feat: add parallelize_in helper function (#46)

Multi-threading of witness generation is tricky because one has to
ensure the circuit column assignment order stays deterministic. To
ensure good developer experience / avoiding pitfalls, we provide a new
helper function for this.

Co-authored-by: Jonathan Wang <[email protected]>

halo2-base/Cargo.toml

@@ -11,6 +11,7 @@ num-traits = "0.2"
 rand_chacha = "0.3"
 rustc-hash = "1.1"
 ff = "0.12"
+rayon = "1.6.1"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 log = "0.4"

halo2-base/src/gates/builder.rs

@@ -17,6 +17,9 @@ use std::{
     env::{set_var, var},
 };
 
+mod parallelize;
+pub use parallelize::*;
+
 /// Vector of thread advice column break points
 pub type ThreadBreakPoints = Vec<usize>;
 /// Vector of vectors tracking the thread break points across different halo2 phases

halo2-base/src/gates/builder/parallelize.rs

@@ -0,0 +1,38 @@
+use itertools::Itertools;
+use rayon::prelude::*;
+
+use crate::{utils::ScalarField, Context};
+
+use super::GateThreadBuilder;
+
+/// Utility function to parallelize an operation involving [`Context`]s in phase `phase`.
+pub fn parallelize_in<F, T, R, FR>(
+    phase: usize,
+    builder: &mut GateThreadBuilder<F>,
+    input: Vec<T>,
+    f: FR,
+) -> Vec<R>
+where
+    F: ScalarField,
+    T: Send,
+    R: Send,
+    FR: Fn(&mut Context<F>, T) -> R + Send + Sync,
+{
+    let witness_gen_only = builder.witness_gen_only();
+    // to prevent concurrency issues with context id, we generate all the ids first
+    let ctx_ids = input.iter().map(|_| builder.get_new_thread_id()).collect_vec();
+    let (outputs, mut ctxs): (Vec<_>, Vec<_>) = input
+        .into_par_iter()
+        .zip(ctx_ids.into_par_iter())
+        .map(|(input, ctx_id)| {
+            // create new context
+            let mut ctx = Context::new(witness_gen_only, ctx_id);
+            let output = f(&mut ctx, input);
+            (output, ctx)
+        })
+        .unzip();
+    // we collect the new threads to ensure they are a FIXED order, otherwise later `assign_threads_in` will get confused
+    builder.threads[phase].append(&mut ctxs);
+
+    outputs
+}

halo2-ecc/src/ecc/fixed_base.rs

@@ -2,7 +2,7 @@
 use super::{ec_add_unequal, ec_select, ec_select_from_bits, EcPoint, EccChip};
 use crate::fields::{FieldChip, PrimeField, Selectable};
 use group::Curve;
-use halo2_base::gates::builder::GateThreadBuilder;
+use halo2_base::gates::builder::{parallelize_in, GateThreadBuilder};
 use halo2_base::{gates::GateInstructions, utils::CurveAffineExt, AssignedValue, Context};
 use itertools::Itertools;
 use rayon::prelude::*;
@@ -107,6 +107,7 @@ where
     curr_point.unwrap()
 }
 
+/* To reduce total amount of code, just always use msm_par below.
 // basically just adding up individual fixed_base::scalar_multiply except that we do all batched normalization of cached points at once to further save inversion time during witness generation
 // we also use the random accumulator for some extra efficiency (which also works in scalar multiply case but that is TODO)
 pub fn msm<F, FC, C>(
@@ -212,6 +213,7 @@ where
         .collect_vec();
     chip.sum::<C>(ctx, scalar_mults)
 }
+*/
 
