gimp_display_shell_render() writes to a GeglBuffer backed by allocated memory
(shell->profile_data). Unfortunately while converting prevision in
gimp_image_convert_precision(), we change the "precision" property (hence the
source format) first, hence end up trying to write data in a too small buffer.
This crash was hard to find as it was not showing up on my machine (though it
did produce rendering artifacts!), unless I built both GIMP and babl with
`b_sanitize=address`.
Note that an alternate fix was to make sure that the profile_data buffer is big
enough (by calling gimp_display_shell_profile_update() before rendering), but
anyway the image is in an inconsistent state while conversion is in progress:
whereas the `src_format` is the new one, the `src_profile` is still the old one
(and cannot be changed before we finish converting).
Moreover the render happen regularly on progress signals, once after each
converted drawable. So each of these rendering step happens in an inconsistent
state, with the wrong profile set, some of the drawables converted and others
not yet.
We could still render properly if each drawable's buffer used space-aware format
(thus allowing different drawables to use different profiles/spaces), but it
feels over-engineering the problem. It might be much better to ignore rendering
steps while converting the image precision. Moreover it would obviously make a
faster conversion.
See discussions in #9136 for this crash, which didn't have dedicated report
AFAIK.
(cherry picked from commit de25be9210)
Note: on the `master` branch, even with sanitized code, I don't get the crash.
Yet this change seems relevant enough that I'm adding it.
In GimpDisplayShell, scale the render cache by the window's scale
factor, and render its content in device pixels, instead of scaled
application pixels. When painting the cache to the screen, unscale
the cairo context by the same factor, so that it's painted in the
native resolution. Note that the various
gimp_display_shell_render_foo() functions still speak in
application pixels, and the scaling happens internally in
gimp_display_shell_render().
Aside from rendering at native resolution on HiDPI, this also fixes
an issue where grid-like display artifacts would appear when the
render cache was not fully validated due to the non-native scaling.
Add a "show all" mode to GimpDisplayShell, controlled through a
corresponding "View -> Show All" menu item. When enabled, the
entire image content is displayed, instead of cropping the image
to the canvas size. More generally, the display behaves as if the
canvas were infinite. The following commits improve the overall
behavior in this mode.
Add a prefernces option to control the default "show all" state.
When set to FAST we do nearest neighbor from the next bigger mipmap
level instead of linear or box filtering - this gives a slight and
permanent boost to painting, and all updates, having one that combines
best of both worlds and reblits in high quality after a timeout would
be even more desirable.
Introduce a render cache that keeps the result of scaling, color
management, display filters and shell mask (for tools like fuzzy
select).
Change gimpdisplayshell-render.[ch] to only render to the cache and
manage a cairo region of the cache's valid area. Call cache
invalidation functions form various places. Change the API of all
render functions to be in display coordinates.
Also get rid of gimpdisplayxfer.[ch] because we now have a
canvas-sized cairo surface which is a surface similar to the
destination surface.
As suggested by massimo in bug #694917, move unconditional creation/destruction
of a wrapper GeglBuffer object from top-level scope of the function to the
single conditional scope where it is used.
This reverts commit 36258a671a.
This commit was making the rotated canvas rendering quite horrible to
the point that I think it would make the canvas rotation feature barely
usable. See bug 759287, comments 13 to 18.
I think we will need to find other ways to accelerate rendering.
Compromise on quality is possible, but I think that in this case, this
was more than just a compromise. It was more like completely abandonning
quality. We could even see the lines "spiking" while you were rotating!
Like your drawing was alive!
Use CAIRO_FILTER_FAST when painting the xfer surface to the
screen. This notably improves performance when the canvas is
rotated, at the cost of lower filtering quality.
Based on a patch by Massimo.
Move the entire image-space/screen-space transformation logic from
gimp_display_shell_render() to gimp_display_shell_draw_image(), so
that the former works entirely in image space, and do the chunking
and clipping in screen-space, making sure that image-space chunks
are never larger than
GIMP_DISPLAY_RENDER_BUF_WIDTH x GIMP_DISPLAY_RENDER_BUF_HEIGHT,
even when the window's scale factor is greater than 1.
Add a GIMP_BRICK_WALL environment variable, which, when set, shows
the screen-space chunk bounds.
When we have display filters, break the color profile transform in
two: first, convert from the image profile to sRGB, then apply the
filters, then convert from sRGB to the monitor profile.
