Add unsafe and Arc implementation

Add BABELSTREAM_NUM_THREADS env option
Refactor driver
Bump dependencies

rust-stream/Cargo.lock

@@ -30,9 +30,9 @@ checksum = "cdb031dd78e28731d87d56cc8ffef4a8f36ca26c38fe2de700543e627f8a464a"
 
 [[package]]
 name = "bitflags"
-version = "1.2.1"
+version = "1.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cf1de2fe8c75bc145a2f577add951f8134889b4795d47466a54a5c846d691693"
+checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
 
 [[package]]
 name = "cfg-if"
@@ -102,9 +102,9 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-deque"
-version = "0.8.0"
+version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "94af6efb46fef72616855b036a624cf27ba656ffc9be1b9a3c931cfc7749a9a9"
+checksum = "6455c0ca19f0d2fbf751b908d5c55c1f5cbc65e03c4225427254b46890bdde1e"
 dependencies = [
  "cfg-if",
  "crossbeam-epoch",
@@ -186,18 +186,18 @@ dependencies = [
 
 [[package]]
 name = "hermit-abi"
-version = "0.1.18"
+version = "0.1.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "322f4de77956e22ed0e5032c359a0f1273f1f7f0d79bfa3b8ffbc730d7fbcc5c"
+checksum = "62b467343b94ba476dcb2500d242dadbb39557df889310ac77c5d99100aaac33"
 dependencies = [
  "libc",
 ]
 
 [[package]]
 name = "instant"
-version = "0.1.9"
+version = "0.1.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61124eeebbd69b8190558df225adf7e4caafce0d743919e5d6b19652314ec5ec"
+checksum = "7a5bbe824c507c5da5956355e86a746d82e0e1464f65d862cc5e71da70e94b2c"
 dependencies = [
  "cfg-if",
 ]
@@ -220,15 +220,15 @@ checksum = "e2abad23fbc42b3700f2f279844dc832adb2b2eb069b2df918f455c4e18cc646"
 
 [[package]]
 name = "libc"
-version = "0.2.97"
+version = "0.2.108"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "12b8adadd720df158f4d70dfe7ccc6adb0472d7c55ca83445f6a5ab3e36f8fb6"
+checksum = "8521a1b57e76b1ec69af7599e75e38e7b7fad6610f037db8c79b127201b5d119"
 
 [[package]]
 name = "lock_api"
-version = "0.4.4"
+version = "0.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0382880606dff6d15c9476c416d18690b72742aa7b605bb6dd6ec9030fbf07eb"
+checksum = "712a4d093c9976e24e7dbca41db895dabcbac38eb5f4045393d17a95bdfb1109"
 dependencies = [
  "scopeguard",
 ]
@@ -253,9 +253,9 @@ dependencies = [
 
 [[package]]
 name = "mio"
-version = "0.7.11"
+version = "0.7.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cf80d3e903b34e0bd7282b218398aec54e082c840d9baf8339e0080a0c542956"
+checksum = "8067b404fe97c70829f082dec8bcf4f71225d7eaea1d8645349cb76fa06205cc"
 dependencies = [
  "libc",
  "log",
@@ -303,9 +303,9 @@ dependencies = [
 
 [[package]]
 name = "parking_lot"
-version = "0.11.1"
+version = "0.11.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d7744ac029df22dca6284efe4e898991d28e3085c706c972bcd7da4a27a15eb"
+checksum = "7d17b78036a60663b797adeaee46f5c9dfebb86948d1255007a1d6be0271ff99"
 dependencies = [
  "instant",
  "lock_api",
@@ -314,9 +314,9 @@ dependencies = [
 
 [[package]]
 name = "parking_lot_core"
-version = "0.8.3"
+version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa7a782938e745763fe6907fc6ba86946d72f49fe7e21de074e08128a99fb018"
+checksum = "d76e8e1493bcac0d2766c42737f34458f1c8c50c0d23bcb24ea953affb273216"
 dependencies = [
  "cfg-if",
  "instant",
@@ -361,18 +361,18 @@ dependencies = [
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.27"
+version = "1.0.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0d8caf72986c1a598726adc988bb5984792ef84f5ee5aa50209145ee8077038"
+checksum = "ba508cc11742c0dc5c1659771673afbab7a0efab23aa17e854cbab0837ed0b43"
 dependencies = [
  "unicode-xid",
 ]
 
