Added abs_diff for integer types.

[orlp]

Closes #62111.

[rust-highfive]

r? @m-ou-se

(rust-highfive has picked a reviewer for you, use r? to override)

[the8472]

Does either of these compile to a psadbw (or equivalent SSE/AVX instructions for wider arrays) on x86?

fn sad_iter(a: &[u8; 8], b: &[u8; 8]) -> u32 {
  a.iter().zip(b).map(|&a, &b| a.abs_diff(b) as u32).sum()
}

fn sad_loop(a: &[u8; 8], b: &[u8; 8]) -> u32 {
  let mut sum = 0;
  for i in 0..a.len() {
      sum += a[i].abs_diff(b[i]) as u32;
  }
  sum
}

llvm does have some pattern detection to generate psadbw, so we should make sure we fall into that.

[orlp]

@the8472 I can't get Rust to generate psadbw without intrinsics at all. I'm not sure if your linked thread was actually merged into LLVM or is still in the code.

[the8472]

Yes, it's still there. Based on comments in tickets the pattern it expects for absolute difference should look like abs(a - b), it must be in a loop that can be vectorized and the aggregation variable must be 32bit or 64bit wide.

https://github.com/llvm/llvm-project/blob/main/llvm/test/CodeGen/X86/sad.ll
https://raw.githubusercontent.com/llvm/llvm-project/main/llvm/lib/Target/X86/X86ISelLowering.cpp

[thomcc]

It's a really fragile optimization. I've never seen it in practice, but: https://godbolt.org/z/qffbWvh4v

[orlp]

@the8472 I've tried a bunch of things, and I can't generate it.

Note that abs(a-b) is not valid at all unless you widen the type first (which I've also tried without luck).

EDIT: it appears to only work for exactly this:

pub fn abs_diff(slf: u8, other: u8)  -> u8 {
    (slf as i32).wrapping_sub(other as i32).abs() as u8
}

If you widen to anything except i32 it doesn't work at all.

[the8472]

It's a really fragile optimization.

I suspect it's written to match C's integer conversion rules.

[orlp]

If we consider this worthwhile I could add that implementation for u8 specifically. It does mean I have to pull out from the uint_macros and create a macro myself to add the current implementation for everything except u8.

EDIT: or I could do a little hack using if mem::size_of::<Self>() == 1 and see if it affects codegen at all.

EDIT2: appears to work to generate the psad instruction.

[the8472]

SAD is really important in video processing. Yes, all serious implementations will use platform-specific intrinsics but they have generic fallback paths and if those get used for any reason (wasm? riscv?) it should be in a pattern that optimizers recognize.

[orlp]

@the8472 Implemented psadbw support.

[m-ou-se]

This looks good to me.

Can you open a tracking issue, and add its number in the #[unstable] attributes? Thanks!

[orlp]

@m-ou-se Done: #89492.

[m-ou-se]

Thanks!

@bors r+

[bors]

📌 Commit 6dd6e7c has been approved by m-ou-se