With large steps in hair editing strokes the solution to distance-constrained
hair displacement becomes inaccurate, leading to a gradual lengthening of strands.
To counteract this the length constraints can enforce a velocity relative to the
constant target length of each segment. This results in an overall constant
length, even when the constraint solution is inaccurate for nonlinear constraints.
Note that the result would become much better with backward Euler integration,
currently there is a noticable 1-step lag between strokes and the distance fix.
it is incorrect to set ROOT_DIR and then call find_package.
If something is expected to be in a given location use hardcoded
locations (for the precompiled libraries). Otherwise just use
find_package() and let users to set ROOT_DIT when it is required.
The first character of the ID type was used to tag IDs for updates which
is weak since different IDs can have the same first character (for
example meshes, materials and metaballs), causing unnecessary updates of
unrelated IDs.
Now we use a unique index per ID type to tag for updates, unifying IDs
arrays indexing along the way.
Reviewers: sergey, mont29
Differential Revision: https://developer.blender.org/D2139
Depth buffer values are used by the viewport pan and zoom code to
adjust response scaling factors between mouse and viewport movement.
Letting smoke write to the buffer confuses it and causes the camera
to get stuck and move very slowly inside smoke domains, because it
thinks it is very close to an object.
Another issue with the modified particle motion blur fix: since
pre and post are used as validity markers, they must be set even
if there is no actual motion, like the original bool flags were.
Otherwise an object starting to move or stopping is interpreted
as having invalid blur data and hidden.
Wasn't copying back local view bytes to object after changing layer.
Steps to reproduce were:
* Split 3D view in default startup.blend
* Enter local view in one of the 3D views
* Move default cube to different layer *in the other 3D view* (the one that's not in local view)
* Try transforming object from 3D View that's in local view (should lag)
Displacement is now a per material setting, which means old files will have to
be updated if they had used displacement. Cool side effect of this change is
material previews now show displacement.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2140
Enables Catmull-Clark subdivision meshes with support for creases and attribute
subdivision. Still waiting on OpenSubdiv to fully support face varying
interpolation for subdividing uv coordinates tho. Also there may be some
inconsistencies with Blender's subdivision which will be resolved at a
later time.
Code for reading patch tables and creating patch maps is borrowed
from OpenSubdiv.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2111
From source code history, it seems this behavior is a relic of a very
old limitation when the same object couldn't be both a Softbody and a
Collision object. In those times if somehow both modifiers were added
to an object, Softbody auto-disabled itself and buttons were hidden.
Now however there is no problem having both modifiers on an object,
so there is no point hiding the buttons. The same exact buttons are
available on the physics tab in any case.
As reported in T48928, From Dupli UV is not supported for grid
distribution, and running the normal code as usual simply produces
nonsense data, because fuv is used to hold orco instead of
interpolation factors for uv, and num is zero.
Since support won't be added in 2.78, just stop outputting nonsense.
This uses the Lagrange multiplier method for correcting the "velocity" of
vertex displacements as described in
"Linear-Time Dynamics using Lagrange Multipliers" (Baraff, 1996)
Future additions would include stiffness forces for creating more natural
hair shapes, and handling of collisions (deflection).
All in all, this patch adds an Alembic importer, an Alembic exporter,
and a new CacheFile data block which, for now, wraps around an Alembic
archive. This data block is made available through a new modifier ("Mesh
Sequence Cache") as well as a new constraint ("Transform Cache") to
somewhat properly support respectively geometric and transformation data
streaming from alembic caches.
A more in-depth documentation is to be found on the wiki, as well as a
guide to compile alembic: https://wiki.blender.org/index.php/
User:Kevindietrich/AlembicBasicIo.
Many thanks to everyone involved in this little project, and huge shout
out to "cgstrive" for the thorough testings with Maya, 3ds Max, Houdini
and Realflow as well as @fjuhec, @jensverwiebe and @jasperge for the
custom builds and compile fixes.
Reviewers: sergey, campbellbarton, mont29
Reviewed By: sergey, campbellbarton, mont29
Differential Revision: https://developer.blender.org/D2060
There were two issues here. One is that the fix done originally for this
bug only checks for colliding with the same face as the single preceeding
hit. If the particle hits an edge or vertex of the collider, it in fact
hits two or more faces, so the loop ends up cycling between first two
of them and reaches the max collision limit.
The fix is to disable the collider for the sim step once a permeability
roll succeeds, by adding it to a skip list. Skipping just one face causes
some particles to bounce at odd angles in case of partial permeability.
The second problem was that the collider bounced back a small percentage
of particles, and the cause seemed to be that the code was set to flip
the velocity if the particle was just past the collider but still within
collision distance. Inverting both values causes a half permeable collider
to stop particles, so it seems that this if branch shouldn't bounce at all.
Test file: {F327322}
Reviewers: lukastoenne, brecht
Reviewed By: brecht
Subscribers: brecht, #physics
Maniphest Tasks: T26658
Differential Revision: https://developer.blender.org/D2120
Adds a descriptor for attributes that can easily be passed around and extended
to contain more data. Will be used for attributes on subdivision meshes.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2110
* "Flip direction" -> "Flip Direction"
* "Show drawing direction" -> "Show Directions"
* "Grease Pencil Curves" -> "Brush Curves"
(I was considering "Brush Response Curves" instead, but that seemed like too much
of a mouthful)
* "X" for removing a palette. The UI there was more similar to a standard datablock
selector, so it should use the "+X" combo instead of "+-" combo for consistency.
(Note though, presets tend to use "+-" instead - e.g. see the Render Settings)
Previously, it would insert "invisible" points after the endpoints of the strokes,
so that they wouldn't appear to be joined, but that behaviour could also get quite
confusing as you wouldn't be sure whether the strokes were really joined or not.
To keep the previous behaviour, simply enable the "Leave Gaps" option on the operator
after running it. This setting will get saved between runs of the operator.
Group membership testing for including/excluding feature lines was not
accounting for object names possibly further qualified by library file
paths.
Also fixed a few potential (but unlikely) references of uninitialized
variables.
A big thank to Bastien Montagne for the insight on the cause of the
problem and how to fix it.
There were actually two issues, one in recent changes and one existing... forever?
I) id_make_local() would never proceed over indirectly linked data, which is required in specific 'append' context.
II) BKE_image_copy() was not setting id->newid pointer of newly copied ID.
About II: don't really know why image copy does its own cooking instead of using generaic ID copy helpers.
Think this should be changed, but that would be after 2.78 now.
Drawing both text and the wave onto a sound strip makes both hard to read,
which is a concrete issue for Hjalti at the moment. This was the simplest
fix I could think of to give him control over what he sees.
Currently cycles cannot correctly render motion blur for objects that appear or
disappear during the shutter window. Until that can be fixed properly, it may be
better to hide such particles rather than let them render as if they were
stationary for half of the frame.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2125
It is possible that compilation will fail without giving anything in the
log buffer. For this cases giving a tip about error code will be really
handy.
Patch by @Ilia, thanks!
Now all consistent and using term "properties region" instead of "properties panel". Ideally we had a more generic operator for all those.
Fixes T49006.
This patch improves UI of a recently added solver preference in boolean modifier:
{F331776}
Issue with the current UI is that it shows user unnecessary information and breaks established grid layout.
Reviewers: carter2422
Reviewed By: carter2422
Subscribers: carter2422
Tags: #user_interface, #bf_blender
Differential Revision: https://developer.blender.org/D2136
This changes actually lead to 2x slowdown. It's getting a bit annoying
because those are the changes to make pre-maxwell cards render with the
same speed.
The problem happens because smoke collides only with the surface of the
collider and uses incompressible fluid solver. This means that scaling
the collider tries to compress or decompress fluid within the volume of
the collider, which can't be handled by the simulation. Fast rotation
likely also causes transient scaling due to emulation of arcs by chords.
This can be fixed by finding compartments completely isolated by obstacles
from the rest of the domain, and forcing total divergence within each one
to be zero so that equations are solvable. Physical validity is somewhat
dubious, but without this the solver simply breaks down.
From the physics point of view, the effect of the correction should be
similar to opening a hole from every cell to another dimension that lets
an equal amount of air to pass through to balance the change in volume.
Reviewers: miikah, lukastoenne
Reviewed By: lukastoenne
Subscribers: dafassi, scorpion81, #physics
Maniphest Tasks: T43220, T47551
Differential Revision: https://developer.blender.org/D2112
When the colliders are joined, each cell that touches them only gets
velocity contribution once. When there are multiple objects, velocities
are summed, which causes some cells to get 2x, 3x or more than the actual
object velocity.
Fix this by using the average velocity of all colliders touching a cell.
Reviewers: miikah, lukastoenne
Reviewed By: lukastoenne
Subscribers: dafassi, scorpion81, #physics
Maniphest Tasks: T43782
Differential Revision: https://developer.blender.org/D2112
As a result of other folding simplifications it may happen that
two type conversion nodes end up directly connected. In some
cases it may be possible to then remove both. A realistic case
might be an optimized out Mix RGB node used to blend vectors.
It seems it's safe to optimize when B is a float3 type
(color, vector), and A is float3 or float.
Reviewers: #cycles, sergey
Reviewed By: #cycles, sergey
Subscribers: sergey
Differential Revision: https://developer.blender.org/D2134
(Re)-setting custom normals could cause some unwanted splitting of some of them, leading
to slightly different tangent space. Simply enlarged slightly the threshold detecting
similar normals as identical ones for now, afarid this is the kind of issue that cannot
get a full complete solution for until we drop floats...
This way we can easily switch between toolkits without worrying
whether some kernel was compiled with old or new CUDA toolkit.
It's also now possible to switch machine architecture and have
proper cached kernel detected. Not as if it happens every day,
but i did such a bitness switch back in the days :)
Make 'Laplace' filter an edge filter operation, since this what it is
typically used for, and such operation does not affect the input's alpha
channel.