 [[package]]
 name = "quote"
-version = "1.0.9"
+version = "1.0.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c3d0b9745dc2debf507c8422de05d7226cc1f0644216dfdfead988f9b1ab32a7"
+checksum = "38bc8cc6a5f2e3655e0899c1b848643b2562f853f114bfec7be120678e3ace05"
 dependencies = [
  "proc-macro2",
 ]
@@ -404,9 +404,9 @@ dependencies = [
 
 [[package]]
 name = "redox_syscall"
-version = "0.2.8"
+version = "0.2.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "742739e41cd49414de871ea5e549afb7e2a3ac77b589bcbebe8c82fab37147fc"
+checksum = "8383f39639269cde97d255a32bdb68c047337295414940c68bdd30c2e13203ff"
 dependencies = [
  "bitflags",
 ]
@@ -502,9 +502,9 @@ dependencies = [
 
 [[package]]
 name = "smallvec"
-version = "1.6.1"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fe0f37c9e8f3c5a4a66ad655a93c74daac4ad00c441533bf5c6e7990bb42604e"
+checksum = "1ecab6c735a6bb4139c0caafd0cc3635748bbb3acf4550e8138122099251f309"
 
 [[package]]
 name = "strsim"
@@ -514,9 +514,9 @@ checksum = "8ea5119cdb4c55b55d432abb513a0429384878c15dde60cc77b1c99de1a95a6a"
 
 [[package]]
 name = "structopt"
-version = "0.3.21"
+version = "0.3.25"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5277acd7ee46e63e5168a80734c9f6ee81b1367a7d8772a2d765df2a3705d28c"
+checksum = "40b9788f4202aa75c240ecc9c15c65185e6a39ccdeb0fd5d008b98825464c87c"
 dependencies = [
  "clap",
  "lazy_static",
@@ -525,9 +525,9 @@ dependencies = [
 
 [[package]]
 name = "structopt-derive"
-version = "0.4.14"
+version = "0.4.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5ba9cdfda491b814720b6b06e0cac513d922fc407582032e8706e9f137976f90"
+checksum = "dcb5ae327f9cc13b68763b5749770cb9e048a99bd9dfdfa58d0cf05d5f64afe0"
 dependencies = [
  "heck",
  "proc-macro-error",
@@ -538,9 +538,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "1.0.73"
+version = "1.0.82"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f71489ff30030d2ae598524f61326b902466f72a0fb1a8564c001cc63425bcc7"
+checksum = "8daf5dd0bb60cbd4137b1b587d2fc0ae729bc07cf01cd70b36a1ed5ade3b9d59"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -573,15 +573,15 @@ checksum = "56dee185309b50d1f11bfedef0fe6d036842e3fb77413abef29f8f8d1c5d4c1c"
 
 [[package]]
 name = "unicode-segmentation"
-version = "1.7.1"
+version = "1.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bb0d2e7be6ae3a5fa87eed5fb451aff96f2573d2694942e40543ae0bbe19c796"
+checksum = "8895849a949e7845e06bd6dc1aa51731a103c42707010a5b591c0038fb73385b"
 
 [[package]]
 name = "unicode-width"
-version = "0.1.8"
+version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9337591893a19b88d8d87f2cec1e73fad5cdfd10e5a6f349f498ad6ea2ffb1e3"
+checksum = "3ed742d4ea2bd1176e236172c8429aaf54486e7ac098db29ffe6529e0ce50973"
 
 [[package]]
 name = "unicode-xid"

rust-stream/README.md

@@ -10,15 +10,21 @@ Currently, we support three CPU threading API as devices:
   see [rayon_stream.rs](src/rayon_stream.rs)
 * [Crossbeam](https://github.com/crossbeam-rs/crossbeam) - Parallel with partitions per thread,
   see [crossbeam_stream.rs](src/crossbeam_stream.rs)
+* Arc - Parallel with `Vec` per thread (static partitions) wrapped in `Mutex` contained in `Arc`s,
+  see [crossbeam_stream.rs](src/arc_stream.rs)
+* Unsafe - Parallel with unsafe pointer per thread (static partitions) to `Vec`,
+  see [crossbeam_stream.rs](src/unsafe_stream.rs)
 
 In addition, this implementation also supports the following extra flags:
-
+****
 ```
 --init    Initialise each benchmark array at allocation time on the main thread
 --malloc  Use libc malloc instead of the Rust's allocator for benchmark array allocation
 --pin     Pin threads to distinct cores, this has NO effect in Rayon devices
 ```
 
