Closed XiaPZ closed 4 months ago
Can you give more details on how you "tried" hmmer and how you relate IPC of hmmer agaist spec2006 score?
Sure, hmmer is built with gcc 11, O3 option, with rv64imafdc, score is calculated as ref_time / ((1 / IPC) num_insts / 10^9 ). For hmmer, there ref1 and ref2, so the score is ref_time / ((((1/IPC_ref1) num_insts_ref1) + (1/IPC_ref2) * num_insts_ref2) / 10^9).
checkpoint is collected with NEMU, and run with util/warmup_scripts/simple_gem5.sh
hmmer's performance is sensitive to B-extension, try gcc 12 with rv64gc_zba_zbb_zbc_zbs
After recompiled hmmer with B-extension, now the IPC is around 4.9, you are right, hmmer is sensitive to B-extension.
Thanks for your help @shinezyy
I am trying to run some spec2006 benchmarks on xiangshan GEM5, currently I tried hmmer on backport branch, the weighted IPC is around 2.5 (spec2006 score is around 5.6 / GHz), that is far from 11 / GHz, so I am wondering what is the performance of current backport branch? Thanks