Things like `BLI_uniquename` had nothing, but really nothing to do in
BLI_path_util files!
Also, got rid of length limitation in `BLI_uniquename_cb`, we can use
alloca here to avoid overhead of malloc while keeping free size (within
reasonable limits of course).
It is quite likely in a triangulated mesh that the actual island edge
belongs to a different triangle than the current pixel; for example
consider corners of a triangulated axis aligned rectangle face that
have the additional edge: a pixel there will have to be assigned to
one of the triangles, but one of the edges of the original rectangle
can only be accessed through the other triangle.
Thus for robust operation it is necessary to do a recursive search.
The search is limited by requiring that it only goes through edges
that bring it closer to the target point, and also by depth as a
safeguard.
Differential Revision: https://developer.blender.org/D2409
The code requires the pixel on the other side of the seam to be assigned
precisely to the expected triangle. This can cause false negatives around
vertices, where a pixel is likely to touch multiple triangles and thus
cannot be said to unambiguously belong to any one of them, so check
distance to the intended triangle and accept the result if it's close.
1. Forcibly symmetrize the neighbor relations, so that if A is neighbor
of B, B is neighbor of A. The existing code is guaranteed to violate
this if texture resolution is different between the sides of a seam.
2. In texture mode dynamic paint adds a 1 pixel wide border around the
islands. These pixels aren't really part of the dynamic paint domain
and thus by design can't have symmetrical neighbor relations. This
means they can't be treated by effects like normal pixels.
The simplest way to handle it in a consistent way is to exclude
them from effects, but add an additional pass that recomputes them
as average of their non-border neighbors, located on both sides of
the seam.
This splits `interp_weights_face_v3` into `interp_weights_tri_v3` and
`interp_weights_quad_v3`, in order to properly handle three sided polygons
without needing a useless extra index in your weight array. This also
improves clarity and consistency with other math_geom functions, thus
reducing potential future errors.
Reviewed By: mont29
Differential Revision: https://developer.blender.org/D2461
If a very low wetness absolute alpha brush is used with spread and
drying effects enabled, some pixels will rapidly accumulate paint.
This happens because paint drying code applies a minimal wetness
threshold that causes the paint to instantly dry out.
Specifically, every frame the brush adds paint at the specified
absolute alpha and wetness set to the minimal threshold, spread
drops it below threshold, and finally drying moves all paint to
the dry layer. This drastically accelerates the rate of flow of
paint into the affected pixels.
Fortunately, the reason paint spread actually ends up decreasing
wetness turns out to be a simple floating point precision problem,
which can be easily fixed by restructuring the affected expression.
Duplicates can happen at UV seams in case of resolution mismatch
or other complications. It's better not to store them in case it
confuses some math later on.
Reviewers: mont29
Differential Revision: https://developer.blender.org/D2261
1. When adding one pixel border to UV islands prefer grid directions.
This is more logical as it means neighbor pixels will commonly
share a border, opposed to just corners.
2. Don't subtract 0.5 when converting float UVs to int in computing
adjacency across a UV seam. Pixels cover a square with corners
at int positions and adding/subtracting 0.5 is only for dealing
with the center point of a pixel.
3. Use the neighbour_pixel field from the correct pixel, and check
it's not back to the origin point.
4. In the connected UV case, ensure that the returned index actually
refers to a valid active pixel.
This fixes paint spread not traversing some UV seams, while at the
same time spreading to random unrelated points on other seams.
The first problem is primarily fixed by 2, while the second one
is addressed by item 4.
Differential Revision: https://developer.blender.org/D2261
Regression from rBa4a968f, we would adjust current point's wetness without actually protecting it
in new multi-threaded context, leading to concurrent access mess.
Now delay applying wetness reduction to current point to end of function, allows us to avoid having
to lock current point twice together with neighbor one (and reducing spinlock awainting too).
To be backported to 2.78a.
Point cache read code contains checks designed to prevent it reading
stale data when the relevant simulation code should instead compute
the next frame from the previous one. However in some situations like
motion blur subframes the simulation can't possibly do it and just
exits. This causes completely incorrect motion blur at or after the
last cached frame.
To fix, add a parameter that tells the cache code whether it should
apply the checks and exit, or read what it can even if stale (true
means exactly same as old behavior).
Doing this in cache rather than clamping the frame number better in
the caller lets it handle the case of incomplete cache that stops
before the official last frame.