Reviewers: sergey, campbellbarton
Reviewed By: sergey
Differential Revision: https://developer.blender.org/D1817
The cause seems to be that despite dt_frac being computed as
1/(subframes+1) with integer subframes value, it doesn't always
add up to exactly 1.0 due to precision limitations. If the sum
is similar to 1.00000???, the last subframe is skipped, and all
particles that were supposed to be emitted in that interval are
emitted next frame, with the code working incorrectly due to
skewed time range.
To fix, separate the code from the dynamic timestep feature that
adjusts the last subframe length into a separate function, and
use it even when dynamic timestep is disabled.
Code coverage of different combinations of secondary conditions
is obviously not complete because there are so many of them, but
all main rules should be there.
The reason for CORRECT vs INVALID is that both words have the same
number of characters so calls line up, but look quite different.
Reviewers: #cycles, sergey
Reviewed By: #cycles, sergey
Subscribers: dingto, sergey, brecht
Differential Revision: https://developer.blender.org/D2130
All the changes are mainly giving explicit tips on inlining functions,
so they match how inlining worked with previous toolkit.
This make kernel compiled by CUDA 8 render in average with same speed
as previous kernels. Some scenes are somewhat faster, some of them are
somewhat slower. But slowdown is within 1% so far.
On a positive side it allows us to enable newer generation cards on
buildbots (so GTX 10x0 will be officially supported soon).
Enabled the Undo option for the operation "Add background Image". This is to treat adding a background image like opening an image. This behavior is useful when a background image is add using 'drag and drop'.
Reviewers: Severin
Reviewed By: Severin
Subscribers: brecht
Tags: #bf_blender
Maniphest Tasks: T47050
Differential Revision: https://developer.blender.org/D1725
The redundant terms were harmless but check an expression we already
know to be true (from earlier in the same conditional).
Found by PVS-Studio T48917
These values often either turn the node into a no-op, or even make it
evaluate to 0 no matter what the other input value is, thus allowing
deletion of a branch of the node graph that otherwise is not constant.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2085
These are complex nodes, and it's conceivable they may end up constant
in some circumstances within node groups, so folding support is useful.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2084
Drivers can use this to refer to the data which the driver is applied to,
useful for objects, bones, to avoid having to create a variable pointing to its self.
This adds the ability for cloth simulations to respect changes in the underlying mesh.
So you can for instance, animate shape keys, armatures, or add any deformation modifiers (above the cloth modifier).
This is mainly useful for (but not limited to) cartoon animations,
where your character might stretch or change shape, and you want the clothes to follow accordingly.
D1903 by @LucaRood
As title says... Copying tracking data from movieclip was not the simplest thing...
Reviewers: sergey
Differential Revision: https://developer.blender.org/D2126
This commit adds some easy to use way to check whether the graph
was properly optimized. The idea is based on using mock glog sync
which keeps track on all messages and expects specific log message
to appear in the logging.
This means each optimization step should report that it optimized
something (specifying what exactly was optimized) and what was used
for optimization.
There's also some ease-ish way to define shader networks which we
might want to make a bit more global and available for all tests
in the future.
For now only RGBToBW node is covered as an example, really hope
to get help from active guys in the community to finish covering
all existing optimization cases. Ready to fix any possible issues
with builder when needed tho :)
Initial report/patch by Alexander Gavrilov (@angavrilov), found more on the road.
Nice demo of why we should use libquery ID looper in read code too - but that's for another day
(also because read code needs to take care of some deprecated pointers sometimes)...
Hating all those not-so-real ID types... Here there were two causes for the issue:
1) Linked shapekey ID was made local twice (once from mesh's make local, once by itself).
Solved by not explicitely making shapekeys (nor any other non-linkable datatype) local.
2) Key->from 'back pointer' to its owner was messing 'still in used' detection of linked data
after localization. Fixed with a hack for now, thinking correct solution might actually
be to not consider this pointer at all in libquery ID looper, since it's nothing like
and actual usage of mesh/lattice/curve.
Again, shapekeys as ID is a joke, those should be mere struct, they have absolutely nothing to do in Main, period. :(
Point 2) still demonstrates the need for better handling of IDs dependencies though,
so far we only hit corner cases, but think there could also be valid cases generating those
'dependency cycles' between IDs (ID a using ID b which uses ID a), this will have to be addressed some day...
This method is just a placeholder, in the future it would be worthwhile to use Bullet collision
detection with ghost objects for this.
Also the code should eventually be abstracted a bit to make it agnostic to mesh formats, so it
can be used for BMesh as well as DNA strand data or physics simulation.
This should serve as a universal interface between possible collision
detection systems (BVH, Bullet) and the various collision response systems.
Using an new data structure means that the two sides of a collision
handling feature (detection and response) can be treated independently,
reducing code dependencies and complexity.
The properties of contact points are modeled after btManifoldPoint from
Bullet and may need to be adjusted in the future.
The depsgraph component node for invalidating the strand shader was located in
the GEOMETRY component of the object. This component by default gets tagged along
with every other object part during editing. Only standardized way of avoiding this
would be putting it in a different datablock, but the shader is very much tied to
the particular object due to the displacement function from its node tree.
For now the solution is a new component GEOMETRY_SHADER, which does not get tagged
when the ID node is tagged. It still gets regular updates from node tree changes.
This seems to be the least invasive way of avoiding unnecessary performance drops,
and can be removed and replaced easily later on without disrupting other code work.
This makes no sense since user usually only has release build... And those noisy printings are
behind debug flag anyway, so really no reason to only print them out in debug build.
There was some smart code in the transform which would alter between translation
and rotation based on whether bone is connected or not and whether translation is
locked or not.
It could be handy to also fallback to scale if both rotation and translation are
locked.
Subdivision options can now be found in the subsurf modifier. The modifier must
be the last in the stack or the options will be unavailable. Catmull-Clark
subdivision is still unavailable and will fallback to linear subdivision instead
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2109
This adds support for ngons and attributes on subdivision meshes. Ngons are
needed for proper attribute interpolation as well as correct Catmull-Clark
subdivision. Several changes are made to achieve this:
- new primitive `SubdFace` added to `Mesh`
- 3 more textures are used to store info on patches from subd meshes
- Blender export uses loop interface instead of tessface for subd meshes
- `Attribute` class is updated with a simplified way to pass primitive counts
around and to support ngons.
- extra points for ngons are generated for O(1) attribute interpolation
- curves are temporally disabled on subd meshes to avoid various bugs with
implementation
- old unneeded code is removed from `subd/`
- various fixes and improvements
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2108
This is a bit weak, but better than tagging whole mesh manager for update.
Maybe we'll solve such dual-look up in the future.
This commit finally solves T48963: Noise when changing Diffuse node to Emission node
This mostly affects physics panels. Any engines relying on
RenderEngine.use_game_engine flag to show/hide panels will need to be
updated. The COMPAT_ENGINES technique is how we usually deal with this.
One issue with use_game_engine is that I cannot find a way to set it; it
appears only the BGE can set it. This means (without this commit)
external RenderEngines cannot get rid of the default physics panels.
The RE_GAME flag (the C flag behind use_game_engine) is pretty hacky
and we should look into removing its usage where possible.
Happens because it was always using regionbase of active/visible area, not the correct one of the (possibly inactive) 3D view. Was pretty hidden because you had to be in local view to trigger it.
Could also remove ScrArea * argument from SpaceType.id_remap now, but leaving in there to avoid bigger changes (might also be useful for some case).
Steps to reproduce:
* Default startup.blend
* Enter local view with default cube (/-key)
* Render (or open a different editor where the 3D View is)
* Close Blender - should invoke crash
Enable triple buffer by default for AMD cards on opensource driver.
Should be safe now, since it was only old cards causing problems,
which we don't support now most likely anyway.
This patch fixes the annoyance that when creating a node group where one of its nodes is connected
to several other nodes, a separate output will be created for each link, even though they're all
connected to the same socket in the group.
Now, before adding an output for an outgoing link, the existing outputs are checked to find whether
any output is already connected to the same socket. If such an output is found, it is reused instead of
creating a new one.
Reviewers: Severin
Subscribers: Blendify
Differential Revision: https://developer.blender.org/D1836
- Nodification of shader nodes changed default strength from 10 to 1.
If that was intentional, IMO should happen as a separate commit.
- Fixed indentation.
This will allow easier addition of other shader features later on. The system
should be merged with the GPUMaterial code, but currently the material system
is too unwieldy to integrate the strand displacement code easily. With the 2.8
viewport rewrite all of this will have to be rewritten as well anyway.
Report popups simply reuse popup menu code, so their blocks got "UI_popup_menu_begin" assigned as name, which was a bit misleading. Now uses "UI_popup_menu_reports".
Ideally they'd have their own popup code.
Commit rB709ca0ece changed how rotation was handled for particles so
that unless actual rotation physics is enabled, there is no rotation.
However it only updated ptcache_particle_read, forgetting to change
exactly the same code in ptcache_particle_interpolate.
This means that for subframes old code that computes a rotation from
velocity is used, resulting in completely different rotation than for
integer frames. This causes rotational motion blur by itself, and also
mangles motion blur paths of child particles.
Reviewers: sergey, lukastoenne
Maniphest Tasks: T45936
Differential Revision: https://developer.blender.org/D2124
For now use 'brute force' and refresh whole UI when new icons get rendered.
See comment of T48813 for ideas about how to handle that in future (2.8 project ?).
While they prevent legit write past the array boundary error
those fixes introduced regression in behavior when having exact
max_hits transparent intersections and nothing else.
Previous code would have considered such case a totally opaque,
but it's not correct.
Fixes T48941: Some materials don't get transparent shadows anymore
In the triangle intersection refinement code, rays that are parallel to the triangle caused a divide by zero.
These rays might initially hit the triangle due to the watertight intersection test, but are very rare - therefore, just skipping the refinement for them works fine.
Also, a few remaining issues in the MultiGGX code are fixed that were caused by rays parallel to the surface (which happened more often there due to smooth shading).
This is an alternative method for fitting a curve which incrementally simplifies the curve, then re-fits.
Generally gives better results, also improves corner detection.
