load -pin_load 10 [all_outputs]:

Untar cts-nobuffer.tar.gz

$ ./run-me-mock-array_Element-asap7-base.sh
OpenROAD v2.0-12488-g85f541bb6
openroad> report_checks -to io_lsbOuts_7
Startpoint: _755_ (rising edge-triggered flip-flop clocked by clock)
Endpoint: io_lsbOuts_7 (output port clocked by clock_vir)
Path Group: clock_vir
Path Type: max

  Delay    Time   Description
---------------------------------------------------------
   0.00    0.00   clock clock (rise edge)
  74.98   74.98   clock network delay (propagated)
   0.00   74.98 ^ _755_/CLK (DFFHQNx2_ASAP7_75t_R)
  45.97  120.95 v _755_/QN (DFFHQNx2_ASAP7_75t_R)
  13.06  134.01 ^ _435_/Y (INVx4_ASAP7_75t_R)
  18.04  152.05 ^ output265/Y (BUFx2_ASAP7_75t_R)
   0.01  152.06 ^ io_lsbOuts_7 (out)
         152.06   data arrival time

 300.00  300.00   clock clock_vir (rise edge)
  70.00  370.00   clock network delay (ideal)
 -20.00  350.00   clock uncertainty
   0.00  350.00   clock reconvergence pessimism
-240.00  110.00   output external delay
         110.00   data required time
---------------------------------------------------------
         110.00   data required time
        -152.06   data arrival time
---------------------------------------------------------
         -42.06   slack (VIOLATED)

maliberty

How do I stop "don't touch" from removing the output buffer so I get the comparison I'm after?

Set don't touch after the output buffer is created.

If I delete set_load -pin_load 10 [all_outputs], the extra buffer goes away.

Probably the timing improves and the net no longer needs repair. Similarly with the set_driving_cell changing the timing.