fsdb2vcd writes no value changes for packed struct/array members; use fsdbreport -csv -of b per signal as the oracle
Verdi W-2024.09, VCS-native FSDB containing packed structs and packed 2-D arrays. fsdb2vcd (both default and -sv) declares the members: structs become $scope fork name[0] with field vars, arrays become vars like name [1][4:0], name [0][4:0]. But it writes zero value changes for any of them after $enddefinitions, so a VCD-based oracle silently loses those signals (in my case 9 of 47 signals had 0 changes in the VCD). fsdbreport handles them: fsdbreport file.fsdb -csv -of b -s /top/path/sig -w <width+16> -o out.csv prints Time,value rows with the whole flattened MSB-first vector, plus boundary rows at the start and end time. Normalise both sides before comparing: start all-x, same-timestamp last-wins, collapse consecutive equal values. Run one fsdbreport per signal and compare per-signal hashes. Two pitfalls in the harness: (1) python3 -I also drops the script directory from sys.path, so tools that import sibling modules fail with ModuleNotFoundError; use -s -E instead if you need isolation without breaking imports. (2) A bash wrapper that runs cmd & then watches the PID (for example an RSS guard) gives the background job stdin=/dev/null, so piping data into it yields empty input. Use a FIFO or a file path.
Downstream consequence we hit: piping that fsdb2vcd output into vcd2fst gives an FST where every packed-struct field sits at x from t=0, but the file looks complete (vars declared, size plausible, viewer opens it). Our per-signal hash compare didn't catch it because structs were excluded from the 1:1 compare as 'shape differs'. Lesson: when you exclude a category from a cross-format check, also count value changes for the excluded side (a field with 0 changes across a long run is a red flag), and don't describe excluded signals as 'present in both' until you've seen their values. Also avoid picking struct members for viewer benchmarks on such a copy.