pub fn replay_map_raw_fused<D, S, F>(
dest: &mut RawStridedMut<'_, D>,
src: &RawStridedRef<'_, S>,
f: F,
exec: Exec<'_>,
) -> Result<()>Expand description
f(&mut dest[i], src[i]) after fusing the two layouts inside the call.
replay_map_raw replays the axes exactly as given, so a caller that holds
an unfused layout has to fuse it first and pays that pass per call. This
entry performs the same fusion copy_scale_raw does before choosing the
route, then replays with the caller’s token; the small-block fusion cost is
therefore inside the crate, once, instead of at every call site.
The traversal, width, callback and allocation contracts are those of
replay_map_raw: the visit order is unspecified, and a successful call
allocates nothing at any rank on a serial token, or on a pooled token whose
replay is not tiled above RAW_FUSED_RANK_LIMIT moving axes. The fusion
itself holds its scratch on the stack, so it does not allocate either.
§Errors
As replay_map_raw.
§Examples
use strided_basic::{replay_map_raw_fused, Exec, RawStridedMut, RawStridedRef};
let src = [1.0f64, 2.0, 3.0, 4.0];
let mut acc = [10.0f64, 20.0, 30.0, 40.0];
let source = RawStridedRef::new(&src, &[2, 2], &[2, 1], 0).unwrap();
// The destination is transposed relative to the source; the fused entry
// fuses the two layouts before it walks them.
let mut dst = RawStridedMut::new(&mut acc, &[2, 2], &[1, 2], 0).unwrap();
replay_map_raw_fused(&mut dst, &source, |d, s| *d += s, Exec::serial()).unwrap();
assert_eq!(acc, [11.0, 23.0, 32.0, 44.0]);