We need to leave tweak mode before trying to modifiy the action as doing
so will leave Blender in a semi-corrupted state.
Reviewers: #animation
Reviewed by: aligorith
Maniphest Tasks: T48397
Differential Revision: https://developer.blender.org/D2119
3D view
There was a misusage of the `outcol` and `texcol` params. The actual
formula should have been:
incol = facm * outcol + fact * ((one - outcol) * texcol * outcol +
outcol * scr);
To make sure the result is consistent with material mode, reuse the
material blend function (mix_soft), similarly to what most other texture
blend modes do.
The depsgraph now invalidates the strand shader when the nodes are changed.
As mentioned in the previous commit, this happens through an operation in
the object GEOMETRY component for each strand modifier. This is messy, only
useful until we have truly granular modifier integration in the depsgraph.
Invalidating the shader is currently done just by setting a flag. The actual
freeing and subsequent lazy-rebuild of the shader is done in the drawing
thread, because during the threaded depsgraph update we don't have a valid
OpenGL context. This needs to be sorted out for future viewport recode.
This reverts commit ca85cb436d.
Reverting the commit will break depsgraph updates for regular bvm functions again, but is necessary
for making shader updates working. The problem here is that the update operations for _dependent_
data (bvm functions, shaders, etc.) should not actually be part of the node tree, but of the
respective dependent data components! So the previous method of invalidating shaders as part
of the node tree operations is incorrect and the operation should be moved to some other place.
Perhaps we need a dedicated depsnode for the bvm function cache, since there is no single "owner"
of these functions that could take care of invalidation in the depsgraph.
Suspicious conditional found by PVS-Studio T48917
Original code (from Blender’s initial open-source commit!) looks like
it’s testing inputs to atan2f, to avoid undefined function values.
Passing xn=0 does *not* always evaluate to 0 though… I’m not sure if
this is a coding error or was done for a desired visual result.
Also changed xn==0 to xn>=0 to avoid function call in more cases.
Good description and visualization of atan2f function:
http://en.cppreference.com/w/c/numeric/math/atan2
It became rather annoying to have those functions duplicated.
Surely, it's not really nice it's actually a bad level call,
but similar thing is happening in OCIO and Cycles.
IMO, it's better than having functions re-implemented, and
have this solved with new OpenGL pipeline.
Redundant conditional (line 939) found with PVS-Studio T48917
There’s still a lot of true/false, 1/0, SELECT/DESELECT usage nearby.
Not a problem, just confusing to read.
Checking a condition right after we’ve checked it (and it hasn’t
changed). Most of these are trivial to understand.
split_quads in convertblender.c:
It seems quads should be processed and triangles should be marked as
not needing split. So I removed the outer vlr->v4 check.
Found with PVS-Studio T48917
Some of these check that dimensions match before running code that
assumes they do match.
For imb_stereo3d_write_anaglyph I *assume* this change reflects the
intended behavior. Before it was always grabbing alpha from buffer 0.
Found with PVS-Studio T48917
Check to ensure we do not delete last instancing of an indirectly used object (forbidden
because it creates 'ghost' objects user have no real way to re-instance) was defeated by
the backward pointer 'proxy_from', which generated a 'false' indirect usage of (local)
proxy object itself (the one we are trying to delete).
Fixed by actually considering that ID usage as local if proxy object itself is local
(because that pointer actually does not 'pertain' to the object holding it, but to its proxy).
Yeah, it's... complicated, twisted and ugly - it's proxy.
(A - A).norm() is always 0 so condition is always true.
(A - B).norm() and (B - A).norm() both compute the same distance so I
picked one to match surrounding code.
Found with PVS-Studio T48917
The value is calculated without any derivative inputs (this is not a differential equation).
The derivatives in both variables also depend on the partial derivatives of the inputs.
I don’t think any other codec enum makes sense here, so probably just
an extra copy/paste. Here are the video codecs:
AV_CODEC_ID_NONE
AV_CODEC_ID_MPEG4
AV_CODEC_ID_MJPEG
AV_CODEC_ID_DNXHD
AV_CODEC_ID_MPEG2VIDEO
AV_CODEC_ID_MPEG1VIDEO
AV_CODEC_ID_DVVIDEO
AV_CODEC_ID_THEORA
AV_CODEC_ID_PNG <— alpha
AV_CODEC_ID_QTRLE <— alpha
AV_CODEC_ID_FFV1 <— alpha (if enabled)
AV_CODEC_ID_HUFFYUV <— alpha
AV_CODEC_ID_H264
AV_CODEC_ID_FLV1
Found with PVS-Studio T48917
Note that for the hair displacement function we would only be interested
in a single derivative (the tangent). For now we just use the existing
2-variable duals, later on this can be generalized to avoid unnecessary
and confusing code emission.
This is really not elegant solution, but simplest for now.
Ideally, we'd check all IDs first and directly 'make local' (without need to copy)
those only indirectly used by other datablocks to be made local. Would also save us from the
need of the extra 'lib_local' parameter recently added to id_make_local.
Current code seems to work well enough though, and this is not high priority to fix imho.
Those bone pointers in object's pose bite again - turns out they can be accessed
before pose actually gets rebuilt in some cases (e.g. from undo writefile), so
we need to clear the pointers immediately.
Adding make_local for vfont/gpencil/group/freestyle linestyle/mask/scene/sound/text.
Note that there are still some lose ends here, since some type are not handled by id_copy
(Scene, Sound and VFont), which means in case a datablock of that type is used both
directly and indirectly, localization will fail.
Scene copying might be doable though (maybe with a simple new 'full shalow' mode to existing BKE_scene_copy?),
not sure about sounds and vfonts... Situations where this becomes an issue should be very rare corner cases anyway.
Turns out most BKE_foo_make_local datablock-specific functions are actually doing
exactly the same thing, only two currently need special additional operations
(object and brush ones). So added a BKE_id_make_local_generic instead
of copying same code over and over.
Also, changed a bit how make_local works in case we are localizing a whole library.
We need to do the 'remap' step (from old linked ID to new local one) in the second loop,
otherwise we miss some dependencies. This fixes main part of T48907.
Mainly the changes are related on establishing API to feed UV islands
to OpenSubdiv, so it will know all the connectivity information and
will be able to do proper interpolation.
Island calculation is currently rather slow, not sure how to make it
fast and not use lots of allocations.
Shouldn't be THAT bad, since this code is only runs once when creating
OSD mesh, and it's probably still faster than our orientation code.
Note that issue has several levels here actually, first one was metaball's materials
not being properly copied into new mesh (code was commented out because of some crash it
seems, made it a bit closer to mesh one and got no crash at all...).
Then, we were calling test_object_materials when ob->data is actually *not* new tmpmesh!
Will remove this call completely in next commit (to make it easier to bisect), I cannot see
any case where object would be assigned with newly generated tmpmesh in this func.
A bit work in progress, currently the following limitations:
- Texture shading only, Material shading will come later
- No UVs subdivision yet
- Always uses active UV and currently changing active UV will
not properly update the viewport.
Well, need to start somewhere :)
This is just a stub atm. The BVM system will use nodes to generate
a GLSL function, which can then be inserted into and compiled with
a framework shader. The deformation function itself is just a simple
expression (with autodiff for tangents), so it doesn't need to know
about the topology of index buffers and the like.
The flatten brush depended on accumulate being disabled,
Adding dynotopo support for accumulate caused problems for this tool (see T44390).
Enable for existing files.
- remove edge scaling, instead avoid checking intersections with connected edges.
- replace local line intersection functions with BLI_math
- center the projection for more precise calculation.
BKE_previewimg_copy() would simply copy PreviewImage's icon_id, without bothering about
ID one.
When we duplicate an ID, we want to reset its icon_id to zero (and regenerate it on-demand),
not keep same icon_id as original, so added new BKE_previewimg_id_copy helper to handle that.
This probably makes code somewhat slower, but we can't easily know
whether we can use a shortcut and only use directional lighting from
the scene.
Need some better integration between GPU and OpenSubdiv for that.
This patch addresses the following issues in bf_intern_ghost
```
Warning C4312 'type cast': conversion from 'GHOST_TEmbedderWindowID' to 'HWND' of greater size bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_WindowWin32.cpp 179
Warning C4312 'type cast': conversion from 'GHOST_TEmbedderWindowID' to 'HWND' of greater size bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_WindowWin32.cpp 198
```
GHOST_TEmbedderWindowID is defined as long, handles are however of pointer size,
so this should have been an issue when we moved to 64 bits, guess we got lucky.
fixed by turning GHOST_TEmbedderWindowID from long into void*
```
Warning C4302 'reinterpret_cast': truncation from 'HKL' to 'LANGID' bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_ImeWin32.cpp 67
```
reinterpret_cast emits warnings on truncation, LOWORD does the job just
as well with no warnings.
```
Warning C4838 conversion from 'int' to 'DWORD' requires a narrowing conversion bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_ContextWGL.cpp 734
Warning C4838 conversion from 'int' to 'BYTE' requires a narrowing conversion bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_ContextWGL.cpp 734
```
Weird warning, it does a really bad job at telling you what parameter is
causing the warning , tuns out there's a bunch of parameters that cause it
but it still only yields a single warning, the problem is that every
(somevar ? a : b) construct results in an integer type. which needs to be
properly cast to get rid of the warning.
```
Warning C4996 'GetVersionExA': was declared deprecated bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_WindowWin32.cpp 105
Warning C4996 'GetVersionExA': was declared deprecated bf_intern_ghost K:\BlenderGit\blender\intern\ghost\intern\GHOST_WindowWin32.cpp 107
```
The warning was clear, the code not as much. The version check in place
here is quite convoluted and could be replaced by including VersionHelpers.h
and calling IsWindows7OrGreater, However, CreateInstance will just return NULL
in m_Bar if the interface is not supported, so the whole check is useless.
