From 82bbfa864dd2731aa8301273f531fd071fd70352 Mon Sep 17 00:00:00 2001
From: CPerezz <carlos@dusk.network>
Date: Thu, 22 Apr 2021 01:32:08 +0200
Subject: [PATCH] Fix proving performance regressions

Adds back `rayon` under `std` feature just for the `quotient_poly` module
and just for a few lines which do not make the code harder to mantain or
less readable, but provide a boost in the performance of a 25%~30%.

Resolves: #512
---
 Cargo.toml                        | 11 ++++++++++-
 src/proof_system/quotient_poly.rs | 29 +++++++++++++++++++++++------
 2 files changed, 33 insertions(+), 7 deletions(-)

diff --git a/Cargo.toml b/Cargo.toml
index fbd16a70..f6ca4e1e 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -28,6 +28,7 @@ dusk-bls12_381 = {version = "0.8.0-rc.0", default-features = false, features = [
 dusk-jubjub = {version = "0.10.0-rc.0", default-features = false}
 itertools = {version = "0.9", default-features = false}
 hashbrown = {version = "0.9", default-features=false, features = ["ahash"]}
+rayon = {version = "1.3", optional = true}
 cfg-if = "1.0"
 # Dusk related deps for WASMI serde
 canonical = {version = "0.6", optional = true}
@@ -47,7 +48,8 @@ std = [
     "dusk-jubjub/default",
     "itertools/default",
     "hashbrown/default",
-    "alloc"
+    "alloc",
+    "rayon"
 ]
 alloc = ["dusk-bls12_381/alloc"]
 nightly = []
@@ -55,3 +57,10 @@ trace = []
 trace-print = ["trace"]
 canon = ["dusk-bls12_381/canon", "dusk-jubjub/canon", "canonical", "canonical_derive"]
 
+
+[profile.release]
+debug = false
+panic = 'abort'
+lto = true
+incremental = false
+codegen-units = 1
diff --git a/src/proof_system/quotient_poly.rs b/src/proof_system/quotient_poly.rs
index 2bf78c6a..60cd7789 100644
--- a/src/proof_system/quotient_poly.rs
+++ b/src/proof_system/quotient_poly.rs
@@ -11,6 +11,8 @@ use crate::{
 };
 use alloc::vec::Vec;
 use dusk_bls12_381::BlsScalar;
+#[cfg(feature = "std")]
+use rayon::prelude::*;
 
 /// This quotient polynomial can only be used for the standard composer
 /// Each composer will need to implement their own method for computing the
@@ -94,8 +96,13 @@ pub(crate) fn compute(
         (alpha, beta, gamma),
     );
 
-    let quotient: Vec<_> = (0..domain_4n.size())
-        .into_iter()
+    #[cfg(not(feature = "std"))]
+    let range = (0..domain_4n.size()).into_iter();
+
+    #[cfg(feature = "std")]
+    let range = (0..domain_4n.size()).into_par_iter();
+
+    let quotient: Vec<_> = range
         .map(|i| {
             let numerator = t_1[i] + t_2[i];
             let denominator = prover_key.v_h_coset_4n()[i];
@@ -129,8 +136,13 @@ fn compute_circuit_satisfiability_equation(
     let domain_4n = EvaluationDomain::new(4 * domain.size()).unwrap();
     let pi_eval_4n = domain_4n.coset_fft(pi_poly);
 
-    let t: Vec<_> = (0..domain_4n.size())
-        .into_iter()
+    #[cfg(not(feature = "std"))]
+    let range = (0..domain_4n.size()).into_iter();
+
+    #[cfg(feature = "std")]
+    let range = (0..domain_4n.size()).into_par_iter();
+
+    let t: Vec<_> = range
         .map(|i| {
             let wl = &wl_eval_4n[i];
             let wr = &wr_eval_4n[i];
@@ -212,8 +224,13 @@ fn compute_permutation_checks(
         compute_first_lagrange_poly_scaled(domain, alpha.square());
     let l1_alpha_sq_evals = domain_4n.coset_fft(&l1_poly_alpha.coeffs);
 
-    let t: Vec<_> = (0..domain_4n.size())
-        .into_iter()
+    #[cfg(not(feature = "std"))]
+    let range = (0..domain_4n.size()).into_iter();
+
+    #[cfg(feature = "std")]
+    let range = (0..domain_4n.size()).into_par_iter();
+
+    let t: Vec<_> = range
         .map(|i| {
             prover_key.permutation.compute_quotient_i(
                 i,