The problem occurs when the calculation could lead to a final row (or
column) of grid squares with no pixels in them (and hence, NaNs).
One specific case is a Pi 5 with an image width (or height) of 1364,
so that's 682 Bayer quads. To give 32 grid squares it was calculating
22 quads per cell. However, 31 * 22 = 682 leaving nothing in the final
column.
The fix is to do a rounding-down division by the number of cells minus
one, rather than a rounding-up division by the number of cells. This
turns the corner case from one where the final row/column has no
pixels to one where we don't quite cover the full image, which is how
we have to handle these cases.
Bug: https://github.com/raspberrypi/libcamera/issues/254
Signed-off-by: David Plowman <david.plowman@raspberrypi.com>
Reviewed-by: Naushir Patuck <naush@raspberrypi.com>
Acked-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
Signed-off-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
36ba0e5515