 /// # Assumptions
 /// * `points.len() = scalars.len()`
@@ -269,25 +271,23 @@ where
     C::Curve::batch_normalize(&cached_points_jacobian, &mut cached_points_affine);
 
     let field_chip = chip.field_chip();
-    let witness_gen_only = builder.witness_gen_only();
 
     let zero = builder.main(phase).load_zero();
-    let thread_ids = (0..scalars.len()).map(|_| builder.get_new_thread_id()).collect::<Vec<_>>();
-    let (new_threads, scalar_mults): (Vec<_>, Vec<_>) = cached_points_affine
-        .par_chunks(cached_points_affine.len() / points.len())
-        .zip_eq(scalars.into_par_iter())
-        .zip(thread_ids.into_par_iter())
-        .map(|((cached_points, scalar), thread_id)| {
-            let mut thread = Context::new(witness_gen_only, thread_id);
-            let ctx = &mut thread;
-
+    let scalar_mults = parallelize_in(
+        phase,
+        builder,
+        cached_points_affine
+            .chunks(cached_points_affine.len() / points.len())
+            .zip_eq(scalars)
+            .collect(),
+        |ctx, (cached_points, scalar)| {
             let cached_points = cached_points
                 .iter()
                 .map(|point| chip.assign_constant_point(ctx, *point))
                 .collect_vec();
             let cached_point_window_rev = cached_points.chunks(1usize << window_bits).rev();
 
-            debug_assert_eq!(scalar.len(), scalar_len);
+            assert_eq!(scalar.len(), scalar_len);
             let bits = scalar
                 .into_iter()
                 .flat_map(|scalar_chunk| {
@@ -319,9 +319,8 @@ where
                     field_chip.gate().mul_add(ctx, is_started, is_zero_window, not_zero_window)
                 };
             }
-            (thread, curr_point.unwrap())
-        })
-        .unzip();
-    builder.threads[phase].extend(new_threads);
+            curr_point.unwrap()
+        },
+    );
     chip.sum::<C>(builder.main(phase), scalar_mults)
 }

halo2-ecc/src/ecc/mod.rs

@@ -971,6 +971,7 @@ impl<'chip, F: PrimeField, FC: FieldChip<F>> EccChip<'chip, F, FC> {
         self.field_chip.assert_equal(ctx, P.y, Q.y);
     }
 
+    /// None of elements in `points` can be point at infinity.
     pub fn sum<C>(
         &self,
         ctx: &mut Context<F>,
@@ -1153,21 +1154,15 @@ impl<'chip, F: PrimeField, FC: FieldChip<F>> EccChip<'chip, F, FC> {
         #[cfg(feature = "display")]
         println!("computing length {} fixed base msm", points.len());
 
-        // heuristic to decide when to use parallelism
-        if points.len() < 25 {
-            let ctx = builder.main(phase);
-            fixed_base::msm(self, ctx, points, scalars, max_scalar_bits_per_cell, clump_factor)
-        } else {
-            fixed_base::msm_par(
-                self,
-                builder,
-                points,
-                scalars,
-                max_scalar_bits_per_cell,
-                clump_factor,
-                phase,
-            )
-        }
+        fixed_base::msm_par(
+            self,
+            builder,
+            points,
+            scalars,
+            max_scalar_bits_per_cell,
+            clump_factor,
+            phase,
+        )
 
         // Empirically does not seem like pippenger is any better for fixed base msm right now, because of the cost of `select_by_indicator`
         // Cell usage becomes around comparable when `points.len() > 100`, and `clump_factor` should always be 4

halo2-ecc/src/ecc/pippenger.rs

@@ -7,11 +7,13 @@ use crate::{
     fields::{FieldChip, PrimeField, Selectable},
 };
 use halo2_base::{
-    gates::{builder::GateThreadBuilder, GateInstructions},
+    gates::{
+        builder::{parallelize_in, GateThreadBuilder},
+        GateInstructions,
+    },
     utils::CurveAffineExt,
-    AssignedValue, Context,
+    AssignedValue,
 };
-use rayon::prelude::*;
 