+Max thread count is controlled by the environment variable `BABELSTREAM_NUM_THREADS` which is compatible for all devices (avoid setting `RAYON_NUM_THREADS`, the implementation will issue a warning if this happened).   
+
 There is an ongoing investigation on potential performance issues under NUMA situations. As part of
 the experiment, this implementation made use of the
 provisional [Allocator traits](https://github.com/rust-lang/rust/issues/32838) which requires rust

rust-stream/rustfmt.toml

@@ -54,7 +54,7 @@ use_field_init_shorthand = false
 force_explicit_abi = true
 condense_wildcard_suffixes = false
 color = "Auto"
-required_version = "1.4.36"
+required_version = "1.4.38"
 unstable_features = false
 disable_all_formatting = false
 skip_children = false

rust-stream/src/arc_stream.rs

@@ -0,0 +1,254 @@
+use std::iter::Sum;
+use std::sync::{Arc, Mutex};
+
+use self::core_affinity::CoreId;
+use crate::stream::{AllocatorType, ArrayType, RustStream, StreamData};
+
+struct ArcHeapData<T: ArrayType, A: AllocatorType> {
+  a_chunks: Vec<Arc<Mutex<Vec<T, A>>>>,
+  b_chunks: Vec<Arc<Mutex<Vec<T, A>>>>,
+  c_chunks: Vec<Arc<Mutex<Vec<T, A>>>>,
+}
+
+pub struct ArcDevice<T: ArrayType, A: AllocatorType> {
+  pub(crate) ncore: usize,
+  pub(crate) pin: bool,
+  pub(crate) core_ids: Vec<CoreId>,
+  data: ArcHeapData<T, A>,
+}
+
+impl<T: ArrayType, A: AllocatorType> ArcDevice<T, A> {
+  pub fn new(ncore: usize, pin: bool, alloc: A) -> Self {
+    let mut core_ids = match core_affinity::get_core_ids() {
+      Some(xs) => xs,
+      None => {
+        colour::e_red_ln!("Cannot enumerate cores, pinning will not work if enabled");
+        (0..ncore).map(|i| CoreId { id: i }).collect()
+      }
+    };
+    core_ids.resize(ncore, core_ids[0]);
+
+    let lift =
+      || (0..ncore).map(|_| return Arc::new(Mutex::new(Vec::new_in(alloc)))).collect::<Vec<_>>();
+    let data = ArcHeapData { a_chunks: lift(), b_chunks: lift(), c_chunks: lift() };
+
+    ArcDevice { ncore, pin, core_ids, data }
+  }
+
+  pub fn ref_a(&self, t: usize) -> Arc<Mutex<Vec<T, A>>> { self.data.a_chunks[t].clone() }
+
+  pub fn ref_b(&self, t: usize) -> Arc<Mutex<Vec<T, A>>> { self.data.b_chunks[t].clone() }
+
+  pub fn ref_c(&self, t: usize) -> Arc<Mutex<Vec<T, A>>> { self.data.c_chunks[t].clone() }
+
+  // divide the length by the number of cores, the last core gets less work if it does not divide
+  fn chunk_size(&self, len: usize, t: usize) -> usize {
+    assert!(t < self.ncore);
+    let chunk = (len as f64 / self.ncore as f64).ceil() as usize;
+    if t == self.ncore - 1 {
+      len - (t * chunk)
+    } else {
+      chunk
+    }
+  }
+}
+
+extern crate core_affinity;
+
+// Arc+Mutex threaded version, it should be semantically equal to the single threaded version
+impl<T: 'static + ArrayType + Sync + Send + Sum, A: AllocatorType + Sync + Send + 'static>
+  RustStream<T> for StreamData<T, ArcDevice<T, A>, A>
+{
+  fn init_arrays(&mut self) {
+    let init = self.init;
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(&|t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_b = self.device.ref_b(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          ref_a.lock().unwrap().resize(n, init.0);
+          ref_b.lock().unwrap().resize(n, init.1);
+          ref_c.lock().unwrap().resize(n, init.2);
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+  fn read_arrays(&mut self) {
+    let range = self.size;
+    let unlift = |drain: &mut Vec<T, A>, source: &Vec<Arc<Mutex<Vec<T, A>>>>| {
+      let xs =
+        source.into_iter().flat_map(|x| x.lock().unwrap().clone().into_iter()).collect::<Vec<_>>();
+      for i in 0..range {
+        drain[i] = xs[i];
+      }
+    };
+    unlift(&mut self.a, &self.device.data.a_chunks);
+    unlift(&mut self.b, &self.device.data.b_chunks);
+    unlift(&mut self.c, &self.device.data.c_chunks);
+  }
+
+  fn copy(&mut self) {
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(move |t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = ref_a.lock().unwrap();
+          let mut c = ref_c.lock().unwrap();
+          for i in 0..n {
+            c[i] = a[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn mul(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(move |t| {
+        let ref_b = self.device.ref_b(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let mut b = ref_b.lock().unwrap();
+          let c = ref_c.lock().unwrap();
+          for i in 0..n {
+            b[i] = scalar * c[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn add(&mut self) {
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(&|t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_b = self.device.ref_b(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = ref_a.lock().unwrap();
+          let b = ref_b.lock().unwrap();
+          let mut c = ref_c.lock().unwrap();
+          for i in 0..n {
+            c[i] = a[i] + b[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn triad(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(&|t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_b = self.device.ref_b(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let mut a = ref_a.lock().unwrap();
+          let b = ref_b.lock().unwrap();
+          let c = ref_c.lock().unwrap();
+          for i in 0..n {
+            a[i] = b[i] + scalar * c[i]
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn nstream(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(&|t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_b = self.device.ref_b(t);
+        let ref_c = self.device.ref_c(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let mut a = ref_a.lock().unwrap();
+          let b = ref_b.lock().unwrap();
+          let c = ref_c.lock().unwrap();
+          for i in 0..n {
+            a[i] += b[i] + scalar * c[i]
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn dot(&mut self) -> T {
+    let pin = self.device.pin;
+    (0..self.device.ncore)
+      .map(&|t| {
+        let ref_a = self.device.ref_a(t);
+        let ref_b = self.device.ref_b(t);
+        let core = self.device.core_ids[t];
+        let n = self.device.chunk_size(self.size, t);
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = ref_a.lock().unwrap();
+          let b = ref_b.lock().unwrap();
+          let mut p = T::default();
+          for i in 0..n {
+            p += a[i] * b[i];
+          }
+          p
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .map(|t| t.join().unwrap())
+      .sum()
+  }
+}

