Remove large vector allocations #144
Conversation
src/r1cs/mod.rs
Outdated
| let u_1_cdot_CZ_2 = U1.u * CZ_2[i]; | ||
| AZ_1_circ_BZ_2 + AZ_2_circ_BZ_1 - u_1_cdot_CZ_2 - CZ_1[i] |
There was a problem hiding this comment.
Very small nit: subtraction is ever so slightly more expensive than addition. Computing
u_1_cdot_Cz_2_plus_Cz_1 = U1.u * CZ_2[I] + CZ_1[i] and returning
AZ_1_circ_BZ_2 + AZ_2_circ_BZ_1 - u_1_cdot_Cz_2_plus_Cz_1 would replace one subtraction by an addition.
There was a problem hiding this comment.
I think this is a good change. I would also like to see (in a PR, or in separate commits attached to this PR) showing a slightly more extensive variant of that where we delete (instead of adapting, as you do here with multiply_witness_unchecked) the allocating methods that this would make obsolete.
If we're going to go for fine-grained allocation control, let's do it fully and not maintain "shadow" allocating methods that would be left unused.
src/lib.rs
Outdated
| // save the inputs before proceeding to the `i+1`th step | ||
| let r_U_primary_i = self.r_U_primary.clone(); | ||
| let r_U_secondary_i = self.r_U_secondary.clone(); | ||
| // let l_u_primary_i = self.l_u_primary.clone(); |
There was a problem hiding this comment.
Oops, one second
|
@huitseeker The allocating methods are still required by SuperNova, and I wanted to avoid the PR from spilling over. My plan was to wait once this is merged in, then open the same set of PRs for SuperNova, and everything should be cleaned up. |
winston-h-zhang
force-pushed
the
large-vector-allocations
branch
from
cda7c6f
to
74a44a0
Compare
winston-h-zhang
force-pushed
the
large-vector-allocations
branch
3 times, most recently
from
a300f84
to
d599a53
Compare
winston-h-zhang
force-pushed
the
large-vector-allocations
branch
from
d599a53
to
8f7f8e6
Compare
* remove large vector allocations * add suggestions
Addresses #137.
This PR addresses the cloning/allocation of large vectors used in the intermediary steps of each Nova
prove_stepcomputation. In particular, we focus onNIFS::prove, which allocates new vectors to compute the fold step, when technically we could be reusing and overwriting the already allocated vectors from the previous step. Concretely, we add the functionsNIFS::prove_mutandR1CSRelaxedWitness/Instance::prove_mut. We also add aResourceSinkwith preallocated buffers forABC_Z{1,2}andTs to help as stratch memory for thecommit_T_intooperation.