Aux computation now deduplicates evaluation of Linear Combinations shared across the auxiliary variables. 28% of Linear Combinations are currently reused in jolt_constraints.rs and were previously computed from scratch. 35% speed up over existing on 2^22 trace length 50bps reduction e2e.
In the process removes AuxComputation.{flat_vars, inputs_to_flat}.
Separately speeds up + cleans up CombinedUniformBuilder::compute_spartan_Az_Bz_Cz by moving the batch LC evaluation logic onto LC.
sragss
commented
3 weeks ago
Aux computation now deduplicates evaluation of Linear Combinations shared across the auxiliary variables. 28% of Linear Combinations are currently reused in
jolt_constraints.rsand were previously computed from scratch. 35% speed up over existing on 2^22 trace length 50bps reduction e2e.In the process removes
AuxComputation.{flat_vars, inputs_to_flat}.Separately speeds up + cleans up
CombinedUniformBuilder::compute_spartan_Az_Bz_Czby moving the batch LC evaluation logic onto LC.