This however did require that the CreateInstance call actually asked for
ITaskbarList3 and not ITaskBarlist . (You're not really allowed to assign
different interface types to each-other, a roundtrip through QueryInterface
is required there, we were violating spec here by asking for ITaskBarlist and
storing it in ITaskbarList3* )
Reviewers: sergey
Reviewed By: sergey
Subscribers: sergey
Differential Revision: https://developer.blender.org/D2094
This diff adds a 6th drawing mode to the Waveform Scope.
The new mode shows the RGB colour channels overlaid as a "Full colour" waveform.
The old "Red Green Blue" mode is renamed "Parade" which is the standard industry
term for RGB channels shown side-by-side.
This full colour style of waveform is very much more useful for colour grading than the
Parade mode and is the default waveform for many artists.
Files from older Blender versions which show scopes open as expected.
Patch by John Cox (johnedwardcox), thanks!
Reviewers: sergey
Reviewed By: sergey
Subscribers: campbellbarton, tmw, Blendify
Differential Revision: https://developer.blender.org/D1936
I'm not quite sure where the codegen bug gets triggered but it's easily
noticeable in the barcelona scene. (extra saturated leafs and illumination
on the right ledge of the pool)
2013 buildbot reference image:
{F320792}
2015 With no updates (compiler version 19.00.23026)
{F320793}
2015 With Update 2 (Compiler version 19.00.23918)
{F320794}
2015 With Update 3 (Compiler version 19.00.24210)
{F320795}
This patch blocks all compiler builds before update 3 in a similar way we
did for msvc 2013 update 4
Reviewers: campbellbarton, brecht, juicyfruit
Reviewed By: juicyfruit
Tags: #bf_blender
Differential Revision: https://developer.blender.org/D2100
We recently added delay-loading of the openmp dll's so we no longer had to use
the stub loader, we however put these linker flags on the global linker flags
causing any sub projects not using openmp to spit linker warning 4199 while building,
```
Warning LNK4199 /DELAYLOAD:vcomp140.dll ignored; no imports found from vcomp140.dll datatoc k:\BlenderGit\build_windows_2015a\source\blender\datatoc\LINK 1
```
This patch makes the delay-load only apply to the blender project.
Reviewers: sergey
Subscribers: sergey
Tags: #bf_blender
Differential Revision: https://developer.blender.org/D2092
The issue was caused by SSS intersection code gathering all
intersections without check for duplicated ones. This caused
situations when same intersection will be recorded twice in
the case if triangle is shared by several BVH nodes.
Usually this is handled by checking intersection distance
after sorting intersections (in shadow_blocked for example)
but for SSS we don't do such sorting and using number of
intersections to calculate various things.
Didn't find anything smarter than to check intersection
distance in triangle_intersect_subsurface().
This solves render artifacts in the cost of 1.5% slowdown
of extreme case rendering (SSS object filling in whole
FullHD screen).
Reviewers: brecht
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2105
The previous one probably only works because there are currently no more
than 3 valid control curves for each fiber, so the 4th one resets the
cumulative values to 0.0.
- In fresnel_dielectric, the differentials calculation sometimes divided by zero.
- When the normal map was (0.5, 0.5, 0.5), the code would try to normalize a zero vector. Now, it just uses the regular normal as a fallback.
- The approximate error function used in Beckmann sampling sometimes overflowed to inf while calculating r^16. The final value is 1 - 1/r^16, however,
so now it just returns 1 if the computation would overflow otherwise.
Also, this fixes a numerical issue where A would be inf.
Since later G is set to 1 if A is larger than 1.6, the code now checks the reciprocal of A for being smaller than 1/1.6 - same effect, but no inf involved.
This moves most of the data access logic for this data out of the GPU code and into
the respective BKE files. The concept is similar to how DerivedMesh is used in other
GPU code.
also changes the name of the callbacks:
- walk_parent_snap_project_cb to cb_walk_parent_snap_project;
- cb_leaf_snap_vert to cb_walk_leaf_snap_vert;
- cb_leaf_snap_edge to cb_walk_leaf_snap_edge;
- test_vert_depth_cb to test_vert_ray_dist_cb;
- test_edge_depth_cb to test_edge_ray_dist_cb;
and cleanup:
Number buttons & sliders weren't updating when shift/ctrl were pressed.
Note that this is a regression in 2.77 but hard to track down since it worked unreliably for some time.
This commit contains all the changes required for most optimal maximum threads
number bump. This is needed to avoid possibly unneeded initialization or data
allocation on systems with lower threads count.
TODO: Still need to review arrays in render data structures from render_types.h,
P.S. We might remove actual bump of max threads from this patch, so when we'll
be applying the patch we can do all the preparation work and then do actual
bump of max threads.
Reviewers: mont29, campbellbarton
Reviewed By: mont29, campbellbarton
Maniphest Tasks: T43306
Differential Revision: https://developer.blender.org/D1343
While it's an extra option added to the interface which might not be
fully obvious for artists, it allows to save up to 20% of memory in
hairy scenes.
This is high enough memory saver in my opinion which might become
handy for some production files where it's more important to make
scene to fit into memory rather than trying to use more optimal BVH
structure but go into swap or crash.
Reviewers: dingto, brecht
Reviewed By: dingto, brecht
Differential Revision: https://developer.blender.org/D2090
Currently lines and ribbons (triangle strips) are supported.
The ribbon width can be controlled with a parameter. Ribbons are
oriented toward the camera, similar to billboards, but shaded like lines
with the Kajiya shader imitating a reflective tube.
This function was only a wrapper around id_clear_lib_data(), and shapekeys
are not linkable nor shareable anyway, no point keeping this currently,
was only adding confusion about shapekey 'status' as a datatblock.
Currently only used in its broadest form for freeing the whole buffer,
but could be used in the future to selectively invalidate parts of the
data, to avoid costly uploading of large but static buffers.
Many checks for out->rect_float being [non-]NULL are done back-to-back.
Combining them into a single check for slightly more efficient code and
less code clutter for easier readability/understanding.
Differential Revision: https://developer.blender.org/D2097
The gpu_buffers.c file is becoming too large, and incorporates all the logic
from meshes and now strands, which should be separate from the API.
Making the strands buffer code separate also allows much more control,
e.g. for partially invalidating draw data and avoiding unnecessary buffer
uploads to the GPU.
After reformulation of SSS indirect rays it became possible to
try accessing dimension higher than was pre-calculated on scene
preparation.
This is because we're traversing rays backwards, which means we
are using higher dimensions first now.
It was possible to miss bounces termination criteria in this functions,
mainly when max_hits was set to 0.
Made the check more robust in traversal functions (which should not
affect performance, it's an operation of same complexity AFAIK).
Also avoid doing ray-scene intersection from shadow_blocked when
limit of transparent bounces was already reached.
Applying cflags globally can be problematic especially with extern, intern libs.
Now flags from target named will be used when defined,
allowing for developers to define flags for modules they maintain.
Convention is CMAKE_CFLAGS_${UPPERCASE_TARGET_NAME}, (CXXFLAGS for C++).
eg: CMAKE_CFLAGS_BF_BLENDER, CMAKE_CFLAGS_MAKESDNA, CMAKE_CXXFLAGS_CYCLES_KERNEL
On Linux run `make help` for full list of names, MSVC shows these in the solution.
Accessing `bpy.app.binary_path_python does search, then swaps its getset with the string it finds.
This caused a freed pointer to be stored in bpy.app's dictionary.
Fix by using the same string for get/set access.
Recently the snap to edges has been changed to work with coordinates 2d. Thus, the lambda of the snap pointon on edge becomes different from the actual lambda in a 3d edge.
The solution to correct the lambda was this:
```
if (is_persp) {
const float fac = depth_a / (depth_a + depth_b);
lambda *= (1.0f + (fac - 0.5f) * (1.0f - lambda));
````
...
But the lambda was still not very accurate.
Now the function uses the `dist_squared_ray_to_seg_v3` to get the coordinate of the snap point on the edge. And then tests that point.
The accuracy returned to normal values.
from D2104: reference all repeated (and strange) equations (example: mul_m4_m4m4(pmat_local, pmat, obmat)) in the function `precalc_project.
This is useful for maintenance.
Current code changes the selection then then runs ED_armature_sync_selection
to copy to the bones rootsel flag from the parents tipsel.
Use this logic for entering edit-mode too.
This reverts commit 47d0d9cca4.
Reverting the commit. Not only it did not solve all the cases of proxy popping,
but also broke real cases with single proxy involved.
The snapping functions when performed in the perspective view,
have some problems in the threshold (a distortion) and in the clip plane (the normal is incorrect).
These problems can be only observed when making the snap to edges or to vertices (nearest to ray function).
This patch propose a totally different solution.
The idea is to project the edges of bvh nodes and test the 2d projection of the snap element.
For this it used the BLI_bvhtree_walk_dfs function.
It is important to pay particular attention also to the changes in `ED_transform_snap_object_project_view3d_ex`
- icon_id from ID and PreviewImage were not guaranteed to be in sync.
- PreviewImage one was not reset on file read.
- Through RNA e.g., it was possible to ensure an ID icon via its preview image,
which was running code designed for custom previews/icons system, instead
of generating correct 'auto ID icon'.
This reverts commit f12204196f.
Campbell, sorry. have to revert this for the time being.
We've missed some very important bits, such as:
- FFmpeg is usually linked against OpenJPEG
- OIIO needs OpenJPEG as well.
For FFmpeg issues we can either disable OpenJPEG there (since
we don't really use it), or bump FFmpeg to version 3.1.1 which
can use either of OpenJPEG 1.5 or 2.1.
For OIIO we do need OpenJPEG support (otherwise Cycles will
not be able to use j2k/j2c textures) and currently there is
NO solution to make OIIO working with OpenJPEG 2.1.
According to Matthias Fauconneau (aka mfv) Larry is working
on the patch to get OIIO work with OpenJPEG 2.1, but it'll
take some time still.
I've tried to look into support of some sort of build system
flag and do ifdefs, but it all becomes quite nasty, especially
with bundled OpenJPEG bumped to 2.1.
Surely such an update is something we'll have to apply to
but at this exact moment it causes quite some pain for all
developers.