Reviewed By: mont29, lukastoenne
Maniphest Tasks: T49004
Differential Revision: https://developer.blender.org/D2144
Two were actual bugs, though they existed only in unused code:
* In Freestyle it was unintentionally copying a scene rather than referencing it.
* In BLI_array_store_is_valid there was use of uninitialized memory.
Replaces `G.is_rendering` with `use_render_params` argument.
This is needed for Cycles, which attempts to restore render-preview settings from particles,
after it gets its own particle data, but fails to restore because
`G.is_rendering` was being checked in psys_cache_paths (and other places).
This commit makes Dynamicpaint modifier evaluation (during playback) a few percents quicker.
However, it makes dynapaint's 'image sequence' baking about 33% quicker (from 119 to 77 seconds
in own heavy test), partly due to switch to BLI_task itself (about 20%), and partly due to
optimizations (remaining ~13%).
As usual, did a lot of tests here to ensure nothing is broken, but a lot more users' testing would definitively
be welcome too! ;)
Note that some quite meaningless omp forloops have been removed (parallelizing thousands of vec copy does
make it two or three times quicker, but the few hundreds of microseconds gained do not make any difference
in a hundreds millisecond process).
Also, this code could still use a lot more cleanup (naming etc.), the way it (tries to) handle malloc faults
is also totally flacky and makes the code horribly verbose and convoluted in some places - without actually
catching all possible faults (memarena could make it more easy to handle here), etc.
Gives 3-4% speedup in pre-bake step (from 112 to 108 seconds with own heavy test file).
Note that here we have a huge potential performance boost if we replace the flat
`Bounds2D *faceBB` array of UV tris bounding boxes by a real 2D AABB tree (right now,
we can loop over all UV triangles times the number of pixels of the surface times 5...).
Was not so far, because this effect is not modifying its 'own' PaintPoint, which means
it's not threadsafe. Since a global lock (mutex or spinlock) would not be much efficient
(we need to lock a given point pretty much all the computaion cycle), and since locking
a same PaintPOint from different threads at the same time is *very* unlikely,
solution here is to use an 'array of locks', one for each PaintPoint (same thing as BLI_bitmap,
using atomic ops to set/clear bits).
Here in own test (complex dynapaint over a huge sphere combining all dynapaint types), it gives
20% speedup of the whole dynapaint simulation!
Note: maybe we'd want to move that kind of bitlock into BLI lib some day - not totally sure how,
so let's keep it local for now...
Vertex's normal is always normalized - and its conversion from short to float should not change that!
On the other hand, linear interpolation of three normals **does not give a normalized vector**
(unless all three inputs are exactly the same).
Also, minor optimization, avoid recomputing that interpolated normal twice for each PaintUVPoint.
Line lengths, monolined 'if' statements, int -> bool, etc.
Also, replaced some internal cooked stuff by BLI helpers (most notably, the
'is inside UV triangle' code in `dynamicPaint_createUVSurface()`), and some
other minor optimizations.
`dynamicPaint_doMaterialTex` was called from inside an omp parallel section with brush->dm
itself, and not the local dm copy. Generating looptri data is not thread safe at all...
We cannot use FLT_MAX as initi distance for raycast...
Renamed TRANSFORM_DIST_MAX_RAY to BVH_RAYCAST_DIST_MAX, moved it into BLI_kdopbvh,
and use in RNA raycast callbacks (and all other places using that API).
As in the title. In the smoke version, there was also an extra
'for_render' parameter that wasn't used, and wasn't used by the callers
either, so it was removed altogether.
Reviewers: brecht
Reviewed By: brecht
Subscribers: brecht
Differential Revision: https://developer.blender.org/D1718
- Add blentranslation `BLT_*` module.
- moved & split `BLF_translation.h` into (`BLT_translation.h`, `BLT_lang.h`).
- moved `BLF_*_unifont` functions from `blf_translation.c` to new source file `blf_font_i18n.c`.
ImBuf types were getting stored as bitflags in a 32bit integer which had
already run out of space. Solved the problem by separating file type to
an ftype enum, and file specific options to foptions.
Reviewed by Campbell, thanks a lot!
The issue was caused by the conflict between preview render which would set
R_NO_IMAGE_LOAD flag on the renderer and texture samplers called outside of
the render pipeline trying to use this flag.
Now the sampler functions accepts extra argument so render pipeline can
still skip image load, but calls outside of the pipeline will nicely load
all the images.
Not cleanest change in the world but good enough to unlock gooseberry team,
and assuming we already had pool passed all over the place it should be all
fine.
Will need to reshuffle arguments into SamplerOptions structure later.