When processing display filters, shift the filter buffer to the
top-left corner of the render area, and pass the actual render
area, instead of an area whose top-left coords are (0, 0), to the
display filter. This allows for position-dependent display
filters.
which encapsulates a cmsHTRANSFORM and does all the pixel format
conversion magic. It has API to create transforms and proofing
transforms, and to convert pixels arrays and GeglBuffers.
Before, each place which has a transform had to keep around the
transform and its input and output Babl formats, and had to implement
lots of stuff itself. Now all that lives in GimpColorTransform,
removing lots of logic from many places, and pretty much removing lcms
from the public API entirely.
This removes including <lcms2.h>, LCMS_LIBS and LCMS_CFLAGS from
almost all directories and potentially allows to replace lcms by
something else.
but they aren't clipped
Add gimp_display_shell_profile_can_convert_to_u8() which returns
whether the lcms transform can safely write directly into an u8 buffer
without destroying out-of-gammut pixel values, which we assume is the
case for all integer formats. If the function returns FALSE, always
convert via the R'G'B'A float filter_buffer.
Also connect to the image's "precision-changed" signal and update the
profile transform when it's emitted.
- disable auto-adding of the lcms display filter module
- change profile convert dest formats to be always R'G'B'A, a display
profile transform outputs something that can be displayed directly,
so no additional gamma transform must happen when the pixels are
copied to a cairo-ARGB32 buffer
- add a medium forest of if() branches to gimpdisplayshell-filter.c
which cover all combinations of profile and display filter
transforms
- all of this is still very broken when changing an image to linear,
because the configured RGB profile from prefs will do horrible
nonsense (things work fine though with a per-image profile that is
for linear data)
Enhance the existing but unused display scaling (hidpi/retina) support
to work independently in x and y direction, and adjust the scaling
factors accordingly when dot-for-dot is off and xres != yres.
Increase GIMP_DISPLAY_RENDER_MAX_SCALE from 2.0 to 4.0 and adjust the
rendering chunk size dynamically so we never render chunks that do
not fit into the GimpDisplayXfer buffers.
This removes the obsolete check which makes the tool fail from
gimp_display_shell_set_mask(). Also change the foreground select tool
and the display mask from using GimpChannel to GeglBuffer, because
that's what it needs, simply buffers. Most changed files simply newly
include <gegl.h> because a GeglBuffer appeared in two headers.
- stroke around the drawn area to work around the impossibility of
tiling antialiased stuff.
- don't use the display xfer stuff because if we overdraw we must
not have anything in the source pattern outside the drawn region
Recent Cairo uses SHM transports when available, and exposes the ability
for its users to manage images shared between it and the display.
This allows us to eliminate copies, and if the architecture supports it
even to upload directly into GPU addressable memory without any copies
(all in normal system memory so we suffer no performance penalty when
applying the filters). The caveat is that we need to be aware of the
synchronize requirements, the cairo_surface_flush and
cairo_surface_mark_dirty, around access to the transport image. To
reduce the frequency of these barriers, we can subdivide the transport
image into small chunks as to satisfy individual updates and delay the
synchronisation barrier until we are forced to reuse earlier pixels.
Note this bumps the required Cairo version to 1.12, and please be aware
that the XSHM transport requires bug fixes from cairo.git (will be
1.12.12)
v2: After further reflections with Mitch, we realized we can share the
transport surface between all canvases by attaching it to the common
screen.
v3: Fix a couple of typos in insert_node() introduced when switching
variables names.
v4: Encapsulating within an image surface rather than a subsurface was
hiding the backing SHM segment from cairo, causing it to allocate
further SHM resources to stream the upload. We should be able to use a
sub-surface here, but it is more convenient to wrap the pixels in an
image surface for rendering the filters (and conveniently masking the
callee flushes from invalidating our parent transport surface).
Cc: Michael Natterer <mitch@gimp.org>
And along with it a lot of stuff like the drawable preview cache, the
gegl tile manager backend, temporary gimp_gegl_buffer_foo() stuff, and
the remaining bits of performance.
The projection is in an evil semi-ported state which makes it work
ok-ish for stuff like layer moving, but absolutely unbearable for
painting, there is also an off-by-one rendering glitch at some zoom
levels.
Create and use Cairo-compatible API for display filters. Also
includes logic changes to the display filters to deal with cairo's
ARGB32 pre-multiplied buffer format.
Jehan