rust-stream/src/crossbeam_stream.rs

@@ -6,13 +6,13 @@ use crossbeam::thread;
 use self::core_affinity::CoreId;
 use crate::stream::{AllocatorType, ArrayType, RustStream, StreamData};
 
-pub struct ThreadedDevice {
+pub struct CrossbeamDevice {
   pub(crate) ncore: usize,
   pub(crate) pin: bool,
   pub(crate) core_ids: Vec<CoreId>,
 }
 
-impl ThreadedDevice {
+impl CrossbeamDevice {
   pub fn new(ncore: usize, pin: bool) -> Self {
     let mut core_ids = match core_affinity::get_core_ids() {
       Some(xs) => xs,
@@ -22,15 +22,13 @@ impl ThreadedDevice {
       }
     };
     core_ids.resize(ncore, core_ids[0]);
-    ThreadedDevice { ncore, pin, core_ids }
+    CrossbeamDevice { ncore, pin, core_ids }
   }
 }
 
-impl ThreadedDevice {
+impl CrossbeamDevice {
   // divide the length by the number of cores, the last core gets less work if it does not divide
-  fn chunk_size(&self, len: usize) -> usize {
+  fn chunk_size(&self, len: usize) -> usize { (len as f64 / self.ncore as f64).ceil() as usize }
-    (len as f64 / self.ncore as f64).ceil() as usize
-  }
 
   // make a mutable chunk from the vec
   fn mk_mut_chunks<'a, T, A: AllocatorType>(&self, xs: &'a mut Vec<T, A>) -> ChunksMut<'a, T> {
@@ -48,7 +46,7 @@ extern crate core_affinity;
 
 // Crossbeam threaded version, it should be semantically equal to the single threaded version
 impl<T: ArrayType + Sync + Send + Sum, A: AllocatorType + Sync + Send> RustStream<T>
-  for StreamData<T, ThreadedDevice, A>
+  for StreamData<T, CrossbeamDevice, A>
 {
   fn init_arrays(&mut self) {
     thread::scope(|s| {

rust-stream/src/lib.rs

@@ -2,6 +2,7 @@
 #![feature(vec_into_raw_parts)]
 
 use std::alloc::System;
+use std::env;
 use std::fmt::{Debug, Display};
 use std::iter::Sum;
 use std::mem::size_of;
@@ -11,15 +12,19 @@ use num_traits::abs;
 use structopt::StructOpt;
 use tabular::{Row, Table};
 
