search

The-OpenROAD-Project / OpenROAD

OpenROAD's unified application implementing an RTL-to-GDS Flow. Documentation at https://openroad.readthedocs.io/en/latest/
https://theopenroadproject.org/
BSD 3-Clause "New" or "Revised" License
1.5k stars 527 forks source link

rtlmp should complete in a reasonable amount of time or fail quickly if it is hopeless #4335

Closed oharboe closed 9 months ago

oharboe commented 9 months ago

Description

Never completes...

To reproduce(I suppose it should be possible to make a small text only based unit-test that reproduces the problem, but I don't know how to):

unzip https://drive.google.com/file/d/1e2ebRJ80NInmfF3xhgY96wOHQsz_7Czc/view?usp=sharing

./run-me-BoomTile-asap7-base.sh 
OpenROAD v2.0-11160-g042cdcfa4 
This program is licensed under the BSD-3 license. See the LICENSE file for details.
Components of this program may be licensed under more restrictive licenses which must be honored.
[INFO ORD-0030] Using 16 thread(s).
HierRTLMP Flow enabled...
Call Macro Placer  -halo_width 20 -report_directory bazel-bin/build//objects/asap7/BoomTile/base/rtlmp
Floorplan Outline: (0.0, 0.0) (4740.732, 4740.732),  Core Outline: (2.052, 2.16) (4738.716, 4738.5)
Traversed logical hierarchy
    Number of std cell instances: 2130872
    Area of std cell instances: 273473.66
    Number of macros: 7
    Area of macros: 4194455.50
    Area of macros with halos: 4534412.00
    Area of std cell instances + Area of macros: 4467929.00
    Core area: 22434450.00
    Design Utilization: 0.20
    Core Utilization: 0.01
    Manufacturing Grid: 1

[INFO MPL-0024] [Multilevel Autoclustering] Creating clustered netlist.
[INFO MPL-0039] [Coarse Shaping] Determining shape functions for clusters.
[INFO MPL-0028] [Hierarchical Macro Placement] Placing clusters and macros.
[no further output]

Suggested Solution

Actionable error message if it is hopeless(what is wrong and what needs to be fixed) or complete in a practical amount of time. At most hours, not days.

Also, it would be great to have a write_macro_placement command so as to store the macro placement. Once the floorplan has been identified, there may not be a reason and also reasons not to recompute it. Also writing out the macro placement allows for manual tinkering afterewards.

Additional Context

No response

AcKoucher commented 9 months ago

@oharboe Write macro placement was already implemented in #4120. It was reported in #4068.

I'm updating Makefile options in ORFS PR #1680 with -write_macro_placement included.

oharboe commented 9 months ago

@oharboe Write macro placement was already implemented in #4120. It was reported in #4068.

I'm updating Makefile options in ORFS PR #1680 with -write_macro_placement included.

Ah yes, I recall now, thanks!

This issue is really about running times of rtlmp.

maliberty commented 9 months ago

How are you running synthesis? I'm surprised to see modules as small as

[DEBUG MPL-multilevel_autoclustering] Setting Cluster Metrics for core/ftq_arb: Num Macros: 0 Num Std Cells: 27
maliberty commented 9 months ago

Are you setting export SYNTH_HIERARCHICAL = 1 export RTLMP_FLOW = True ?

oharboe commented 9 months ago

Are you setting export SYNTH_HIERARCHICAL = 1 export RTLMP_FLOW = True ?

Yes, core/ftq_arb Verilog:

module Arbiter_19(
  input        io_in_0_valid,
  input  [5:0] io_in_0_bits,
  input        io_in_1_valid,
  input  [5:0] io_in_1_bits,
               io_in_2_bits,
  output       io_in_2_ready,
  output [5:0] io_out_bits
);

  assign io_in_2_ready = ~(io_in_0_valid | io_in_1_valid);
  assign io_out_bits = io_in_0_valid ? io_in_0_bits : io_in_1_valid ? io_in_1_bits : io_in_2_bits;
endmodule
maliberty commented 9 months ago

I see things have become inconsistent around MAX_UNGROUP_SIZE 

./designs/asap7/riscv32i/config.mk\013:export MAX_UNGROUP_SIZE ?= 1000
./designs/gf12/ariane/config.mk\05:export MAX_UNGROUP_SIZE ?= 10000
./designs/gf12/swerv_wrapper/config.mk\04:export MAX_UNGROUP_SIZE ?= 10000
./designs/gf12/tinyRocket/config.mk\06:export MAX_UNGROUP_SIZE ?= 1000
./designs/gf12/bp_quad/config.mk\07:export MAX_UNGROUP_SIZE ?= 1000
./designs/gf12/ariane133/config.mk\06:export MAX_UNGROUP_SIZE ?= 10000
./designs/nangate45/tinyRocket/config.mk\06:export MAX_UNGROUP_SIZE ?= 5000
./designs/tsmc65lp/tinyRocket/config.mk\06:export MAX_UNGROUP_SIZE ?= 5000
./designs/sky130hd_fakestack/ariane136/config.mk\06:export MAX_UNGROUP_SIZE = 10000
./platforms/nangate45/config.mk\025:export MAX_UNGROUP_SIZE ?= 10000
./scripts/synth_hier_report.tcl\031:  if { [info exist ::env(MAX_UNGROUP_SIZE)] && $::env(MAX_UNGROUP_SIZE) > 0 } {
./scripts/synth_hier_report.tcl\032:    set ungroup_threshold $::env(MAX_UNGROUP_SIZE)
./Makefile\0185:export MAX_UNGROUP_SIZE ?= 0

So in asap7 (and all non-ng45) it is defaulting to 0 which isn't great as every small module will be kept. This will need some sorting out.

AcKoucher commented 9 months ago

@oharboe FYI Even though #4360 solves the problem here, I think is important to comment that for this specific design we end up exhausting the mechanism that tries to estimate the best position for both macros and std cells in mpl2. As consequence, we use a more extreme mechanism to do so and this can deteriorate QoR.

Within this context, I think #4354 can have a significant impact for us to get better results out of this type of situation.

The same applies for #4206.