Suggest to wait for until OIIO supports OpenJPEG 2.1 and then
go with the updates for real.
Idea looked good, but we have too much custom situations here (some half-fake-sub-ID
being copied with their 'owner', animdata, etc.), let's let datablock copy functions
handle that themselves.
Also allows to safely call BKE_id_expand_local from all copy functions now (only when
copying linked data).
ntree is once more time a PITA - actually, all sub-IDs that get copied together with the 'main' one are
(shapekey was workedaround, as was animdata/action issue, but nodes are more touchy).
For now, better not to touch that, needs careful check & rethink.
Reason is, typically those actions datablock usercounts have already been adjusted in BKE_animdata_copy()
(called by generic iddata copy function).
Think this needs to be reworked a bit too, there are way too much ways to copy animdata currently,
it's rather confusing. But not the goal here, so for now we'll live with the hack!
will be used by both make_local() and copy() datablock functions.
Note that this new func make 'extern' all IDs used by localized datablock,
not only refcounted ones as it used to be in each type's functions (with a few exceptions).
Geometry shaders easily become performance bottlenecks, for in-depth explanation see e.g.
http://gamedev.stackexchange.com/a/48434http://rastergrid.com/blog/2010/09/history-of-hardware-tessellation/
The geometry shader was used so far for generating the fiber geometry (line strips)
from just the root vertices. Without a geometry shader every vertex has to be put into
a buffer and uploaded to the GPU (we could use a tesselation shader to cut down on size,
but that remains to be decided later if OpenGL 4.x is available).
To limit the size of this vertex buffer, the vertices contain only minimal necessary
information: an index for the fiber curve, and an interpolation curve parammeter in [0.0, 1.0].
The locations of vertices are interpolated in the vertex shader using texture lookups.
A range of textures encode all the necessary per-fiber attributes, in particular the
interpolation indices and weights wrt. control curves. These control curves are stored in
yet another texture (as is the case with the geometry shader too).
Initial performance seems to improve drastically without the geometry shader.
This way restrict can be used for CUDA and OpenCL as well.
From quick tests in areas i've been testing this it might give some
barely measurable %% of speedup, but it increases registers pressure.
So use of this qualifier is still really limited.
Using camel case for variables is something what didn't came from our original
code, but rather from third party libraries. Let's avoid those as much as possible.
Matches better naming of volume traversal files, where we've got
optimized versions of a single step of volume intersection and
traversal which will gather all volume intersections.
BVH traversal is not really that much a geometry and we've got
quite some traversals now. Makes sense to keep them separate in
the name of source structure clarity.
When an edge was collapsed, one of the vertices would be removed.
Edges attached to the deleted vertex wouldn't be considered for collapsing again,
making the outcome from collapsing edges depend on the edge-vertex order.
Use a hash to lookup the final vertex when collapsing edges, instead of skipping them.
The shader sync code used the anisotropic version of the Ashikhmin-Shirley closure for both
Anisotropic and Glossy BSDF, which caused a failed assert because the anisotropic closure
expects an addiitonal SVM node.
The orientation of the hair root ("root space") as well as a parametric
distance on the scalp to the primary deforming hair are passed to the
vertex shader. These value will allow deformation modifiers to be applied
for things like noise and clumping.
The vertex attributes for the fiber buffer use the gl_Vertex builtin variable
of GLSL instead of an explicit named attribute input. The corresponding index
then is -1. The following attributes still need to be disabled however.
Mostly pass bmain and do not check for NULL key, keys' make_local is
suspiciously simple in fact, but think until those behave like real
full-featured IDs, it's doing enough!
Available choices are the "classic" particle shading (bogus, but useful
for comparison), Kajiya, and Marschner. The Marschner model is not yet
implemented.
* Alt-G, Alt-R, Alt-S --> These don't clear delta transforms by default anymore,
making it possible to use these to properly store a "rest" pose
* Alt-Shift-G, Alt-Shift-R, Alt-Shift-S --> These WILL clear both the normal
transform and the delta, should you need to do so.
At first thought it was own recent work, but think issue is there since ages actually...
Basically, id_make_local() would always localize mesh/curve/lattice shapekeys, even in case
obdata localization actually made a local copy instead of localizing original datablock.
This was causing shapekeys being localized twice, and other odd nasty effects.
libquery now passes an extra flag info to the callback, in case that specific
ID usage is considered as indirect.
In most cases, it's just set from ID_IS_LINKED_DATABLOCK() result on datablock owner,
but in proxy object case we also consider ob->data, materials and pose usages as indirect.
Does not fixes all issues yet, but should already make behavior with proxy object saner.
Now using modern features from libquery/libremap areas.
Provides same kind of fixes/improvements as for BKE_object_make_local() (see rBd1a4ae3f395a6).
Note: this enlightened broken case of proxy objects regarding make_local
(and also whole remapping, in fact). Will be fixed in near future.
Not really happy with this, could lead to a **lot** of callback executions...
But libquery foreach looper must go over **all** ID pointers,
otherwise remapping & co is broken!
A possible fix for future would be to have 'image slots' defined at root of poly/facetex layers,
and just a (short) index in each face/poly item - would also save a reasonable amount of memory...
Now using modern features from libquery/libremap areas.
Also, it should handle much better cases where localized ID was also indirectly used by non-refcounting users
(typical case: object used as modifier/constraint/whatever target from another linked object, previous
code would not take those into account and just localize original object instead of making a local copy.
Would result in local object used by linked one, which would be partially 'undone' on next file reload... Crappy behavior).
And it fixes some obvious errors too (nullifying all proxy pointers unconditionnaly,
some missing refcounted usages cases in extern_local_object(), etc.).
Totally stupid to not pass it, and then let (some) BKE_foo_make_local() use G.main!
Note: unused for now, much more refactoring still to come in make_local area!
Now test_object_materials only handles one object. New test_all_objects_materials
checks the whole bmain->object list for cases where it is actually needed.
Should avoid some useless looping over all objects!
Main issue was that BKE_libblock_relink_ex was pretty much ignoring all those...
Also, unlinking of objects was not handling correctly indirect-related flags.
Refactored code into helper functions to avoid too much duplicated code.
Control strands are now drawn without subdivision. Control and fiber
strands get their own vertex buffers for this purpose.
The intermediate StrandData has been removed, because it is just
redundant storage of the draw data buffers. The results in StrandData
are not used for further editing, so there is really no point in
keeping them around.
Toggling edit-node after joining objects is redundant as far as I can see -
(dates back to first revision).
Also caused all edges to be drawn w/ meshes, redundant undo step w/ curves.
`my_memcmp` didn't work properly comparing memory sizes not aligned to 4 bytes,
this worked while we used guarded-alloc (which always wrote a guard at the end of each allocation).
Since moving to lockfree allocator it could read uninitialized memory.
It also consistently performed ~10-30% worse then glibc's.
This is typically well optimized, no need to do ourselves.
Scaling a bone on the Y axis (for a stretch effect), wasn't supported by the IK solver.
Support this by solving as uniform scaled bones by matching the X,Z axis to the Y.
Requested by @pepeland, see D2088 for details.
Not sure how useful it is to track this. Just making it consistent.
"bound" only stays in sync with GL if we are careful to use the
GPU_strand_shader_bind/unbind functions. Rogue calls to GPU_shader_bind
update GL's state without updating strand shader's "bound" flag.
This should allow us to avoid a lot of useless processing when iterating over the
whole main database (unlink/remap/usages checks/etc.).
Note that some ID types report they can use any type for now, due to
fuzzyness/indefined nature of some usages (like constraints/modifiers/game logic,
or ID pointer of nodes...). Maybe we could address this (like e.g. adding defines
in relevant headers to restrict ID types used by constraints, by modifiers, etc.).
But don’t think this is top priority for now.
This commit enables new unaligned BVH builder and traversal for scenes
with hair. This happens automatically, no need of manual control over
this.
There are some possible optimization still to happen here and there,
but overall there's already nice speedup:
Master Hair BVH
bunny.blend 8:06.54 5:57.14
victor.blend 16:07.44 15:37.35
Unfortunately, such more complexity is not really coming for free,
so there's some downsides, but those are within acceptable range:
Master Hair BVH
classroom.blend 5:31.79 5:35.11
barcelona.blend 4:38.58 4:44.51
Memory usage is also somewhat bigger for hairy scenes, but speed
benefit pays well for that. Additionally as was mentioned in one
of previous commits we can add an option to disable hair BVH and
have similar render time but have memory saving.
Reviewers: brecht, dingto, lukasstockner97, juicyfruit, maiself
Differential Revision: https://developer.blender.org/D2086
This commit implements traversal of unaligned BVH nodes.
QBVH traversal is fully SIMD optimized and calculates orientation
for all 4 children at a time, regular BVH might probably be optimized
a bit more.
This is a special builder type which is allowed to orient nodes to
strands direction, hence minimizing their surface area in comparison
with axis-aligned nodes. Such nodes are much more efficient for hair
rendering.
Implementation of BVH builder is based on Embree, and generally idea
there is to calculate axis-aligned SAH and oriented SAH and if SAH
of oriented node is smaller than axis-aligned SAH we create unaligned
node.
We store both aligned and unaligned nodes in the same tree (which
seems to be different from what Embree is doing) so we don't have
any any extra calculations needed to set up hair ray for BVH
traversal, hence avoiding any possible negative effect of this new
BVH nodes type.
This new builder is currently not in use, still need to make BVH
traversal code aware of unaligned nodes.
This seems to be straightforward way to support heterogeneous nodes
in the same tree.
There is some penalty related on 4gig limit of the address space now,
but here's are the thing:
Traversal code was already using ints to store final offset, so
there can't be regressions really.
This is a required commit to make it possible to encode both aligned
and unaligned nodes in the same array. Also, in the future we can use
this to get rid of __leaf_nodes array (which is a bit tricky to do since
trickery in pack_instances().
