Parallelize s-column generation #326
Conversation
Al-Kindi-0
force-pushed
the
al-parallelize-s-col-generation
branch
from
6215f33 to
6ddec5f
Compare
prover/src/logup_gkr/mod.rs
Outdated
| // note that `deltas[0]` is set `0` and thus `deltas` satisfies the conditions for invoking | ||
| // the function | ||
| let mut cumulative_sum = deltas; | ||
| cumulative_sum[0] = E::ZERO; |
There was a problem hiding this comment.
nit: can we set deltas[0] - E::ZERO on line 391 so that it's clear before we get to the cumulative_sum part that all elements on deltas were properly initialized?
There was a problem hiding this comment.
Makes sense, updated
prover/src/logup_gkr/mod.rs
Outdated
| batch_iter_mut!(&mut deltas[1..], 1024, |batch: &mut [E], batch_offset: usize| { | ||
| let mut query = vec![E::BaseField::ZERO; num_oracles]; | ||
| let mut main_frame = EvaluationFrame::<E::BaseField>::new(num_cols); | ||
|
|
||
| evaluator.build_query(&main_frame, &mut query); | ||
| let cur_value = last_value - mean + gkr_data.compute_batched_query(&query) * *item; | ||
| for (i, v) in batch.iter_mut().enumerate() { | ||
| trace.read_main_frame(i + batch_offset, &mut main_frame); | ||
|
|
||
| result.push(cur_value); | ||
| last_value = cur_value; | ||
| } | ||
|
|
||
| result | ||
| evaluator.build_query(&main_frame, &mut query); | ||
| *v = gkr_data.compute_batched_query(&query) * lagrange_kernel_col[i + batch_offset] | ||
| - mean; | ||
| } | ||
| }); |
There was a problem hiding this comment.
Did you benchmark the serial version of this approach vs the old? In the case where this version turns out to be slower, we could keep the old version for the serial case, and this one for the parallel case.
There was a problem hiding this comment.
Good point! Updated to include the old implementation in the serial case as it a bit faster than the other one and also since I updated the implementation for the parallel case to take the number of threads into account
Al-Kindi-0
force-pushed
the
al-parallelize-s-col-generation
branch
from
6ddec5f to
74b5a5c
Compare
Al-Kindi-0
force-pushed
the
al-parallelize-s-col-generation
branch
from
74b5a5c to
b3eb72a
Compare
This brings about 60% improvement in s-column generation times.
Fixes also a bug in the construction of the MLEs.