better batch size

plonky2x/examples/mapreduce.rs

@@ -1,6 +1,7 @@
 use ethers::types::U64;
 use itertools::Itertools;
 use jemallocator::Jemalloc;
+use log::debug;
 use plonky2::plonk::config::{AlgebraicHasher, GenericConfig};
 use plonky2x::backend::circuit::{Circuit, PlonkParameters};
 use plonky2x::backend::function::VerifiableFunction;
@@ -17,7 +18,7 @@ static GLOBAL: Jemalloc = Jemalloc;
 const BLOCK_ROOT: &str = "0x4f1dd351f11a8350212b534b3fca619a2a95ad8d9c16129201be4a6d73698adb";
 
 /// The number of balances to fetch.
-const NB_BALANCES: usize = 8192;
+const NB_BALANCES: usize = 131072;
 
 /// The batch size for fetching balances and computing the local balance roots.
 const BATCH_SIZE: usize = 2048;
@@ -61,6 +62,7 @@ impl Circuit for MapReduceCircuit {
                     while leafs.len() != 1 {
                         let mut tmp = Vec::new();
                         for i in 0..leafs.len() / 2 {
+                            debug!("calling sha256 pair w/ curta");
                             tmp.push(builder.curta_sha256_pair(leafs[i*2], leafs[i*2+1]));
                         }
                         leafs = tmp;

plonky2x/src/frontend/builder/mod.rs

@@ -10,6 +10,7 @@ use backtrace::Backtrace;
 use ethers::providers::{Http, Middleware, Provider};
 use ethers::types::U256;
 use itertools::Itertools;
+use log::debug;
 use plonky2::iop::generator::SimpleGenerator;
 use plonky2::iop::target::{BoolTarget, Target};
 use plonky2::plonk::circuit_builder::CircuitBuilder as CircuitAPI;
@@ -114,6 +115,7 @@ impl<L: PlonkParameters<D>, const D: usize> CircuitBuilder<L, D> {
     /// Build the circuit.
     pub fn build(mut self) -> CircuitBuild<L, D> {
         if !self.sha256_requests.is_empty() {
+            debug!("curta constraining sha256");
             self.curta_constrain_sha256();
         }
 

plonky2x/src/frontend/hash/sha/sha256_curta.rs

@@ -2,6 +2,7 @@ use curta::chip::hash::sha::sha256::builder_gadget::{SHA256Builder, SHA256Builde
 use curta::chip::hash::sha::sha256::generator::SHA256HintGenerator;
 use curta::math::field::Field;
 use itertools::Itertools;
+use log::debug;
 use plonky2::iop::target::Target;
 
 use crate::backend::circuit::PlonkParameters;
@@ -207,6 +208,7 @@ impl<L: PlonkParameters<D>, const D: usize> CircuitBuilder<L, D> {
         let mut nb_chunks = 0;
         let mut curr_rq = 0;
         let mut num_rqs = self.sha256_requests.len();
+        debug!("num_rqs: {}", num_rqs);
 
         let zero = self.constant::<ByteVariable>(0u8);
         let zero_chunk = [zero; 1];
@@ -253,6 +255,7 @@ impl<L: PlonkParameters<D>, const D: usize> CircuitBuilder<L, D> {
             (0..nb_chunks / 1024)
                 .map(|_| self.api.init_sha256())
                 .collect_vec();
+        debug!("allocated {} curta sha256 gadgets", gadgets.len());
 
         let mut rq_idx = 0;
         for i in 0..gadgets.len() {

succinct.json

@@ -1,5 +1,5 @@
 {
     "preset": "plonky2",
-    "build_command": "rm -rf ./build && mkdir build && rustup override set nightly && RUST_LOG=debug cargo run --release --example mapreduce build && cp ./target/release/examples/mapreduce ./build/mapreduce",
+    "build_command": "rm -rf ./build && mkdir build && rustup override set nightly && RUSTFLAGS=-Ctarget-cpu=native RUST_LOG=debug cargo run --release --example mapreduce build && cp ./target/release/examples/mapreduce ./build/mapreduce",
     "prove_command": "./build/mapreduce prove --input-json input.json"
 }