There are several internal changes for this:
First idea is to make __tri_verts to behave similar to __tri_storage,
meaning, __tri_verts array now contains all vertices of all triangles
instead of just mesh vertices. This saves some lookup when reading
triangle coordinates in functions like triangle_normal().
In order to make it efficient needed to store global triangle offset
somewhere. So no __tri_vindex.w contains a global triangle index which
can be used to read triangle vertices.
Additionally, the order of vertices in that array is aligned with
primitives from BVH. This is needed to keep cache as much coherent as
possible for BVH traversal. This causes some extra tricks needed to
fill the array in and deal with True Displacement but those trickery
is fully required to prevent noticeable slowdown.
Next idea was to use this __tri_verts instead of __tri_storage in
intersection code. Unfortunately, this is quite tricky to do without
noticeable speed loss. Mainly this loss is caused by extra lookup
happening to access vertex coordinate.
Fortunately, tricks here and there (i,e, some types changes to avoid
casts which are not really coming for free) reduces those losses to
an acceptable level. So now they are within couple of percent only,
On a positive site we've achieved:
- Few percent of memory save with triangle-only scenes. Actual save
in this case is close to size of all vertices.
On a more fine-subdivided scenes this benefit might become more
obvious.
- Huge memory save of hairy scenes. For example, on koro.blend
there is about 20% memory save. Similar figure for bunny.blend.
This memory save was the main goal of this commit to move forward
with Hair BVH which required more memory per BVH node. So while
this sounds exciting, this memory optimization will become invisible
by upcoming Hair BVH work.
But again on a positive side, we can add an option to NOT use Hair
BVH and then we'll have same-ish render times as we've got currently
but will have this 20% memory benefit on hairy scenes.
It was initially unsupported because initial idea of checking visibility
of all children was slowing scenes down a lot. Now the idea has changed
and we only perform visibility check of current node. This avoids huge
slowdown (from tests here it seems to be withing 1-2%, but more tests
would never hurt) and gives nice speedup of ray traversal for complex
scenes which utilized ray visibility.
Here's timing of koro.blend:
Without visibility check With visibility check
Original file 4min 20sec 4min 23sec
Camera rays only 1min 43 sec 55sec
Unfortunately, this doesn't come for free and requires extra data in
BVH node, which increases memory usage of BVH nodes by 15%. This we
can solve with some future trickery of avoiding __tri_storage created
for curve segments.
Those unwrap operators are a bit tricky, some are available from both 3DView and UVEditor, others only from 3DView...
Hacked around this by returning Mesh keymap for UV_OT ops for specific 3DView/MeshEditMode context.
This reverts commit 4fd78bb06f.
After further testing, it turns out that these options are less-broken than
I remember them being (and have been hearing about). Specifically, as long
as you disable all 3-axes of a transform component (i.e. all location, all
rotation, all scale) you're not likely to have problems, whereas if you only
disabled one axis (i.e. y-rotation), you may have problems in some cases.
So, restoring these to the UI.
Currently "long keyframes" are only useful for indicating where stationary
holds occur. If however you try to create a "moving hold" (where the values
are slightly different, but in terms of overall effect, it's still a hold)
then it could get tricky to keep track of where these occur.
Now it's possible to tag such keyframes (using the keyframe types - RKEY)
as being part of a moving hold. These will not only be drawn differently
from normal keyframes, but they will also result in a "long keyframe"
being drawn between each pair of them, just like if they had been completely
stationary instead.
Currently the theming/styling of these is a bit rough. They reuse the existing
theme colours for long keyframes.
* Added new operators to take the current transform value (loc/rot/scale or all 3)
and convert/apply that transform to a corresponding delta transform value. By default,
the transform value will be zeroed out again afterwards, so you don't end up with a
double transform.
* These operators can be found in the "Apply" menu (Ctrl-A)
* The "Animated Transforms to Deltas" (which does a similar job, except it adjusts all
existing animation data instead of the current transform) has also been moved to the
Apply menu (it was in the Transform menu instead)
A long requested feature has been to have objects appear in alphabetical order
in the animation editors, so that it is easier to find where they occur. This
commit implements support for this.
The main sticking point has been the performance impact of having this sorting
happening all the time (as the actual list of "bases" cannot be modified, as the
old depsgraph still needs random-looking unsorted order of objects for scheduling
updates). However, it recently occurred to me that perhaps by restricting it to
the one case where the ordering actually matters (i.e. when we're getting the channel
list for drawing all channels, vs operating on them), and adding a toggle to turn the
sorting off in heavy scenes when it might bog down things, that it will probably
be acceptable enough in general. Furthermore, if things get really bad, we can investigate
putting in place some sort of caching scheme for this too - though hopefully the
new depsgraph will make that unnecessary (i.e. it doesn't sort the bases directly anymore).
This commit introduces a scale factor setting for scaling all keyframe indicators
in the Dopesheet Editor up/down, in order to make them easier to select. It is perhaps
most useful for keyframe types which are usually indicated using smaller keyframes
(e.g. breakdown), which may get tricky to quickly select.
To make it easier to synchronise timing across multiple strips, if you add
markers locally to an action, these will show up in the NLA strip in the
NLA Editor. These markings can then be used to line up the start/end of
another strip, or even to make sure that the markers from two different
strips end up lining up.
By default, this is turned on, but it can be disabled (via the View menu)
if it adds too much visual noise.
This is a common simple shading model defined by Kajiya and Kay in
"Rendering fur with three dimensional textures." (1989)
This model in particular defines specular highlights for reflective
thin cylinders (strands), which gives a much more hair-like appearance
in the viewport. More advanced models (Marschner et. al.) improve this
further, in particular the diffuse term.
This is a terrible shading method, using the curve tangent as a "normal" vector.
With the old particle hair this was a compromise, because line drawing does not
support nicer shading in the fixed-function pipeline. With GLSL shaders this
is just a quick starting point, and useful for comparison with the old shading.
The strand system contains 2 levels of curve geometry:
- "Strands" are the abstract control strands, editable by the user and thus few in number.
- "Fibers" are renderable curves, generated based on those strands.
"Roots" now is used exclusively for hair follicles, which are attached to
the scalp mesh, i.e. follow the mesh deformation.
This breaks any post-versionning (like IPO conversion, python handler, etc.).
rB961ebfa8c40b9909 mentions some Main being do_versionned several times (which is not desired for sure),
will try to reproduce again and find another fix.
This will allow the edit mode to display render strands as well.
The roots may later be re-scattered or otherwise modified during editing,
and are eventually copied back to the Strands data when applying edits.
Since SDNA was allocated for each undo step,
the new address meant it was considered different and included again.
Add an option not to duplicate the DNA string when calling DNA_sdna_from_data,
as well as avoiding a redundant copy, it writes the same address each time.
Idea is to replace hard-to-track (id->lib != NULL) 'is linked datablock' check everywhere in Blender
by a macro doing the same thing. This will allow to easily spot those checks in future, and more importantly,
to easily change it (see work done in asset-engine branch).
Note: did not touch to readfile.c, since there most of the time 'id->lib' check actually concerns the pointer,
and not a check whether ID is linked or not. Will have a closer look at it later.
Reviewers: campbellbarton, brecht, sergey
Differential Revision: https://developer.blender.org/D2082
Now the object mode and edit mode for strands use the same drawing code.
Derived render strands are not yet supported (they may need re-scattering
during editing) and the tool settings have no effect yet.
Checking if faces exist was a bottleneck,
use a simpler version of this function for triangles.
Gives approx 1.6x overall speedup (when many edges are being collapsed).
When snapping to edge/vert, check the distance to the ray
instead of the screen-space pixel projection.
This also corrects the conversion of `dist_to_ray_sq` to `dist_px` which was being done incorrectly.
While this change was planned, it fixes T48791, caused by error in b01a56ee.
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.
As pointed by Brecht, previous fix in rB61b49de44940 was actually incomplete,
we could still hit float rounding issue and hence same value in more than one consecutive
items of element_sum.
New solution is a bit different, we remove the 'minimal weight' check, and rather simply
ignores an item when the sum of its normalized weight to previous item's sum does not add
anything. Shall be safe and 100% effective this time!
This patch is another step to achieve BI and it's Viewport consistency for cubemap textures.
{F318879}
To test world_space_shading flag D2072 is required.
Alexander (Blend4Web Team)
Reviewers: campbellbarton, brecht
Subscribers: homyachetser, Evgeny_Rodygin, AlexKowel, yurikovelenov
Differential Revision: https://developer.blender.org/D2074
This patch fixes shortcoming of D2046.
The original behavior without world_space_shading flag is that Texture node expects the reflected vector in view space. But with world_space_shading it should be in world space.
In attached file you will see a simple material setup and a node material analogue.
Simple material must have the same behavior regardless world_space_shading flag.
{F318866}
Alexander (Blend4Web Team)
Reviewers: brecht
Reviewed By: brecht
Subscribers: campbellbarton, homyachetser, Evgeny_Rodygin, AlexKowel, yurikovelenov
Differential Revision: https://developer.blender.org/D2072
OpenGL has a separate function for integer attribute pointers.
If the float attribute version is used, the attribute data will be
interpreted and floats and converted to integer values, which
garbles the content of any true int/uint buffer.
Environment lighting (aka ambient) is a key component of any renderer.
It's implemented like the Environment lighting of BI render for Approximate Gather mode. It support "Sky Color" and "White" Environment lighting modes.
It would be great if the user could see actual lighting conditions right in the Blender viewport instead of waiting for the renderer to complete the final image, exporting for external renderer or for a game engine.
Before:
{F113921}
After:
{F113922}
Example file: {F319013}
Original author: valentin_b4w
Alexander (Blend4Web Team)
Reviewers: valentin_b4w, campbellbarton, merwin, brecht
Reviewed By: brecht
Subscribers: panzergame, youle, duarteframos, AlexKowel, yurikovelenov, dingto, Evgeny_Rodygin
Projects: #rendering, #opengl_gfx
Differential Revision: https://developer.blender.org/D810
The geometry shader will need to do 2 texture lookups per control,
first for getting the control curve start and length, then for
reading the appropriate curve vertices.
