Key generation performance of EC-P hybrids drops significantly when OQS_KEM_ENCODERS is set to ON

[baentsch]

Details discussed in #726

Originally posted by @youer0219 in #726

[RodriM11]

Can we assume that this issue is fixed by #778 ? @youer0219 would you mind confirming that now a better performance is achieved?

[youer0219]

Can we assume that this issue is fixed by #778 ? @youer0219 would you mind confirming that now a better performance is achieved?我们能否认为这个问题已经由 #778 解决了呢？请您确认一下，现在的性能是否有所提升。

@RodriM11 Benchmark results confirm that #778 delivers clear performance improvements for P-curve hybrid KEM key generation when encoders are enabled.

I ran tests across three configurations to quantify the impact:

1. Pre-oqsprov: avoid encoder hotpath in P-curve hybrid KEM keygen #778, OQS_KEM_ENCODERS=ON (commit f193678, parent commit of the oqsprov: avoid encoder hotpath in P-curve hybrid KEM keygen #778 merge)
2. Current main branch, OQS_KEM_ENCODERS=ON (includes changes from oqsprov: avoid encoder hotpath in P-curve hybrid KEM keygen #778)
3. Current main branch, OQS_KEM_ENCODERS=OFF

All tests were executed on the same machine using OpenSSL 3.5.4 with an AMD 9700X:

algorithm	pre-#778 ON	post-#778 ON	post-#778 OFF
p256_mlkem512	263.0	755.8	1579.9
p384_mlkem768	238.1	579.0	981.3
p521_mlkem1024	204.2	441.3	635.7
x25519_mlkem512	54014.8	53191.9	53606.1

For P-curve hybrid algorithms, the optimization yields a substantial speedup. For instance, throughput of p256_mlkem512 rises from ~263 to ~756 keygen/s — approximately a 2.9× improvement. This is consistent with the conclusions outlined in #778 (comment).

Even so, performance under OQS_KEM_ENCODERS=ON still lags noticeably behind the OFF configuration after this patch. This strikes me as somewhat puzzling.