What the pipeline last published: which cacheline, and what shape its pixels are in. More...

#include <pixelpipe_hb.h>

Collaboration diagram for dt_backbuf_t:

Data Fields
size_t	bpp

size_t	width

size_t	height

dt_atomic_uint64	hash

dt_atomic_uint64	history_hash

dt_atomic_uint64	generation

Detailed Description

What the pipeline last published: which cacheline, and what shape its pixels are in.

this encapsulates the pixelpipe. a develop module will need several of these: for previews and full blits to cairo and for the export function.

These five fields are ONE fact and must be read as one. ‘hash’ names a cacheline and ‘width’/‘height’ say what shape the pixels in it are – a cacheline records how many BYTES it holds and nothing about their layout, so a consumer that pairs a hash from one publication with dimensions from the next has no way to notice. It computes a cairo stride from the wrong width and paints the diagonal striping of a stride error, or writes a mis-shaped thumbnail into the mipmap cache. The published entry is routinely LARGER than width x height x bpp (aligned allocation, pool reuse), so a size check does not catch it either.

That was reachable: ‘hash’ was atomic and the dimensions were plain fields, so a GUI thread could read the hash, resolve its entry, and then read dimensions the worker had republished in between. It needs a publication whose SHAPE changes to become visible, which is why entering a clipping module's edit mode – where the crop is neutralised and the frame jumps to the full transformed image – is where it was reported.

So the record carries a seqlock, exactly as dt_dev_geometry_store_t does (develop/dev_geometry.h), and dt_dev_backbuf_snapshot is how a consumer reads it. Do not read the fields directly to pair them with pixels; a bare read is fine only for something that pairs with nothing.

WRITERS ARE SERIALISED BY CONVENTION, not by a lock: each backbuffer belongs to one pipe and is published by the thread that runs it, with the view's leave() joining that thread before it touches anything. A second concurrent publisher needs a real lock, not a second odd counter.

Definition at line 218 of file pixelpipe_hb.h.