mathieu/granite-rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

bench: Improve the spectralnorm shootout benchmark

Browse source 
This improves the spectralnorm shootout benchmark through a few vectors after
looking at the leading C implementation:

* The simd-based f64x2 is now used to parallelize a few computations
* RWLock usage has been removed. A custom `parallel` function was added as a
  form of stack-based fork-join parallelism. I found that the contention on the
  locks was high as well as hindering other optimizations.

This does, however, introduce one `unsafe` block into the benchmarks, which
previously had none.

In terms of timings, the before and after numbers are:

```
$ time ./shootout-spectralnorm-before
./shootout-spectralnorm-before  2.07s user 0.71s system 324% cpu 0.857 total
$ time ./shootout-spectralnorm-before 5500
./shootout-spectralnorm-before 5500  11.88s user 1.13s system 459% cpu 2.830 total
$ time ./shootout-spectralnorm-after
./shootout-spectralnorm-after  0.58s user 0.01s system 280% cpu 0.210 tota
$ time ./shootout-spectralnorm-after 5500
./shootout-spectralnorm-after 5500  3.55s user 0.01s system 455% cpu 0.783 total
```
This commit is contained in:
Alex Crichton 2014-10-12 17:25:51 -07:00
parent ff0abf05c9
commit f7b54703d0
1 changed files with 71 additions and 82 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

