feat(groth16): control parallel proving threads

Uses the existing env variable to restrict the number of threads used for parallel proving.

Cleaned up version of #81

src/groth16/mod.rs

@@ -693,7 +693,7 @@ pub struct BatchPreparedVerifyingKey<E: Engine> {
     ic: Vec<E::G1Affine>,
 }
 
-pub trait ParameterSource<E: Engine> {
+pub trait ParameterSource<E: Engine>: Send {
     type G1Builder: SourceBuilder<E::G1Affine>;
     type G2Builder: SourceBuilder<E::G2Affine>;
 

src/groth16/prover.rs

@@ -10,7 +10,7 @@ use rayon::prelude::*;
 use super::{ParameterSource, Proof};
 use crate::domain::{EvaluationDomain, Scalar};
 use crate::gpu::{LockedFFTKernel, LockedMultiexpKernel};
-use crate::multicore::Worker;
+use crate::multicore::{Worker, THREAD_POOL};
 use crate::multiexp::{multiexp, DensityTracker, FullDensity};
 use crate::{
     Circuit, ConstraintSystem, Index, LinearCombination, SynthesisError, Variable, BELLMAN_VERSION,
@@ -180,7 +180,7 @@ where
 
 pub fn create_proof_batch_priority<E, C, P: ParameterSource<E>>(
     circuits: Vec<C>,
-    mut params: P,
+    params: P,
     r_s: Vec<E::Fr>,
     s_s: Vec<E::Fr>,
     priority: bool,
@@ -191,6 +191,20 @@ where
 {
     info!("Bellperson {} is being used!", BELLMAN_VERSION);
 
+    THREAD_POOL.install(|| create_proof_batch_priority_inner(circuits, params, r_s, s_s, priority))
+}
+
+fn create_proof_batch_priority_inner<E, C, P: ParameterSource<E>>(
+    circuits: Vec<C>,
+    mut params: P,
+    r_s: Vec<E::Fr>,
+    s_s: Vec<E::Fr>,
+    priority: bool,
+) -> Result<Vec<Proof<E>>, SynthesisError>
+where
+    E: Engine,
+    C: Circuit<E> + Send,
+{
     let mut provers = circuits
         .into_par_iter()
         .map(|circuit| -> Result<_, SynthesisError> {

src/multicore.rs

@@ -10,9 +10,26 @@ mod implementation {
     use crossbeam::{self, thread::Scope};
     use futures::{Future, IntoFuture, Poll};
     use futures_cpupool::{CpuFuture, CpuPool};
+    use lazy_static::lazy_static;
     use num_cpus;
     use std::env;
 
+    lazy_static! {
+        static ref NUM_CPUS: usize = if let Ok(num) = env::var("BELLMAN_NUM_CPUS") {
+            if let Ok(num) = num.parse() {
+                num
+            } else {
+                num_cpus::get()
+            }
+        } else {
+            num_cpus::get()
+        };
+        pub static ref THREAD_POOL: rayon::ThreadPool = rayon::ThreadPoolBuilder::new()
+            .num_threads(*NUM_CPUS)
+            .build()
+            .unwrap();
+    }
+
     #[derive(Clone)]
     pub struct Worker {
         cpus: usize,
@@ -31,17 +48,7 @@ mod implementation {
         }
 
         pub fn new() -> Worker {
-            let cpus = if let Ok(num) = env::var("BELLMAN_NUM_CPUS") {
-                if let Ok(num) = num.parse() {
-                    num
-                } else {
-                    num_cpus::get()
-                }
-            } else {
-                num_cpus::get()
-            };
-
-            Self::new_with_cpus(cpus)
+            Self::new_with_cpus(*NUM_CPUS)
         }
 
         pub fn log_num_cpus(&self) -> u32 {