Benchmarking tool using cost estimator #235
Conversation
iquerejeta
force-pushed
the
main
branch
2 times, most recently
from
1edf769 to
25238ea
Compare
There was a problem hiding this comment.
Hey! Thanks for the PR. A couple of remarks:
- I'd just like to have it within a
cost-estimatorfeature flag. - I think is a bit unreasonable to be forced to run the
MSMs to then get the cost. This tool is an estimator, hence it should estimate. Not benchmark directly.
I think a much better way is to log in the CSV properties like:
- Num of Advice cells
- Num of lookup-active columns
- Num of fixed columns
- Maximum rotation
- Num of copy-enabled columns
- Maximum lookup degree
- Maximum expression degree
- Degree of the circuit (lenght)
- Proof Size
Note that most of these things can be counted in Commitments (for both, operations and storage).
Hence, we have a universal unit that will be useful across all possible curves/primitives.
Would like to hear more from @davidnevadoc @han0110 @kilic
halo2_proofs/src/dev/cost_model.rs
Outdated
| CommitmentScheme::KZG => { | ||
| // Polycommit KZG: | ||
| // - quotient polynomial commitment (COMM bytes) | ||
| size(1, 0) |
There was a problem hiding this comment.
For GWC19 it'd be size(set(rotations).len(), 0) and without the 1 from multiopen
There was a problem hiding this comment.
I've done using a HashSet, so rotations are not repeated. Is that right? a36267b
I think the same with these 2 bullets. It would be beyond this porting efford but info of number of gates and gate with their add, mul complexity might be usuful to see since it adds up to evaluation time |
|
|
||
| /// Options to build a circuit specifiction to measure the cost model of. | ||
| #[derive(Debug)] | ||
| pub struct CostOptions { |
There was a problem hiding this comment.
We should include ::shuffle arguments as well
There was a problem hiding this comment.
Do you have a pointer to docs of how the shuffle argument works? Looking at the code it seems that it only uses two rotations (\omega X and \omega), is that right? 67440c9
There was a problem hiding this comment.
I think the feature is defined in the README (not sure if it points to any extra docs). IIRC the cost is to commit to the shuffled column which means extra commitment.
Unsure about the rotations though. ping @kilic
There was a problem hiding this comment.
er to docs of how the shuffle argument works? Looking at the code it seems that it only uses two rotations (
\omega Xand\omega), is that right? 67440c9
@iquerejeta Here it goes #185. It's kind of simplified version of lookup argument
There was a problem hiding this comment.
Great, thanks. And I think the computation of the lookup queries was wrong. Fixed it here
iquerejeta
force-pushed
the
main
branch
2 times, most recently
from
c1184bb to
8464681
Compare
iquerejeta
force-pushed
the
main
branch
2 times, most recently
from
419f77c to
3d789b8
Compare
Co-Authored-By: Iñigo Querejeta Azurmendi <[email protected]>
Use built-in function to compute cs max_degree
) * Benchmarking tool using cost estimator Co-Authored-By: Iñigo Querejeta Azurmendi <[email protected]> * Address review comments * Add example of cost-estimator usage * Cost model example behind feature * Make report generic over field size and commitment scheme * Nits * Use serde for printing the report * Calculate max expression degree Use built-in function to compute cs max_degree * Add cost computation for KZG GWC * Add shuffle argument * Nits * Remove CLI * Remove FromStr impl - not needed without CLI * Fix queries of Lookup * There is only one permutation argument * Remove dependency of halo2_gadgets for cost-estimator example --------- Co-authored-by: Henry Blanchette <[email protected]>
This PR moves the functionality used in
halo2_proofs/examples/cost-model.rsinto a new filehalo2_proofs/src/dev/cost_model.rs, which is exposed publicly. We expose two additional functions,from_circuit_to_model_circuitandfrom_circuit_to_cost_model_optionsto get the benchmarks for a given Plonk circuit.