-use crate::crossbeam_stream::ThreadedDevice;
+use crate::arc_stream::ArcDevice;
+use crate::crossbeam_stream::CrossbeamDevice;
 use crate::plain_stream::SerialDevice;
 use crate::rayon_stream::RayonDevice;
 use crate::stream::{AllocatorType, ArrayType, RustStream, StreamData};
+use crate::unsafe_stream::UnsafeDevice;
 
+mod arc_stream;
 mod crossbeam_stream;
 mod plain_stream;
 mod rayon_stream;
 mod stream;
+mod unsafe_stream;
 
 #[derive(Debug, StructOpt)]
 struct Options {
@@ -119,9 +124,7 @@ fn check_solution<T: ArrayType + Display + Sum + Into<f64>, D, A: AllocatorType>
 fn run_cpu<T: ArrayType + Sync + Send + Sum + Into<f64> + Display, D, A: AllocatorType>(
   option: &Options, mut stream: StreamData<T, D, A>,
 ) -> bool
-where
+where StreamData<T, D, A>: RustStream<T> {
-  StreamData<T, D, A>: RustStream<T>,
-{
   let benchmark = match (option.nstream_only, option.triad_only) {
     (true, false) => Benchmark::NStream,
     (false, true) => Benchmark::Triad,
@@ -175,6 +178,7 @@ where
 
   let tabulate = |xs: &Vec<Duration>, name: &str, t_size: usize| -> Vec<(&str, String)> {
     let tail = &xs[1..]; // tail only
+
     // do stats
     let max = tail.iter().max().map(|d| d.as_secs_f64());
     let min = tail.iter().min().map(|d| d.as_secs_f64());
@@ -234,6 +238,7 @@ where
   let solutions_correct = match benchmark {
     Benchmark::All => {
       let (results, sum) = stream.run_all(option.numtimes);
+      stream.read_arrays();
       let correct = check_solution(benchmark, option.numtimes, &stream, Some(sum));
       tabulate_all(vec![
         tabulate(&results.copy, "Copy", 2 * array_bytes),
@@ -246,12 +251,14 @@ where
     }
     Benchmark::NStream => {
       let results = stream.run_nstream(option.numtimes);
+      stream.read_arrays();
       let correct = check_solution(benchmark, option.numtimes, &stream, None);
       tabulate_all(vec![tabulate(&results, "Nstream", 4 * array_bytes)]);
       correct
     }
     Benchmark::Triad => {
       let results = stream.run_triad(option.numtimes);
+      stream.read_arrays();
       let correct = check_solution(benchmark, option.numtimes, &stream, None);
       let total_bytes = 3 * array_bytes * option.numtimes;
       let bandwidth = giga_scale * (total_bytes as f64 / results.as_secs_f64());
@@ -260,7 +267,7 @@ where
       correct
     }
   };
-  &stream.clean_up();
+  stream.clean_up();
   solutions_correct
 }
 
@@ -274,139 +281,138 @@ static START_SCALAR: f32 = 0.4;
 static FLOAT_INIT_SCALAR: f32 = START_SCALAR;
 static FLOAT_INIT: (f32, f32, f32) = (START_A, START_B, START_C);
 
-static DOUBLE_START_SCALAR: f64 = START_SCALAR as f64;
+static DOUBLE_INIT_SCALAR: f64 = START_SCALAR as f64;
 static DOUBLE_INIT: (f64, f64, f64) = (START_A as f64, START_B as f64, START_C as f64);
 
 pub fn run(args: &Vec<String>) -> bool {
+  let opt: Options = Options::from_iter(args);
 
-  let options: Options = Options::from_iter(args);
+  if opt.numtimes < 2 {
-
-  if options.numtimes < 2 {
     panic!("numtimes must be >= 2")
   }
 