153
src/test/bench/shootout-spectralnorm.rs
View file

@ -41,105 +41,94 @@
// no-pretty-expanded FIXME #15189
#![allow(non_snake_case)]
#![feature(unboxed_closures, overloaded_calls)]
use std::from_str::FromStr;
use std::iter::count;
use std::cmp::min;
use std::iter::AdditiveIterator;
use std::mem;
use std::os;
use std::sync::{Arc, RWLock};
use std::raw::Repr;
use std::simd::f64x2;
fn main() {
let args = os::args();
let answer = spectralnorm(if os::getenv("RUST_BENCH").is_some() {
5500
} else if args.len() < 2 {
2000
} else {
from_str(args[1].as_slice()).unwrap()
});
println!("{:.9f}", answer);
}
fn spectralnorm(n: uint) -> f64 {
assert!(n % 2 == 0, "only even lengths are accepted");
let mut u = Vec::from_elem(n, 1.0);
let mut v = Vec::from_elem(n, 1.0);
let mut tmp = Vec::from_elem(n, 1.0);
for _ in range(0u, 10) {
mult_AtAv(u.as_slice(), v.as_mut_slice(), tmp.as_mut_slice());
mult_AtAv(v.as_slice(), u.as_mut_slice(), tmp.as_mut_slice());
}
(dot(u.as_slice(), v.as_slice()) / dot(v.as_slice(), v.as_slice())).sqrt()
}
fn mult_AtAv(v: &[f64], out: &mut [f64], tmp: &mut [f64]) {
mult_Av(v, tmp);
mult_Atv(tmp, out);
}
fn mult_Av(v: &[f64], out: &mut [f64]) {
parallel(out, |&: start, out| mult(v, out, start, |i, j| A(i, j)));
}
fn mult_Atv(v: &[f64], out: &mut [f64]) {
parallel(out, |&: start, out| mult(v, out, start, |i, j| A(j, i)));
}
fn mult(v: &[f64], out: &mut [f64], start: uint, a: |uint, uint| -> f64) {
for (i, slot) in out.iter_mut().enumerate().map(|(i, s)| (i + start, s)) {
let mut sum = f64x2(0.0, 0.0);
for (j, chunk) in v.chunks(2).enumerate().map(|(j, s)| (2 * j, s)) {
let top = f64x2(chunk[0], chunk[1]);
let bot = f64x2(a(i, j), a(i, j + 1));
sum += top / bot;
}
let f64x2(a, b) = sum;
*slot = a + b;
}
}
fn A(i: uint, j: uint) -> f64 {
((i + j) * (i + j + 1) / 2 + i + 1) as f64
}
fn dot(v: &[f64], u: &[f64]) -> f64 {
let mut sum = 0.0;
for (&v_i, &u_i) in v.iter().zip(u.iter()) {
sum += v_i * u_i;
}
sum
v.iter().zip(u.iter()).map(|(a, b)| *a * *b).sum()
}
fn mult(v: Arc<RWLock<Vec<f64>>>, out: Arc<RWLock<Vec<f64>>>,
f: fn(&Vec<f64>, uint) -> f64) {
// We launch in different tasks the work to be done. To finish
// this function, we need to wait for the completion of every
// tasks. To do that, we give to each tasks a wait_chan that we
// drop at the end of the work. At the end of this function, we
// wait until the channel hang up.
// Executes a closure in parallel over the given mutable slice. The closure `f`
// is run in parallel and yielded the starting index within `v` as well as a
// sub-slice of `v`.
fn parallel<'a, T, F>(v: &'a mut [T], f: F)
where T: Send + Sync,
F: Fn(uint, &'a mut [T]) + Sync {
let (tx, rx) = channel();
let size = v.len() / os::num_cpus() + 1;
let len = out.read().len();
let chunk = len / 20 + 1;
for chk in count(0, chunk) {
if chk >= len {break;}
for (i, chunk) in v.chunks_mut(size).enumerate() {
let tx = tx.clone();
let v = v.clone();
let out = out.clone();
// Need to convert `f` and `chunk` to something that can cross the task
// boundary.
let f = &f as *const _ as *const uint;
let raw = chunk.repr();
spawn(proc() {
for i in range(chk, min(len, chk + chunk)) {
let val = f(&*v.read(), i);
*out.write().get_mut(i) = val;
}
let f = f as *const F;
unsafe { (*f)(i * size, mem::transmute(raw)) }
drop(tx)
});
}
// wait until the channel hang up (every task finished)
drop(tx);
for () in rx.iter() {}
}
fn mult_Av_impl(v: &Vec<f64> , i: uint) -> f64 {
let mut sum = 0.;
for (j, &v_j) in v.iter().enumerate() {
sum += v_j / A(i, j);
}
sum
}
fn mult_Av(v: Arc<RWLock<Vec<f64>>>, out: Arc<RWLock<Vec<f64>>>) {
mult(v, out, mult_Av_impl);
}
fn mult_Atv_impl(v: &Vec<f64> , i: uint) -> f64 {
let mut sum = 0.;
for (j, &v_j) in v.iter().enumerate() {
sum += v_j / A(j, i);
}
sum
}
fn mult_Atv(v: Arc<RWLock<Vec<f64>>>, out: Arc<RWLock<Vec<f64>>>) {
mult(v, out, mult_Atv_impl);
}
fn mult_AtAv(v: Arc<RWLock<Vec<f64>>>, out: Arc<RWLock<Vec<f64>>>,
tmp: Arc<RWLock<Vec<f64>>>) {
mult_Av(v, tmp.clone());
mult_Atv(tmp, out);
}
fn main() {
let args = os::args();
let args = args.as_slice();
let n = if os::getenv("RUST_BENCH").is_some() {
5500
} else if args.len() < 2 {
2000
} else {
FromStr::from_str(args[1].as_slice()).unwrap()
};
let u = Arc::new(RWLock::new(Vec::from_elem(n, 1f64)));
let v = Arc::new(RWLock::new(Vec::from_elem(n, 1f64)));
let tmp = Arc::new(RWLock::new(Vec::from_elem(n, 1f64)));
for _ in range(0u8, 10) {
mult_AtAv(u.clone(), v.clone(), tmp.clone());
mult_AtAv(v.clone(), u.clone(), tmp.clone());
}
let u = u.read();
let v = v.read();
println!("{:.9f}", (dot(u.as_slice(), v.as_slice()) /
dot(v.as_slice(), v.as_slice())).sqrt());
}