This method uses the 3 closest points on the scalp for defining interpolation
weights. This method has drawbacks in cases of thin geometry, but it is
simple to implement.
These attributes carry information for interpolating guide strands
in the geometry shader. Each strand can have up to 4 guide strands
(identified by index) with a normalized weight for each.
This is in fact very hairy situation here... Objects are only refcounted by scenes,
any other usage is 'free', which means once all object instanciations are gone Blender
considers it can delete it.
There is a trap here though: indirect usages. Typically, we should never modify linked data
(because it is essencially useless, changes would be ignored and ost on next reload or
even undo/redo). This means indirect usages are not affected by default 'safe' remapping/unlinking.
For unlinking preceeding deletion however, this is not acceptable - we are likely to end with
a zero-user ID (aka deletable one) which is still actually used by other linked data.
Solution choosen here is double:
I) From 'user-space' (i.e. outliner, operators...), we check for cases where deleting datablocks
should not be allowed (indirect data or indirectly used data), and abort (with report) if needed.
II) From 'lower' level (BKE_library_remap and RNA), we also unlink from linked data,
which makes actual deletion possible and safe.
Note that with previous behavior (2.77 one), linked object would be deleted, including from linked data -
but then, once file is saved and reloaded, indirect usage would link back the deleted object,
without any instanciation in scene, which made it somehow virtual and unreachable...
With new behavior, this is no more possible, but on the other hand it means that in situations of dependency cycles
(two linked objects using each other), linked objects become impossible to delete (from user space).
Not sure what's best here, behavior with those corner cases of library linking is very poorly defined... :(
Also define single callback func typedef, cleaner this way!
Note: maybe we want to do that for the other callbacks too (data, etc.), but will be enough for now.
This is quite tricky situation which combined:
- Boolean modifier which accesses other object's derived mesh
(in fact, it's nothing to do with boolean modifier, any other
modifier which uses other's DM will have same bug).
- Dependency cycles in the scene which is rather russian-roulette
from the cycle solver point of view.
- Multiple instanced objects used as boolean operand.
With all this things combined boolean modifier was accessing operand's derived
mesh which was referencing data from Mesh datablock. The issue is that those
references are becoming invalid after EDBM_mesh_load().
This function already had code to make sure object itself does not end up with
dangling pointers from derived mesh. Make it now so no possible instanced objects
are left with dangling pointers.
And who said it's a good idea to reference something from derived mesh..
The Curves widget has buttons to zoom in on the curve. However the
click detection code doesn't take it into account, and at full zoom
in click on curve is detected very far from the actual visible curve.
Change it to compare the position to the actual line segments
in the UI coordinate space, i.e. with curve zoom applied.
This patch moves the PrintHardwareInfo() function in standalone only, not in embedded. Why? Because you can need this infos for debugging
purpose in "compiled" blender files but it was boring to have it displayed each time you launched embedded. So you can have this infos when you
click standalone or when you run your executable app from a console.
Reviewers: moguri, kupoman, panzergame.
This is equivalent to the DerivedMesh stage in the Mesh -> DM -> GPUBuffer
pipeline. The Strands struct is the base editable user data, StrandsData
is preprocessed data with evaluated root matrices and applied for object
space drawing.
changes in BLI_kdopbvh:
- `BLI_bvhtree_find_nearest_to_ray` now takes is_ray_normalized and scale argument.
- `BLI_bvhtree_find_nearest_to_ray_angle` has been added (use for perspective view).
changes in BLI_bvhutils:
- `bvhtree_from_editmesh_edges_ex` was added.
changes in math_geom:
- `dist_squared_ray_to_seg_v3` was added.
other changes:
- `do_ray_start_correction` is no longer necessary to snap to verts.
- the way in which the test of depth was done before is being simulated in callbacks.
These curves are NOT intended for final rendering! They are abstract
editable data (control strands) from which the "actual" visible strands
have to be generated in some way.
This is largely equivalent to the parent/child dichotomy in old particles,
but enforces the separation right from the start (control curves are never
rendered).
Regression caused by own changes to make texture paint more efficient
from workflow point of view.
Now the idea is to use vertex color outside of paint mode, so we don't
break any compatibility here.
For simple cases bitmasks were OK, but didnt work for vert/edge, vert/edge tests.
Tag verts instead, makes logic easier to follow and gives minor speedup.
Make sure we don't perform any implicit address space conversion.
A bit annoying, but less intrusive approaches (like using temp private
variable in .cl kernel) do not work correct here.
Using generic address space will help from code side here, but will
be somewhat slower due to extra things happening as far as i know.
So the error seems to be in cubic_tangent_factor_circle_v3(),
which was introduced with D2001.
I've tweaked the most obvious culprit here - the epsilon factor.
It used to be 10^-7, but I've reduced it down to 10^-5 now,
and it's looking a lot more stable now :)
---------
BTW, about the derivation of the magic 0.390464 factor I briefly subbed back
as a workaround for this bug, see:
http://www.whizkidtech.redprince.net/bezier/circle/
It's probably some numeric precision issue, but until we figure out exactly what's
going wrong here, let's just revert back to the hardcoded value that was used here
successfully for years without issues.
Remapping indirect usage of IDs is forbidden from user space, this is calling for
nice nightmare with libraries handling (and undo crash, among other things).
Not sure why I was 'laxist' about indirect usage cases detection like that,
for now just consider any ID used by another linked datablock as indirect usage case!
Also, added some error/warning reports to Outliner's remap code.
The "W" channel will get a yellowish colour (i.e. a blend between the X/R and
Y/G axis colours), while the XYZ will behave as they do for other transforms.
The create drivers operator should not use the "Transform Channel" variable type
when driving one transform with another on the same object/bone. Otherwise, you
end up with a situation which technically results in a bit of a pseudo-dependency
cycle.
The Ctrl-Shift-C operator to add constraints between a pair of selected items,
for example, between two objects, or between two bones (in the same armature).
This commit makes it possible to use this operator to add a constraint where the
target is a bone from another object - e.g. to make a deform bone follow the control
bone in another armature, or to make an object use a bone as a tracking target.
Usage:
1) Ensure you are in Pose Mode, then select the bone to use as the target
2) Shift-Select the other object and/or the bone that's going to get the constraint
3) Ctrl-Shift-C
Problem was in fact slightly wider, File space was nearly not taking into account
library navigation case and its 'virtual' directoris, except in a few places.
Add a wrapper around BLI_is_dir that also check for lib paths, and used it in
ED_file_change_dir(), such that we now always check path is a
valid directory - in the filebrowser context, not filesytem context. ;)
As far as I can see, the second issue there was that the functions receive a pointer to a member variable of the
ShaderData, which is stored in global memory. However, this means that the pointer points to global memory as well,
therefore OpenCL requires the ccl_addr_space "keyword" in front of the pointer.
With this commit, the OpenCL kernels build on Linux with the Intel CPU OpenCL runtime - however, they already did
without the change and I don't have an AMD card, so I can't really test whether the AMD runtime is happy as well now.
Those (one per ID type!) were uselessly duplicated, and badly inconsistent
(some types were actually unlinking before deletion, others were only working if already unlinked!).
Now we use same func and same API for all types, by default deletion is performed only if ID is no more used,
set `do_unlink` parameter to True to always delete ID even if still in use.
Only exception now is with Scene, since we always want to keep at least one!
Note that this will change default behavior of some types (since unlinking is never done anymore by default).
Using a DerivedMesh directly for mesh data isn't so great because
the lifetime of a DM is not necessarily known at compile time, depending
on branching. node_counted_ptr is basically a simplified shared_ptr variant.
Exiting Blender during argument parsing would leak memory
(tests, documentation generation, utilities).
While harmless, it hides real leaks which should be resolved.
This commit adds a new distribution to the Glossy, Anisotropic and Glass BSDFs that implements the
multiple-scattering microfacet model described in the paper "Multiple-Scattering Microfacet BSDFs with the Smith Model".
Essentially, the improvement is that unlike classical GGX, which only models single scattering and assumes
the contribution of multiple bounces to be zero, this new model performs a random walk on the microsurface until
the ray leaves it again, which ensures perfect energy conservation.
In practise, this means that the "darkening problem" - GGX materials becoming darker with increasing
roughness - is solved in a physically correct and efficient way.
The downside of this model is that it has no (known) analytic expression for evalation. However, it can be
evaluated stochastically, and although the correct PDF isn't known either, the properties of MIS and the
balance heuristic guarantee an unbiased result at the cost of slightly higher noise.
Reviewers: dingto, #cycles, brecht
Reviewed By: dingto, #cycles, brecht
Subscribers: bliblubli, ace_dragon, gregzaal, brecht, harvester, dingto, marcog, swerner, jtheninja, Blendify, nutel
Differential Revision: https://developer.blender.org/D2002
The RNA properties are still there (in case you really need them), except now
they will not be shown in the UI anymore, as this constraint really didn't
work well/at all when any of those was disabled. Most people shouldn't really
need to worry about this change.
If anyone wants a matrix-math challenge, they're welcome to try getting those
working for real, so that we can show these toggles again.
To get this working the least effort, I've had to expose the helper functions
used by the lasso and circle select keyframe-test callbacks (which are generic)
and expose them for use by the GP keyframe editing code too. Hopefully in time
we clean this all up and just write the code once to operate on "keyframes"
This only works in the Action and Dopesheet modes (which operate on FCurve keyframes).
Support for Grease Pencil and Mask Keyframes though is still pending.
As suggested by mendiobox:
* Don't show "enable editing" in the 3D view. You can already do this by switching
into stroke editing mode here, so no need for the duplication. (In other editors
though, this can't be done yet, so we don't do it)
* Make the "Convert" button into a dropdown so that you don't need to deal with a
a separate popup menu
* In the 3D view, don't show the selection + transform operators that can be easily
found in the menus too (as well as having commonly used shortcuts)
This new operator will delete any GP frame it finds on the current frame, regardless
of whether it's on the active layer or not. It will only remove the frames if the
layer is editable, but otherwise, it will just go for it.
