UPDATE : question is about Camera2
I'm trying to figure out who applies the rotation transform when a camera preview is drawn on a SurfaceTexture.
When requesting the preview sizes from the camera you always get pairs where the width is larger than the height (because landscape is the usual orientation when taking a picture).
When using the device on portrait mode and setting the preview size (for ex. 1600 x1200) the frames from the camera are correctly rotated but I can't find out where it's done, is it something CameraDevice does automatically based on Surface type ? or is it the SurfaceTexture who rotates the image?
The answer depends a bit on which camera API you're using; since you mention CameraDevice, I assume you mean camera2.
In camera2, one of the goals was to simplify the rotation handling where possible; so if you draw preview into a SurfaceView, the camera service and the hardware compositor / GPU cooperate and handle all the rotation for you. You always request landscape resolutions from the sensor (which is aligned with the long edge of the sensor matching the long edge of the device), and you just need to make sure your SurfaceView's aspect ratio matches that of the requested resolution (so in portrait, you need 9:16 View to display 1080p correctly; in landscape you need a 16:9 View).
For TextureView, and completely custom GL drawing with SurfaceTexture, the API can't do it all for you. The SurfaceTexture's transform is set by the camera API to rotate the camera output to the device's native orientation (generally portrait for phones, landscape for some tablets), but it can't handle the additional rotation if your app's UI is not in the native orientation.
In that case, you'll need to get the transform matrix from the SurfaceTexure, apply the extra rotation to match your app UI, and set that as the TextureView matrix.