   let alloc = System;
-  let alloc_name = if options.malloc { "libc-malloc" } else { "rust-system" };
+  let alloc_name = if opt.malloc { "libc-malloc" } else { "rust-system" };
+
+  fn mk_data<T: ArrayType, D, A: AllocatorType>(
+    opt: &Options, init: (T, T, T), scalar: T, dev: D, alloc: A,
+  ) -> StreamData<T, D, A> {
+    StreamData::new_in(opt.arraysize, scalar, init, dev, alloc, opt.malloc, opt.init)
+  }
+
+  let num_thread_key = "BABELSTREAM_NUM_THREADS";
+  let max_ncores = num_cpus::get();
+  let ncores = match env::var(num_thread_key) {
+    Ok(v) => match v.parse::<i64>() {
+      Err(bad) => {
+        colour::e_yellow_ln!(
+          "Cannot parse {} (reason: {}), defaulting to {}",
+          bad,
+          num_thread_key,
+          max_ncores
+        );
+        max_ncores
+      }
+      Ok(n) if n <= 0 || n > max_ncores as i64 => {
+        println!("{} out of bound ({}), defaulting to {}", num_thread_key, n, max_ncores);
+        max_ncores
+      }
+      Ok(n) => n as usize,
+    },
+    Err(_) => {
+      println!("{} not set, defaulting to max ({})", num_thread_key, max_ncores);
+      max_ncores
+    }
+  };
 
   let rayon_device = &|| {
-    let dev = RayonDevice { pool: rayon::ThreadPoolBuilder::default().build().unwrap() };
+    let rayon_num_thread_key = "RAYON_NUM_THREADS";
-    if !options.csv {
+    if env::var(rayon_num_thread_key).is_ok() {
+      colour::e_yellow_ln!("{} is ignored, set {} instead", rayon_num_thread_key, num_thread_key)
+    }
+    let dev = RayonDevice {
+      pool: rayon::ThreadPoolBuilder::default().num_threads(ncores).build().unwrap(),
+    };
+    if !opt.csv {
       println!("Using {} thread(s), alloc={}", dev.pool.current_num_threads(), alloc_name);
-      if options.pin {
+      if opt.pin {
         colour::e_yellow_ln!("Pinning threads have no effect on Rayon!")
       }
     }
-    if options.float {
+    if opt.float {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, FLOAT_INIT, FLOAT_INIT_SCALAR, dev, alloc))
-        &options,
-        StreamData::new_in(
-          options.arraysize,
-          FLOAT_INIT_SCALAR,
-          FLOAT_INIT,
-          dev,
-          alloc,
-          options.malloc,
-          options.init,
-        ),
-      )
     } else {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, DOUBLE_INIT, DOUBLE_INIT_SCALAR, dev, alloc))
-        &options,
+    }
-        StreamData::new_in(
+  };
-          options.arraysize,
+
-          DOUBLE_START_SCALAR,
+  let arc_device = &|| {
-          DOUBLE_INIT,
+    if !opt.csv {
-          dev,
+      println!("Using {} thread, pin={}, alloc={}", ncores, opt.pin, alloc_name);
-          alloc,
+    }
-          options.malloc,
+    if opt.float {
-          options.init,
+      let dev = ArcDevice::<f32, _>::new(ncores, opt.pin, alloc);
-        ),
+      run_cpu(&opt, mk_data(&opt, FLOAT_INIT, FLOAT_INIT_SCALAR, dev, alloc))
-      )
+    } else {
+      let dev = ArcDevice::<f64, _>::new(ncores, opt.pin, alloc);
+      run_cpu(&opt, mk_data(&opt, DOUBLE_INIT, DOUBLE_INIT_SCALAR, dev, alloc))
+    }
+  };
+
+  let unsafe_device = &|| {
+    if !opt.csv {
+      println!("Using {} thread, pin={}, alloc={}", ncores, opt.pin, alloc_name);
+    }
+    if opt.float {
+      let dev = UnsafeDevice::<f32>::new(ncores, opt.pin);
+      run_cpu(&opt, mk_data(&opt, FLOAT_INIT, FLOAT_INIT_SCALAR, dev, alloc))
+    } else {
+      let dev = UnsafeDevice::<f64>::new(ncores, opt.pin);
+      run_cpu(&opt, mk_data(&opt, DOUBLE_INIT, DOUBLE_INIT_SCALAR, dev, alloc))
     }
   };
 
