which obtains a view into an image, and then calls to_image() on it to turn it into a new image buffer. And to_image() shuffles pixels one by one in a purely scalar fashion and with bounds checks on every pixel:
A much better way would be iterating over the rows, and copying over entire rows with slice::copy_from_slice(). That would both leverage vector instructions and reduce the amount of bounds checks from n^2 to n.
jamninetyfive
commented
2 weeks ago
In v0.25.2, the crop operation is implemented as:
https://github.com/image-rs/image/blob/e176cd414ac6cc73909ff162e5d0b677f9d4fb08/src/dynimage.rs#L422
which obtains a view into an image, and then calls
to_image()on it to turn it into a new image buffer. Andto_image()shuffles pixels one by one in a purely scalar fashion and with bounds checks on every pixel:https://github.com/image-rs/image/blob/e176cd414ac6cc73909ff162e5d0b677f9d4fb08/src/image.rs#L1131-L1136
A much better way would be iterating over the rows, and copying over entire rows with
slice::copy_from_slice(). That would both leverage vector instructions and reduce the amount of bounds checks from n^2 to n.