PD: Cannot find RMUX cells during synthesis
steveri opened this issue · 12 comments
TSMC CI is currently failing in mem-tile synthesis, with the following error
https://buildkite.com/tapeout-aha/mflowgen/builds/5733
set rmux_cells [get_cells -hier RMUX_T*sel_inst0]
Warning : Could not find requested search value. [SDC-208] [get_cells]
: The 'get_cells' command cannot find any cells named 'RMUX_T*sel_inst0'
set_dont_touch $rmux_cells true
Error : A required object parameter could not be found. [TUI-61] [parse_options]
: An object of type 'subdesign|instance|net|design|libcell' named '' could not be found.
I know at least one way to fix this, and will submit a pull with something like (in common.tcl)
if { "$rmux_cells" != "" } {
set rmux_cells [get_cells -hier RMUX_T*sel_inst0]
}
If someone reading this knows more about RMUX cells and why/whether there is a better fix, please let us know!
Don't do that... That's only kicking the problem down the road. We need to understand what changed because this constraint is important for breaking combinational loops.
I have a potential fix that might satisfy your concerns...I believe this snippet of code works for both old and new RTL.
Instead of...
set rmux_cells [get_cells -hier RMUX_T*sel_inst0]
...we do this...
set rmux_cells [get_cells -hier RMUX_T*sel_*]
In the old RTL, this matches (only) old select instances RMUX_T*sel_inst0 and in the new RTL it matches (only) new select instances RMUX_T*sel_value_0
Ok, that sounds good. Can you please make that change in all of the following places?
-PE synth constraints
-Mem synth constraints
-common/genlibdb-constraints
-common/synopsys-ptpx-gl/loop_break_Interconnect.tcl
Please see new branch missing-rmux-fix, it is running regressions now, after which I will file a pull.
And yes, I came up with the same four files to change, so maybe we got it right :)
I don't think this worked :(
PE tile genlib step now takes 1.5 hours, whereas it used to take about 6 minutes.
(Ref. r7arm-aha:/build/pe5741/17-Tile_PE/25-cadence-genus-genlib)
The resulting lib file is almost 800 MB whereas it used to be less than 80 MB.
The new lib has 240K lines of the form
/* Start of combinational arc of pin SB_T4_WEST_SB_OUT_B16[14] wrt pin SB_T4_WEST_SB_IN_B16[0] */
/* Start of combinational arc of pin SB_T4_WEST_SB_OUT_B16[14] wrt pin SB_T4_WEST_SB_IN_B16[0] */
/* Start of combinational arc of pin SB_T4_WEST_SB_OUT_B16[14] wrt pin SB_T4_WEST_SB_IN_B16[0] */
...
Whreeas the old lib file had none of these.
I'm thinking we have to rejigger the following code in genlibdb-constraints.tcl
# These constraints ... [prevent] all downstream tools
# from analyzing combinational loops that can be realized
# in the interconnect.
set rmuxes [get_cells -hier *RMUX_*_sel_*]
set rmux_outputs [get_pins -of_objects $rmuxes -filter "direction==out"]
set_case_analysis 1 $rmux_outputs
If anyone already know the right thing to do, please speak up...meanwhile/otherwise I think I can track it down fairly quickly...
Yeah it didn't work if those paths are present. The RTL change must have been more than a simple name change, or our constraints somehow didn't cover it properly.
Yeah it didn't work if those paths are present. The RTL change must have been more than a simple name change, or our constraints somehow didn't cover it properly.
I think maybe it really was a simple name change, but when I broadened the search pattern it brought in more cells than before. I.e. instead of get_cells -hier *RMUX_*_sel_* I should have done something more specific like get_cells -hier -regexp .*RMUX_.*_sel_(inst0|value) ... I am testing this now ...
@steveri do you have this fix ready to go?
Yes, and it seems to be working quite well. Have a look if you like.
#904
I went with the more conservative pattern, it matches both old and new rmux names.
I added this fix to the sparsity-mem branch and started a full-chip run, it seems to have finished the tile-array build in record time and is maybe on track to be our fastest TSMC full-chip build.
https://buildkite.com/tapeout-aha/fullchip/builds/422
The fix seems to have worked, I am closing the issue now.