Add basic SIMD support #523
Conversation
src/distributions/float.rs
Outdated
| // those are usually more random. | ||
| let float_size = mem::size_of::<$f_scalar>() * 8; | ||
| let precision = $fraction_bits + 1; | ||
| let scale = $ty::splat(1.0 / ((1 as $u_scalar << precision) as $f_scalar)); |
There was a problem hiding this comment.
Note: most arithmetic and bitwise operations allow you to just use scalar values:
let scale = $ty::splat(1.0 / ((1 as $u_scalar << precision) as $f_scalar));
scale * value
// is the same as
let scale = 1.0 / ((1 as $u_scalar << precision) as $f_scalar);
scale * value
src/distributions/float.rs
Outdated
|
|
||
| let value: $uty = rng.gen(); | ||
| let fraction = value >> (float_size - $fraction_bits); | ||
| fraction.into_float_with_exponent(0) - $ty::splat(1.0 - EPSILON / 2.0) |
There was a problem hiding this comment.
Same here. This would be the same:
fraction.into_float_with_exponent(0) - (1.0 - EPSILON / 2.0)
assert!(low < high, "Uniform::new called with `low >= high`");
This will be fine on most inputs, but for inputs like The assertions need to be of the form |
O wow, that is something easy to get wrong! Thank you for reviewing. I have added a
|
pitdicker
force-pushed
the
simd_support_basic
branch
from
e6e7734 to
4dab1e3
Compare
|
The CI fails because of rust-lang/rust#51699. |
src/distributions/mod.rs
Outdated
| @@ -214,6 +214,7 @@ mod float; | |||
| mod integer; | |||
| #[cfg(feature="std")] | |||
| mod log_gamma; | |||
| mod math_helpers; | |||
There was a problem hiding this comment.
Is this any more mathematical than other parts of the library? You could just call it num_utils (i.e. Number Theory, though arguably that still covers most of the library) or arithmetic.
src/distributions/math_helpers.rs
Outdated
|
|
||
| /// `PartialOrd` for vectors compares lexicographically. We want natural order. | ||
| /// Only the comparison functions we need are implemented. | ||
| pub trait NaturalCompare { |
There was a problem hiding this comment.
What is natural depends on the context which is how PartialOrd confused you. How about SimultaneousOrd?
|
@TheIronBorn do you think this is ready? I'd like to go over it in more detail myself still, but I think it's pretty neat that we can do this (it's well beyond what most random libs can offer)! |
|
Looks good! I'll make an SIMD |
There was a problem hiding this comment.
But other than those two things looks good.
BTW @TheIronBorn you can make a commit on top of @pitdicker's branch and push to your own branch if you have a fix, but best coordinate between yourselves.
| unsafe { | ||
| let ptr = &mut vec; | ||
| let b_ptr = &mut *(ptr as *mut $ty as *mut [u8; $bits/8]); | ||
| rng.fill_bytes(b_ptr); |
There was a problem hiding this comment.
This doesn't look portable to me. Elsewhere we've made an effort to keep things portable; I don't think this needs to be an exception?
Unfortunately it doesn't look like the SIMD types support to_le. @TheIronBorn is this what you mean about using swap_bytes?
There was a problem hiding this comment.
Yes that’s exactly where we’d use it
| macro_rules! simd_impl { | ||
| ($bits:expr,) => {}; | ||
| ($bits:expr, $ty:ty, $($ty_more:ty,)*) => { | ||
| simd_impl!($bits, $($ty_more,)*); |
There was a problem hiding this comment.
Neat usage of recursive macros.
But why do we need to pass $bits here instead of just using mem::size_of? It seems like an unnecessary risk of underfill/overfill.
|
Since @TheIronBorn has already addressed my comments, lets merge this and make a second PR for the fixes. |
This is a rebased version of the branch from #377 (comment).
@TheIronBorn Do you want to review?