Simplify MSM implementation, and small speedup

CHANGELOG.md

@@ -58,6 +58,7 @@ The main features of this release are:
 - #144 (ark-poly) Add serialization for polynomials and evaluations
 - #149 (ark-serialize) Add an impl of `CanonicalSerialize/Deserialize` for `String`.
 - #153 (ark-serialize) Add an impl of `CanonicalSerialize/Deserialize` for `Rc<T>`.
+- #157 (ark-ec) Speed up `variable_base_msm` by not relying on unnecessary normalization.
 - #158 (ark-serialize) Add an impl of `CanonicalSerialize/Deserialize` for `()`.
 
 ### Bug fixes

ec/src/msm/variable_base.rs

@@ -9,13 +9,10 @@ use rayon::prelude::*;
 pub struct VariableBaseMSM;
 
 impl VariableBaseMSM {
-    fn msm_inner<G: AffineCurve>(
+    pub fn multi_scalar_mul<G: AffineCurve>(
         bases: &[G],
         scalars: &[<G::ScalarField as PrimeField>::BigInt],
-    ) -> G::Projective
-    where
-        G::Projective: ProjectiveCurve<Affine = G>,
-    {
+    ) -> G::Projective {
         let c = if scalars.len() < 32 {
             3
         } else {
@@ -28,6 +25,11 @@ impl VariableBaseMSM {
         let zero = G::Projective::zero();
         let window_starts: Vec<_> = (0..num_bits).step_by(c).collect();
 
+        let size = ark_std::cmp::min(bases.len(), scalars.len());
+        let scalars = &scalars[..size];
+        let bases = &bases[..size];
+        let scalars_and_bases = scalars.iter().zip(bases).filter(|(s, _)| !s.is_zero());
+
         // Each window is of size `c`.
         // We divide up the bits 0..num_bits into windows of size `c`, and
         // in parallel process each such window.
@@ -36,42 +38,39 @@ impl VariableBaseMSM {
                 let mut res = zero;
                 // We don't need the "zero" bucket, so we only have 2^c - 1 buckets
                 let mut buckets = vec![zero; (1 << c) - 1];
-                scalars
-                    .iter()
-                    .zip(bases)
-                    .filter(|(s, _)| !s.is_zero())
-                    .for_each(|(&scalar, base)| {
-                        if scalar == fr_one {
-                            // We only process unit scalars once in the first window.
-                            if w_start == 0 {
-                                res.add_assign_mixed(base);
-                            }
-                        } else {
-                            let mut scalar = scalar;
+                scalars_and_bases.clone().for_each(|(&scalar, base)| {
+                    if scalar == fr_one {
+                        // We only process unit scalars once in the first window.
+                        if w_start == 0 {
+                            res.add_assign_mixed(base);
+                        }
+                    } else {
+                        let mut scalar = scalar;
 
-                            // We right-shift by w_start, thus getting rid of the
-                            // lower bits.
-                            scalar.divn(w_start as u32);
+                        // We right-shift by w_start, thus getting rid of the
+                        // lower bits.
+                        scalar.divn(w_start as u32);
 
-                            // We mod the remaining bits by the window size.
-                            let scalar = scalar.as_ref()[0] % (1 << c);
+                        // We mod the remaining bits by the window size.
+                        let scalar = scalar.as_ref()[0] % (1 << c);
 
-                            // If the scalar is non-zero, we update the corresponding
-                            // bucket.
-                            // (Recall that `buckets` doesn't have a zero bucket.)
-                            if scalar != 0 {
-                                buckets[(scalar - 1) as usize].add_assign_mixed(base);
-                            }
+                        // If the scalar is non-zero, we update the corresponding
+                        // bucket.
+                        // (Recall that `buckets` doesn't have a zero bucket.)
+                        if scalar != 0 {
+                            buckets[(scalar - 1) as usize].add_assign_mixed(base);
                         }
-                    });
-                let buckets = G::Projective::batch_normalization_into_affine(&buckets);
+                    }
+                });
 
+                // We could first normalize `buckets` and then use mixed-addition
+                // here, but that's slower for the kinds of groups we care about
+                // (Short Weierstrass curves and Twisted Edwards curves).
                 let mut running_sum = G::Projective::zero();
-                for b in buckets.into_iter().rev() {
-                    running_sum.add_assign_mixed(&b);
-                    res += running_sum;
-                }
-
+                buckets.into_iter().rev().for_each(|b| {
+                    running_sum += &b;
+                    res += &running_sum;
+                });
                 res
             })
             .collect();
@@ -81,7 +80,7 @@ impl VariableBaseMSM {
 
         // We're traversing windows from high to low.
         lowest
-            + window_sums[1..]
+            + &window_sums[1..]
                 .iter()
                 .rev()
                 .fold(zero, |mut total, sum_i| {
@@ -92,11 +91,4 @@ impl VariableBaseMSM {
                     total
                 })
     }
-
-    pub fn multi_scalar_mul<G: AffineCurve>(
-        bases: &[G],
-        scalars: &[<G::ScalarField as PrimeField>::BigInt],
-    ) -> G::Projective {
-        Self::msm_inner(bases, scalars)
-    }
 }