Skip to content

armgit / ente

Public Imported from GitHub
0

πŸ’š End-to-end encrypted cloud for everything.

[rust] Use one threshold for elementwise parallelism

laurenspriemed6ff8a8ce46
diff --git a/rust/crates/ml/src/cv/bilateral.rs b/rust/crates/ml/src/cv/bilateral.rs
index 68855c5c62..64dead5a9b 100644
--- a/rust/crates/ml/src/cv/bilateral.rs
+++ b/rust/crates/ml/src/cv/bilateral.rs
@@ -1,8 +1,8 @@
use rayon::prelude::*;
use super::reflect101;
-use crate::cv::OpResult;
use crate::cv::image::ImageU8;
+use crate::cv::{OpResult, PARALLEL_MIN_ELEMS};
pub(crate) fn bilateral_filter_u8(
src: &ImageU8,
@@ -59,23 +59,30 @@ pub(crate) fn bilateral_filter_u8(
}
let mut out = vec![0u8; w * h];
- out.par_chunks_exact_mut(w)
- .enumerate()
- .for_each(|(y, dst)| {
- let win = &temp[y * stride..];
- let centre = radius * stride + radius;
- for (x, d) in dst.iter_mut().enumerate() {
- let mid = win[centre + x] as i32;
- let mut sum = 0.0f32;
- let mut wsum = 0.0f32;
- for &(off, sw) in &taps {
- let val = win[off + x];
- let weight = sw * color_weight[(val as i32 - mid).unsigned_abs() as usize];
- wsum += weight;
- sum += val as f32 * weight;
- }
- *d = (sum / wsum).round_ties_even() as u8;
+ let row = |y: usize, dst: &mut [u8]| {
+ let win = &temp[y * stride..];
+ let centre = radius * stride + radius;
+ for (x, d) in dst.iter_mut().enumerate() {
+ let mid = win[centre + x] as i32;
+ let mut sum = 0.0f32;
+ let mut wsum = 0.0f32;
+ for &(off, sw) in &taps {
+ let val = win[off + x];
+ let weight = sw * color_weight[(val as i32 - mid).unsigned_abs() as usize];
+ wsum += weight;
+ sum += val as f32 * weight;
}
- });
+ *d = (sum / wsum).round_ties_even() as u8;
+ }
+ };
+ if w * h >= PARALLEL_MIN_ELEMS {
+ out.par_chunks_exact_mut(w)
+ .enumerate()
+ .for_each(|(y, dst)| row(y, dst));
+ } else {
+ for (y, dst) in out.chunks_exact_mut(w).enumerate() {
+ row(y, dst);
+ }
+ }
ImageU8::new(src.width, src.height, 1, out)
}
diff --git a/rust/crates/ml/src/cv/image.rs b/rust/crates/ml/src/cv/image.rs
index 08ab5db7f6..9d8797222e 100644
--- a/rust/crates/ml/src/cv/image.rs
+++ b/rust/crates/ml/src/cv/image.rs
@@ -1,9 +1,6 @@
use rayon::prelude::*;
-use super::{OpResult, saturate_u8_f32};
-
-/// Elementwise work below this many samples is not worth splitting across threads.
-const PARALLEL_MIN_ELEMS: usize = 200_000;
+use super::{OpResult, PARALLEL_MIN_ELEMS, saturate_u8_f32};
#[derive(Clone, Debug, PartialEq, Eq)]
pub(crate) struct ImageU8 {
diff --git a/rust/crates/ml/src/cv/mod.rs b/rust/crates/ml/src/cv/mod.rs
index f77c38cc37..ca329a6d3b 100644
--- a/rust/crates/ml/src/cv/mod.rs
+++ b/rust/crates/ml/src/cv/mod.rs
@@ -37,7 +37,9 @@ pub(crate) use structuring::ellipse_kernel;
pub(crate) use transform::rotate_u8;
pub(crate) use warp::warp_perspective;
-const PARALLEL_MIN_PIXELS: usize = 65_536;
+/// One answer for every elementwise op: below this many samples, threading
+/// costs more than it saves.
+pub(crate) const PARALLEL_MIN_ELEMS: usize = 200_000;
const PIXELS_PER_CHUNK: usize = 65_536;
pub(crate) fn pointwise<T: Sync, U: Send>(
@@ -47,7 +49,7 @@ pub(crate) fn pointwise<T: Sync, U: Send>(
src_stride: usize,
f: impl Fn(&mut [U], &[T]) + Send + Sync,
) {
- if out.len() / out_stride.max(1) < PARALLEL_MIN_PIXELS {
+ if out.len() < PARALLEL_MIN_ELEMS {
f(out, src);
return;
}
@@ -65,7 +67,7 @@ pub(crate) fn pointwise3<T: Sync, V: Sync, U: Send>(
b_stride: usize,
f: impl Fn(&mut [U], &[T], &[V]) + Send + Sync,
) {
- if out.len() / out_stride.max(1) < PARALLEL_MIN_PIXELS {
+ if out.len() < PARALLEL_MIN_ELEMS {
f(out, a, b);
return;
}
diff --git a/rust/crates/ml/src/cv/morph.rs b/rust/crates/ml/src/cv/morph.rs
index 0560dac93a..c4809d34d9 100644
--- a/rust/crates/ml/src/cv/morph.rs
+++ b/rust/crates/ml/src/cv/morph.rs
@@ -1,9 +1,7 @@
use rayon::prelude::*;
-use crate::cv::OpResult;
use crate::cv::image::ImageU8;
-
-const PARALLEL_MIN_PIXELS: usize = 200_000;
+use crate::cv::{OpResult, PARALLEL_MIN_ELEMS};
fn morph(src: &ImageU8, kernel: &ImageU8, erode: bool) -> OpResult<ImageU8> {
if src.channels != 1 || kernel.channels != 1 {
@@ -48,7 +46,7 @@ fn morph(src: &ImageU8, kernel: &ImageU8, erode: bool) -> OpResult<ImageU8> {
}
}
};
- if w * h >= PARALLEL_MIN_PIXELS {
+ if w * h >= PARALLEL_MIN_ELEMS {
out.par_chunks_mut(w)
.enumerate()
.for_each(|(y, dst)| row(y, dst));