Rsz degrades setup rather than improve it

[maliberty]

Describe the bug

I tried changing

--- a/flow/designs/asap7/swerv_wrapper/constraint.sdc
+++ b/flow/designs/asap7/swerv_wrapper/constraint.sdc
@@ -2,7 +2,7 @@ current_design swerv_wrapper
 
 set clk_name core_clock 
 set clk_port_name clk
-set clk_period 2500 
+set clk_period 1000

I observe in 4_1_cts.log that repair_timing makes things worse:

[INFO RSZ-0099] Repairing 930 out of 930 (100.00%) violating endpoints...
   Iter   | Removed | Resized | Inserted | Cloned |  Pin  |    WNS   |   TNS      |  Viol  | Worst
          | Buffers |  Gates  | Buffers  |  Gates | Swaps |          |            | Endpts | Endpt
---------------------------------------------------------------------------------------------------
        0 |       0 |       0 |        0 |      0 |     0 | -150.852 |  -102009.4 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
...
     2200 |       0 |       0 |      153 |      0 |    11 | -636.199 |  -636317.9 |    930 | mem/Gen_dccm_enable_dccm/mem_bank_1__dccm_bank/mem/addr_in[5]
    final |       0 |       0 |      153 |      0 |    11 | -636.199 |  -636317.9 |   1126 | mem/Gen_dccm_enable_dccm/mem_bank_1__dccm_bank/mem/addr_in[5]

Expected Behavior

Improve timing (or at least neutral)

Environment

With ORFS dadde560afdc443d2fb26c81ad4b8f628d398dff and OR 42ba30d730851a73bd2996c5673a05fd79520f87

To Reproduce

as above

Relevant log output

No response

Screenshots

No response

Additional Context

No response

[maliberty]

I don't know if it relates but note that for this design

export LIB_MODEL = CCS

[maliberty]

Note at period=1500 it still hits 0 tns and this problem is not observed.

[precisionmoon]

This may benefit from odb based save/restore. I will look into it.

[precisionmoon]

There are multiple issues here. One of them is RSZ journal save/restore. Since this doesn't preserve the exact order of all the transforms, it's hard to do 100% restoration to a previous snapshot. Another issue is the delay estimation during buffering. It fails to predict worse QoR with additional buffering. With SKIP_PIN_SWAP=1, big QoR degradation is mitigated.

repair_timing -verbose -repair_tns 100 -skip_pin_swap
[INFO RSZ-0094] Found 930 endpoints with setup violations.
[INFO RSZ-0099] Repairing 930 out of 930 (100.00%) violating endpoints...
   Iter   | Removed | Resized | Inserted | Cloned |  Pin  |    WNS   |   TNS      |  Viol  | Worst
          | Buffers |  Gates  | Buffers  |  Gates | Swaps |          |            | Endpts | Endpt
---------------------------------------------------------------------------------------------------
        0 |       0 |       0 |        0 |      0 |     0 | -150.852 |  -102009.4 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
       10 |       0 |       0 |       16 |      0 |     0 | -135.524 |   -88766.1 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
       20 |       0 |       0 |       26 |      0 |     0 | -135.524 |   -88766.1 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
       30 |       0 |       0 |       36 |      0 |     0 | -135.576 |   -88809.1 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D

...
     1190 |       0 |       0 |      125 |      0 |     0 | -342.369 |  -319908.9 |    930 | mem/Gen_dccm_enable_dccm/mem_bank_1__dccm_bank/mem/addr_in[5]
     1193 |       0 |       0 |       79 |      0 |     0 | -134.332 |  -100334.9 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
     1200 |       0 |       0 |       85 |      0 |     0 | -134.332 |  -100334.9 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
     1200 |       0 |       0 |       85 |      0 |     0 | -134.332 |  -100334.9 |    930 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
    final |       0 |       0 |       85 |      0 |     0 | -134.332 |  -100334.9 |    989 | swerv/dec/instbuff/bp0ff.genblock_dff.dout\[29\]$_DFF_PN0_/D
---------------------------------------------------------------------------------------------------

A proper fix will need a project.

1. migrate from RSZ journal to odb save/restore
2. improve buffering (delay estimation and possibly sizing)
3. add codes to guarantee equal or better final QoR

[povik]

Since this thread was active we have moved to odb save/restore and fixed a case of repair_timing not restoring to previous best result. @maliberty I think we can close it.