diff --git a/CHANGELOG.md b/CHANGELOG.md
index 588db17..4e5271b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,12 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Added
+
+- distribute_unit function
+
 ## [0.1.0] - 2025-02-24
 
 ### Changed
diff --git a/src/lib.rs b/src/lib.rs
index 7e84fb8..d3fe851 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -124,6 +124,46 @@ impl Rearranging {
         Ok(ans)
     }
 
+    /// 从候选值中选择一个能整除当前单元规模的值作为新的单元规模
+    pub fn distribute_unit(&self, candidates: impl IntoIterator<Item = usize>) -> Option<Self> {
+        let unit = candidates.into_iter().find(|n| self.unit() % n == 0)?;
+        if unit == self.unit() {
+            return Some(self.clone());
+        }
+
+        let ndim = self.ndim();
+        let mut layout = vec![0isize; 4 + (ndim + 1) * 3].into_boxed_slice();
+        layout[0] = unit as _;
+        layout[1] = self.dst_offset();
+        layout[2] = self.src_offset();
+
+        let (_, tail) = layout.split_at_mut(3);
+        let (idx, tail) = tail.split_at_mut(ndim + 2);
+        let (dst, src) = tail.split_at_mut(ndim + 1);
+
+        let (_, tail) = self.0.split_at(3);
+        let (idx_, tail) = tail.split_at(ndim + 1);
+        let (dst_, src_) = tail.split_at(ndim);
+
+        idx[ndim + 1] = 1;
+        let extra = (self.unit() / unit) as isize;
+        for (new, old) in izip!(idx, idx_) {
+            *new = *old * extra;
+        }
+
+        fn copy_value(new: &mut [isize], old: &[isize], unit: usize) {
+            let [head @ .., tail] = new else {
+                unreachable!()
+            };
+            head.copy_from_slice(old);
+            *tail = unit as _;
+        }
+        copy_value(dst, dst_, unit);
+        copy_value(src, src_, unit);
+
+        Some(Self(layout))
+    }
+
     /// 执行方案维数。
     #[inline]
     pub fn ndim(&self) -> usize {
@@ -233,3 +273,61 @@ fn test_scheme() {
     assert_eq!(scheme.src_strides(), [96, 8, 16]);
     assert_eq!(scheme.shape().collect::<Vec<_>>(), [24, 2, 3]);
 }
+
+#[test]
+fn test_distribute_unit() {
+    // 创建一个测试用的重排方案
+    let shape = [4, 3, 2];
+    let dst = [24, 8, 2];
+    let src = [48, 8, 16];
+    let dst = ArrayLayout::<3>::new(&shape, &dst, 0);
+    let src = ArrayLayout::<3>::new(&shape, &src, 0);
+    let scheme = Rearranging::new(&dst, &src, 2).unwrap();
+
+    // 测试1: 使用相同的单元大小
+    let candidates = vec![2];
+    let new_scheme = scheme.distribute_unit(candidates).unwrap();
+    assert_eq!(new_scheme.unit(), 2);
+    assert_eq!(new_scheme.count(), scheme.count());
+    assert_eq!(new_scheme.idx_strides(), scheme.idx_strides());
+    assert_eq!(new_scheme.dst_strides(), scheme.dst_strides());
+    assert_eq!(new_scheme.src_strides(), scheme.src_strides());
+
+    // 测试2: 使用更小的单元大小
+    let candidates = vec![1];
+    let new_scheme = scheme.distribute_unit(candidates).unwrap();
+    assert_eq!(new_scheme.unit(), 1);
+    assert_eq!(new_scheme.count(), scheme.count() * 2);
+    assert_eq!(
+        new_scheme
+            .idx_strides()
+            .iter()
+            .take(scheme.idx_strides().len())
+            .map(|&x| x / 2)
+            .collect::<Vec<_>>(),
+        scheme.idx_strides()
+    );
+    assert_eq!(
+        new_scheme
+            .dst_strides()
+            .iter()
+            .take(scheme.idx_strides().len())
+            .cloned()
+            .collect::<Vec<_>>(),
+        scheme.dst_strides()
+    );
+    assert_eq!(
+        new_scheme
+            .src_strides()
+            .iter()
+            .take(scheme.idx_strides().len())
+            .cloned()
+            .collect::<Vec<_>>(),
+        scheme.src_strides()
+    );
+
+    // 测试3: 使用多个候选值
+    let candidates = vec![4, 2, 1];
+    let new_scheme = scheme.distribute_unit(candidates).unwrap();
+    assert_eq!(new_scheme.unit(), 2); // 应该选择第一个能整除的值
+}