   let crossbeam_device = &|| {
-    let ncores = num_cpus::get();
+    let dev = CrossbeamDevice::new(ncores, opt.pin);
-    let dev = ThreadedDevice::new(ncores, options.pin);
+    if !opt.csv {
-    if !options.csv {
+      println!("Using {} thread(s), pin={}, alloc={}", ncores, opt.pin, alloc_name)
-      println!("Using {} thread(s), pin={}, alloc={}", ncores, options.pin, alloc_name)
     }
-    if options.float {
+    if opt.float {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, FLOAT_INIT, FLOAT_INIT_SCALAR, dev, alloc))
-        &options,
-        StreamData::new_in(
-          options.arraysize,
-          FLOAT_INIT_SCALAR,
-          FLOAT_INIT,
-          dev,
-          alloc,
-          options.malloc,
-          options.init,
-        ),
-      )
     } else {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, DOUBLE_INIT, DOUBLE_INIT_SCALAR, dev, alloc))
-        &options,
-        StreamData::new_in(
-          options.arraysize,
-          DOUBLE_START_SCALAR,
-          DOUBLE_INIT,
-          dev,
-          alloc,
-          options.malloc,
-          options.init,
-        ),
-      )
     }
   };
+
   let st_device = &|| {
-    let dev = SerialDevice { pin: options.pin };
+    let dev = SerialDevice { pin: opt.pin };
-    if !options.csv {
+    if !opt.csv {
-      println!("Using 1 thread, pin={}, alloc={}", options.pin, alloc_name);
+      println!("Using 1 thread, pin={}, alloc={}", opt.pin, alloc_name);
     }
-    if options.float {
+    if opt.float {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, FLOAT_INIT, FLOAT_INIT_SCALAR, dev, alloc))
-        &options,
-        StreamData::new_in(
-          options.arraysize,
-          FLOAT_INIT_SCALAR,
-          FLOAT_INIT,
-          dev,
-          alloc,
-          options.malloc,
-          options.init,
-        ),
-      )
     } else {
-      run_cpu(
+      run_cpu(&opt, mk_data(&opt, DOUBLE_INIT, DOUBLE_INIT_SCALAR, dev, alloc))
-        &options,
-        StreamData::new_in(
-          options.arraysize,
-          DOUBLE_START_SCALAR,
-          DOUBLE_INIT,
-          dev,
-          alloc,
-          options.malloc,
-          options.init,
-        ),
-      )
     }
   };
+
   let devices: Vec<(String, &'_ dyn Fn() -> bool)> = vec![
-    ("CPU (Rayon)".to_string(), rayon_device),
-    (format!("CPU (Crossbeam, pinning={})", options.pin), crossbeam_device),
     ("CPU (Single threaded)".to_string(), st_device),
+    ("CPU (Rayon)".to_string(), rayon_device),
+    (format!("CPU (Arc, pinning={})", opt.pin), arc_device),
+    (format!("CPU (Unsafe, pinning={})", opt.pin), unsafe_device),
+    (format!("CPU (Crossbeam, pinning={})", opt.pin), crossbeam_device),
   ];
 
-  if options.list {
+  if opt.list {
     devices.iter().enumerate().for_each(|(i, (name, _))| {
       println!("[{}] {}", i, name);
     });
     true
   } else {
-    match devices.get(options.device) {
+    match devices.get(opt.device) {
       Some((name, run)) => {
-        if !&options.csv {
+        if !&opt.csv {
           println!(
             "BabelStream\n\
                               Version: {}\n\
@@ -414,14 +420,14 @@ pub fn run(args: &Vec<String>) -> bool {
             VERSION.unwrap_or("unknown"),
             name
           );
-          if options.init {
+          if opt.init {
             println!("Initialising arrays on main thread");
           }
         }
         run()
       }
       None => {
-        eprintln!("Device index {} not available", options.device);
+        eprintln!("Device index {} not available", opt.device);
         false
       }
     }

rust-stream/src/stream.rs

@@ -124,6 +124,7 @@ impl<T: Default + Clone, D, A: AllocatorType> StreamData<T, D, A> {
 
 pub trait RustStream<T: Default> {
   fn init_arrays(&mut self);
+  fn read_arrays(&mut self) {} // default to no-op as most impl. doesn't need this
   fn copy(&mut self);
   fn mul(&mut self);
   fn add(&mut self);