 // Reference: https://jbootle.github.io/Misc/pippenger.pdf
 
@@ -238,7 +240,6 @@ where
 
     // get a main thread
     let ctx = builder.main(phase);
-    let witness_gen_only = ctx.witness_gen_only();
     // single-threaded computation:
     for scalar in scalars {
         for (scalar_chunk, bool_chunk) in
@@ -250,32 +251,28 @@ where
             }
         }
     }
-    // see multi-product comments for explanation of below
 
     let c = clump_factor;
     let num_rounds = (points.len() + c - 1) / c;
+    // to avoid adding two points that are equal or negative of each other,
+    // we use a trick from halo2wrong where we load a "sufficiently generic" `C` point as witness
+    // note that while we load a random point, an adversary could load a specifically chosen point, so we must carefully handle edge cases with constraints
+    // we call it "any point" instead of "random point" to emphasize that "any" sufficiently generic point will do
     let any_base = load_random_point::<F, FC, C>(chip, ctx);
     let mut any_points = Vec::with_capacity(num_rounds);
     any_points.push(any_base);
     for _ in 1..num_rounds {
         any_points.push(ec_double(chip, ctx, any_points.last().unwrap()));
     }
-    // we will use a different thread per round
-    // to prevent concurrency issues with context id, we generate all the ids first
-    let thread_ids = (0..num_rounds).map(|_| builder.get_new_thread_id()).collect::<Vec<_>>();
-    // now begins multi-threading
 
+    // now begins multi-threading
     // multi_prods is 2d vector of size `num_rounds` by `scalar_bits`
-    let (new_threads, multi_prods): (Vec<_>, Vec<_>) = points
-        .par_chunks(c)
-        .zip(any_points.par_iter())
-        .zip(thread_ids.into_par_iter())
-        .enumerate()
-        .map(|(round, ((points_clump, any_point), thread_id))| {
+    let multi_prods = parallelize_in(
+        phase,
+        builder,
+        points.chunks(c).into_iter().zip(any_points.iter()).enumerate().collect(),
+        |ctx, (round, (points_clump, any_point))| {
             // compute all possible multi-products of elements in points[round * c .. round * (c+1)]
-            // create new thread
-            let mut thread = Context::new(witness_gen_only, thread_id);
-            let ctx = &mut thread;
             // stores { any_point, any_point + points[0], any_point + points[1], any_point + points[0] + points[1] , ... }
             let mut bucket = Vec::with_capacity(1 << c);
             let any_point = into_strict_point(chip, ctx, any_point.clone());
@@ -294,7 +291,7 @@ where
                     bucket.push(new_point);
                 }
             }
-            let multi_prods = bool_scalars
+            bool_scalars
                 .iter()
                 .map(|bits| {
                     strict_ec_select_from_bits(
@@ -304,31 +301,19 @@ where
                         &bits[round * c..round * c + points_clump.len()],
                     )
                 })
-                .collect::<Vec<_>>();
-
-            (thread, multi_prods)
-        })
-        .unzip();
-    // we collect the new threads to ensure they are a FIXED order, otherwise later `assign_threads_in` will get confused
-    builder.threads[phase].extend(new_threads);
+                .collect::<Vec<_>>()
+        },
+    );
 
     // agg[j] = sum_{i=0..num_rounds} multi_prods[i][j] for j = 0..scalar_bits
-    let thread_ids = (0..scalar_bits).map(|_| builder.get_new_thread_id()).collect::<Vec<_>>();
-    let (new_threads, mut agg): (Vec<_>, Vec<_>) = thread_ids
-        .into_par_iter()
-        .enumerate()
-        .map(|(i, thread_id)| {
-            let mut thread = Context::new(witness_gen_only, thread_id);
-            let ctx = &mut thread;
-            let mut acc = multi_prods[0][i].clone();
-            for multi_prod in multi_prods.iter().skip(1) {
-                let _acc = ec_add_unequal(chip, ctx, &acc, &multi_prod[i], true);
-                acc = into_strict_point(chip, ctx, _acc);
-            }
-            (thread, acc)
-        })
-        .unzip();
-    builder.threads[phase].extend(new_threads);
+    let mut agg = parallelize_in(phase, builder, (0..scalar_bits).collect(), |ctx, i| {
+        let mut acc = multi_prods[0][i].clone();
+        for multi_prod in multi_prods.iter().skip(1) {
+            let _acc = ec_add_unequal(chip, ctx, &acc, &multi_prod[i], true);
+            acc = into_strict_point(chip, ctx, _acc);
+        }
+        acc
+    });
 
     // gets the LAST thread for single threaded work
     let ctx = builder.main(phase);