3 ms·

Based on the "We can get faster still by trusting the compiler" part of the article, the author's using Rust. It doesn't have a global flag like `-funsafe-math-

Based on the "We can get faster still by trusting the compiler" part of the article, the author's using Rust. It doesn't have a global flag like `-funsafe-math-optimizations` or `-ffast-math` so the change would have to be a bit more involved. They'd have to change their use of `+`, `*` etc operators on f32 to `std::intrinsics::fadd_fast`, `std::intrinsics::fmul_fast`, etc functions. So, putting it into https://rust.godbolt.org/z/jT6Mb1K13 https://rust.godbolt.org/z/jT6Mb1K13 , it seems to indeed be using the SIMD instructions.

tialaramex 4y ago
They're asking for sum() on a slice f32s. The sum() function actually works via a Trait for this specific purpose, Sum, so you could go like this... New Type wrapper for f32 called like FastFloat, marked #[repr(transparent)], and if necessary (not sure) have the compiler promise you you're getting the same in memory representation as an actual f32. Implement Sum over FastFloat by having it use the faster SIMD intrinsics for this work to give you an answer, accepting the potential loss of accuracy. Now, unsafely transmute the f32 slice into a FastFloat slice (in principle this is zero instructions, it just satisfies the type checking) and ordinary sum() goes real fast because it's now Sum on the slice of FastFloats.
- Arnavion 4y ago
  If you want to go the newtype + Sum impl route, you don't have to make it `#[repr(transparent)]` or transmute the slice. You can just `impl Sum<FastFloat> for f32` and do `f.iter().copied().map(FastFloat).sum()` https://rust.godbolt.org/z/b9s3dna6r https://rust.godbolt.org/z/b9s3dna6r
  - tialaramex 4y ago
    Oh, I didn't think of that, clever. EtA: The attraction of a New Type plus trait impl is that is re-usable. You could imagine (particularly if it was stable which your approach isn't yet) packaging up several speed-ups like this in a crate, enabling people to get faster arithmetic where they can afford any accuracy trade off without them needing to know anything about SIMD and without (like the C or C++ compiler flags) affecting unrelated code where accuracy may be critical.
    - Arnavion 4y ago
      https://crates.io/crates/fast-floats https://crates.io/crates/fast-floats
      - tialaramex 4y ago
        Nice, although I notice it doesn't implement Sum or Product :D

tialaramex 4y ago

They're asking for sum() on a slice f32s. The sum() function actually works via a Trait for this specific purpose, Sum, so you could go like this... New Type wrapper for f32 called like FastFloat, marked #[repr(transparent)], and if necessary (not sure) have the compiler promise you you're getting the same in memory representation as an actual f32. Implement Sum over FastFloat by having it use the faster SIMD intrinsics for this work to give you an answer, accepting the potential loss of accuracy. Now, unsafely transmute the f32 slice into a FastFloat slice (in principle this is zero instructions, it just satisfies the type checking) and ordinary sum() goes real fast because it's now Sum on the slice of FastFloats.

Arnavion 4y ago
If you want to go the newtype + Sum impl route, you don't have to make it `#[repr(transparent)]` or transmute the slice. You can just `impl Sum<FastFloat> for f32` and do `f.iter().copied().map(FastFloat).sum()` https://rust.godbolt.org/z/b9s3dna6r https://rust.godbolt.org/z/b9s3dna6r
- tialaramex 4y ago
  Oh, I didn't think of that, clever. EtA: The attraction of a New Type plus trait impl is that is re-usable. You could imagine (particularly if it was stable which your approach isn't yet) packaging up several speed-ups like this in a crate, enabling people to get faster arithmetic where they can afford any accuracy trade off without them needing to know anything about SIMD and without (like the C or C++ compiler flags) affecting unrelated code where accuracy may be critical.
  - Arnavion 4y ago
    https://crates.io/crates/fast-floats https://crates.io/crates/fast-floats
    - tialaramex 4y ago
      Nice, although I notice it doesn't implement Sum or Product :D