rust-stream/src/unsafe_stream.rs

@@ -0,0 +1,266 @@
+extern crate core_affinity;
+
+use std::alloc::Allocator;
+use std::iter::Sum;
+use std::ops::Range;
+
+use crate::stream::{AllocatorType, ArrayType, RustStream, StreamData};
+
+use self::core_affinity::CoreId;
+
+#[derive(Debug, Copy, Clone)]
+struct UnsafeData<T>(*mut T, usize);
+
+impl<T: ArrayType> UnsafeData<T> {
+  fn empty() -> UnsafeData<T> { UnsafeData(([] as [T; 0]).as_mut_ptr(), 0) }
+  fn new<A: Allocator>(xs: &mut Vec<T, A>) -> UnsafeData<T> {
+    UnsafeData(xs.as_mut_ptr(), xs.len())
+  }
+
+  fn get_slice(&self) -> &mut [T] { unsafe { std::slice::from_raw_parts_mut(self.0, self.1) } }
+}
+
+unsafe impl<T> Send for UnsafeData<T> {}
+unsafe impl<T> Sync for UnsafeData<T> {}
+
+#[derive(Debug, Copy, Clone)]
+struct UnsafeRefs<T> {
+  a: UnsafeData<T>,
+  b: UnsafeData<T>,
+  c: UnsafeData<T>,
+}
+
+unsafe impl<T> Send for UnsafeRefs<T> {}
+unsafe impl<T> Sync for UnsafeRefs<T> {}
+
+pub struct UnsafeDevice<T: ArrayType> {
+  pub(crate) ncore: usize,
+  pub(crate) pin: bool,
+  pub(crate) core_ids: Vec<CoreId>,
+  data: UnsafeRefs<T>,
+}
+
+impl<T: ArrayType> UnsafeDevice<T> {
+  pub fn new(ncore: usize, pin: bool) -> Self {
+    let mut core_ids = match core_affinity::get_core_ids() {
+      Some(xs) => xs,
+      None => {
+        colour::e_red_ln!("Cannot enumerate cores, pinning will not work if enabled");
+        (0..ncore).map(|i| CoreId { id: i }).collect()
+      }
+    };
+    core_ids.resize(ncore, core_ids[0]);
+
+    UnsafeDevice {
+      ncore,
+      pin,
+      core_ids,
+      data: UnsafeRefs { a: UnsafeData::empty(), b: UnsafeData::empty(), c: UnsafeData::empty() },
+    }
+  }
+
+  fn thread_ranges(&self, len: usize) -> Vec<(usize, Range<usize>)> {
+    let chunk = (len as f64 / self.ncore as f64).ceil() as usize;
+    (0..self.ncore)
+      .map(|t| {
+        (t, if t == self.ncore - 1 { (t * chunk)..len } else { (t * chunk)..((t + 1) * chunk) })
+      })
+      .collect::<Vec<_>>()
+  }
+}
+
+// Unsafe threaded version, it should be semantically equal to the single threaded version
+impl<T: 'static + ArrayType + Sync + Send + Sum, A: AllocatorType + Sync + Send> RustStream<T>
+  for StreamData<T, UnsafeDevice<T>, A>
+{
+  fn init_arrays(&mut self) {
+    self.device.data.a = UnsafeData::new(&mut self.a);
+    self.device.data.b = UnsafeData::new(&mut self.b);
+    self.device.data.c = UnsafeData::new(&mut self.c);
+    let init = self.init;
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let b = data.b.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            a[i] = init.0;
+            b[i] = init.1;
+            c[i] = init.2;
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn copy(&mut self) {
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            c[i] = a[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn mul(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let b = data.b.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            b[i] = scalar * c[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn add(&mut self) {
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let b = data.b.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            c[i] = a[i] + b[i];
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn triad(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let b = data.b.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            a[i] = b[i] + scalar * c[i]
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn nstream(&mut self) {
+    let scalar = self.scalar;
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let b = data.b.get_slice();
+          let c = data.c.get_slice();
+          for i in r {
+            a[i] += b[i] + scalar * c[i]
+          }
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .for_each(|t| t.join().unwrap());
+  }
+
+  fn dot(&mut self) -> T {
+    let pin = self.device.pin;
+    let data = self.device.data;
+    self
+      .device
+      .thread_ranges(self.size)
+      .into_iter()
+      .map(|(t, r)| {
+        let core = self.device.core_ids[t];
+        std::thread::spawn(move || {
+          if pin {
+            core_affinity::set_for_current(core);
+          }
+          let a = data.a.get_slice();
+          let b = data.b.get_slice();
+          let mut p = T::default();
+          for i in r {
+            p += a[i] * b[i];
+          }
+          p
+        })
+      })
+      .collect::<Vec<_>>()
+      .into_iter()
+      .map(|t| t.join().unwrap())
+      .sum()
+  }
+}

rust-stream/tests/integration_test.rs

@@ -2,7 +2,7 @@ use rstest::rstest;
 
 #[rstest]
 fn test_main(
-  #[values(0, 1, 2)] device: usize,                             //
+  #[values(0, 1, 2, 3, 4)] device: usize,                             //
   #[values("", "--pin")] pin: &str,                             //
   #[values("", "--malloc")] malloc: &str,                       //
   #[values("", "--init")] init: &str,                           //