[mohabsafey]

The linear algebra no longer counts the F4 operations of the runs applying the trace (#390 removed the application counters from the kernels as unused, and the counts of a run were never added to the meta data of the caller anyway). As a result, the verbose output of msolve_trace_qq and of the lifting of Gröbner bases always reported 0 Gops/sec for the learning and the application phases, and 0 additions, multiplications and reductions at info level 3.

This PR records the work of an application of the trace in the trace itself, when it is learned. This work is the same for all primes, so it does not need to be counted at each run, and the kernels and their inner loops are not changed.

Changes

• Three new fields of trace_t, summed over the rounds of the trace:
  • nmadd: the multiply-adds of the reductions of the rows by the known pivots. construct_trace adds, for each row kept in the trace, the lengths of the known pivots marked in its reducer bit array. This is exactly the work of these reductions when the trace is applied.
  • nterms: the terms of the rows of the matrices, that is the rows to be reduced (multiples of basis elements, whose lengths construct_trace now reads from the basis passed to it) and the reducers used.
  • nmadd_new: an estimate of the multiply-adds of the reductions of the new rows by each other, which nmadd does not include: the number of new elements of a round times their total length, added by add_lms_to_trace.
• The rate of the application phase in the verbose output is computed from nmadd. It is printed only when a trace is learned, which is not the case with the probabilistic linear algebra (-l 42 or 44). The rate of the learning phase and the numbers of additions, multiplications and reductions printed at info level 3, which were always 0, are no longer printed.

These quantities do not depend on the prime nor on the number of threads, which makes them suitable for cost models choosing between strategies. They are meant for the cost model of the p-adic lifting of parametrizations (#391), which used the application counters removed by #390.

Testing

• make check passes (73/73).
• The outputs are identical to those of master on cyclic5, kat7, kat10, eco10 and random2_7_7, with -P 2 -t 2 and with -g 2.
• On kat7, -P 2 and -g 2 run with each of the options -l 2, 42 and 44, at info levels 0, 2 and 3 (-l 1 crashes on master too, independently of this PR).
• The learning run of F4 on kat12 is not slower (4.1 to 4.3 s against 4.6 to 4.8 s for master, 3 seeds).

Parts of this PR were written with the assistance of Claude Code (Claude Opus 5.5); all changes were reviewed and tested by the author.

🤖 Generated with Claude Code