The existing operator is great for use in the panel (where it only applies to the active
frame), but it was not so good for all the other places where tools can be invoked
(e.g. D-X, or Delete) as you're typically thinking about the whole scene more holisticaly
than just caring about a particular layer.
In rare cases intersect would attempt to add edges with the same vertex twice
from edge-vert / edge-edge intersections.
Solve by checking for duplicates when creating vertex-array for these types of intersections
(always under 3x comparisons, so not much overhead).
Node graph inputs/outputs are now defined statically and graph instances
can be constructed by passing these as ArrayRefs. The code for compiling
various node graph types can then be unified.
The function that assigns names to socket types missed an entry, therefore all entries after it were mapped to the wrong name.
Long-term, it might be a better solution to use a map to avoid issues like these, but for now this fix works.
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 adds operators and Outliner menu entries to reload or relocate a library,
and to delete or replace a datablock.
RNA ID API is also extended to allow ID deletion and remapping from python.
Review task: D2027 (https://developer.blender.org/D2027).
Reviewed by campbellbarton, thanks a bunch.
This commit changes a lot of how IDs are handled internally, especially the unlinking/freeing
processes. So far, this was very fuzy, to summarize cleanly deleting or replacing a datablock
was pretty much impossible, except for a few special cases.
Also, unlinking was handled by each datatype, in a rather messy and prone-to-errors way (quite
a few ID usages were missed or wrongly handled that way).
One of the main goal of id-remap branch was to cleanup this, and fatorize ID links handling
by using library_query utils to allow generic handling of those, which is now the case
(now, generic ID links handling is only "knwon" from readfile.c and library_query.c).
This commit also adds backends to allow live replacement and deletion of datablocks in Blender
(so-called 'remapping' process, where we replace all usages of a given ID pointer by a new one,
or NULL one in case of unlinking).
This will allow nice new features, like ability to easily reload or relocate libraries, real immediate
deletion of datablocks in blender, replacement of one datablock by another, etc.
Some of those are for next commits.
A word of warning: this commit is highly risky, because it affects potentially a lot in Blender core.
Though it was tested rather deeply, being totally impossible to check all possible ID usage cases,
it's likely there are some remaining issues and bugs in new code... Please report them! ;)
Review task: D2027 (https://developer.blender.org/D2027).
Reviewed by campbellbarton, thanks a bunch.
In the OSL node compilation code for the Environment Texture, is_linear was used as a socket.
However, there was no socket for it, which caused Blender to crash.
Adding a socket doesn't really make sense since it's an internal value and not a parameter
of the node, so it now just uses the variable directly.
Getting a new edit-derived-bmesh was always creating a deform-vert array, even when it wasn't needed.
Since this was called on redraw, in many cases it was doing it unnecessarily.
Now pass in a custom-data mask and only fill in deform-verts when needed.
Gives noticeable drawing speedup (~10-30% here).
The problem here was that there are five path types internally (diffuse, glossy, transmission, subsurface and volume scatter), but subsurface isn't exposed to the user.
This caused some weird behaviour - if all four types are disabled on the lamp, Cycles doesn't even try sampling it, but if any type was active, the lamp would illuminate
the cube since none of the options set subsurface to zero.
In the future, it might be reasonable to add subsurface visibility as an option - but for now the weird and inconsistent behaviour can be fixed simply by setting both
diffuse and subsurface to zero if the user disables diffuse visibility.
A new option for Font/Text objects vertical alignment:
* Top Base-Line (current mode)
* Top
* Center
* Bottom
The Top is the equivalent as the Top-Baseline with an empty line at the begin of the
text. It's nice to have this option too though, since if we are driving
the alignment via Python we don't want to add extra lines to the text
only to accomodate to the desired vertical alignment.
The Center and Bottom are as intuitive as their name suggest.
When working with text boxes, the vertical alignment only work for
paragraphs that are not vertically full.
Many thanks to Campbell Barton (ideasman42 / @campbellbarton) for the
code review, code comments, and overall suggestions and changes :)
Reviewers: campbellbarton
Differential Revision: https://developer.blender.org/D2061
BM_face_split_edgenet wasn't correctly detecting boundary vertices to walk over,
since vertices may be attached to boundary edges not part of the newly created face.
rB046adde64f16 was actually pretty useless (and broken), since issue ends up not being
in binary search code, but in generation of the 'summed weights' array used to distribute
particles over mesh items - looks like very small weights could lead to null accumulated
weights, wich was breaking binary search.
Fixed simply by adding a minimal, non-zero weight for mesh items to be allowed to emit particles.
Hopefully we are done with this distribution mess!
It is not possible to use a set split by name as valid input to
check_node_input_traversed - it needs a complete set of all nodes visited so
far. On the other hand, the merge comparison loop should only check nodes that
were not just visited, but found unique. This means that there should really be
two separate data structures.
Without the fix, check_node_input_traversed actually never returns true, so
only nodes without any inputs are processed.
For non-branched path tracing with a GTX 960 and CUDA 7.5, this gives a small reduction
in stack usage but mainly: 8% faster render on BMW, 5% on pabellon, 13% on classroom.
This is an initial commit for half texture support in Cycles.
It adds the basic infrastructure inside of the ImageManager and support for these textures on CPU.
Supported:
* Half Float OpenEXR images (can be used for e.g HDRs or Normalmaps) now use 1/2 the memory, when loaded via disk (OIIO).
ToDo:
Various things like support for inbuilt half textures, GPU... will come later, step by step.
Part of my GSoC 2016.
The main compiler class is now itself independent of LLVM and can be used for
future backends. The implementation of specific behavior (values, statements, etc.)
happens through a CodeGenerator implementation. This also uses aggregation rather
than inheritance on the compiler class part (which is not a virtual class itself).
Some area were still not in sync with readfile.c, now should be better.
Note that readfile.c has been used as référence here re us refcounting,
not sure how accurate it is, time will say :|
Optimization in binary search could lead to equality instead of expected strictly greater than value.
Harmless but noisy, and better be strict here.
reported by sergey on irc (with koro cycles benchmark file), thanks.
Rather than a single socket value map, use a "scope" to keep track of which
nodes are already generated. This reintroduces the "block" idea from earlier
versions, but at the later codegen stage rather than the node graph.
Describing scope on the level of nodes becomes quite clumsy and ambiguous,
and moves nodes further away from the UI representation. The compiler can
handle this more elegantly without requiring persistent data.
Scopes can be nested. This is not used yet, but will allow scoped variables
later on, where the variables being used for a node path depend on its scope.
In some cases when:
* the viewport was in the camera mode
* the camera was ortho
* the view was not fitting (as oppose to use HOME)
region_2d_to_origin_3d would misbehave (and consequently region_2d_to_location_3d).
Sample addon to test it:
```
import bpy
from bpy_extras.view3d_utils import (
region_2d_to_location_3d,
)
from mathutils import (
Vector,
)
class MoveXYOperator(bpy.types.Operator):
"""Translate the view using mouse events"""
bl_idname = "view3d.move_xy"
bl_label = "Move XY"
@classmethod
def poll(cls, context):
return context.object
def modal(self, context, event):
if event.type == 'MOUSEMOVE':
self.move(context, event)
elif event.type in {'LEFTMOUSE', 'RIGHTMOUSE', 'ESC'}:
return {'FINISHED'}
return {'RUNNING_MODAL'}
def invoke(self, context, event):
if context.space_data.type == 'VIEW_3D':
self.ob = context.object
context.window_manager.modal_handler_add(self)
return {'RUNNING_MODAL'}
else:
self.report({'WARNING'}, "Active space must be a View3d")
return {'CANCELLED'}
def move(self, context, event):
xy = region_2d_to_location_3d(
context.region,
context.space_data.region_3d,
(event.mouse_region_x, event.mouse_region_y),
Vector(),
).xy
self.ob.location.xy = xy
def register():
bpy.utils.register_class(MoveXYOperator)
def unregister():
bpy.utils.unregister_class(MoveXYOperator)
if __name__ == "__main__":
register()
```
Factorized writing of ID block's data (so far, only IDProps) into own helper func.
This also fixes missing IDProp (aka custom data) saving from GreasePencil and Library
datablocks (and add comment about why we do not save WM IDProps).
Finaly, it ensures all ID-related data are written immediately after the ID itself
(was not the case for all data types previously, some were writting their own data
before IDProps). This is not a fix (.blend file format does not enforce any order
on sub-data of datablocks, they only have to be after datablock struct itself in file),
but makes things more consistent.
Before the intervall was set to 0.125 which effectively resulted in 8 positions across the UV space (per axis).
I halved that value, holding shift enables an even finer movement.
This change was ported over from my soc-2016-uv_tools branch after talking with howardt, ideasman42 and hackerman-
When this flag is set - even when the curve error is under the threshold,
keep attempting a better fit.
Enable this for freehand drawing, since it gives nicer results and isn't noticeably slower.
Also, can use numeric input to set segments and profile when
in those respective value-adjusting modes (as per S or P toggle).
Finally, fixed problem with previous bevel commit: when changing
value-adjusting mode, would like to start off resumed value adjustment
where it was before.
Unsigned int is not supported by OSL as far as i concerned, so should not
really matter here. However, might be wrong and perhaps more proper idea
would be so set it as regular int?
Auto-merge caused all edges between selected vertices to be selected.
This only makes sense in vertex-select-mode.
Correct edge-flag merging code, which now merges flags from multiple edges.
Short story: draw_lamp would add itself to delayed transp drawing list from 'xray' drawing step.
This was broken, since delayed transp drawing list is always handled **before** delayed xray one.
After undo it lead to segfault crash, v3d->afterdraw_transp still having reference to old freed scene's base.
Also added asserts that those afterdraw list are empty at end of drawing step, should help
avoiding that kind of issue in future.
Add a new fallback method that uses offset distance from the curve to the line between both points,
for freehand drawing it typically only fives minor improvements (1-3% fewer points),
for curve dissolve the improvements are more noticeable.
D1988 by @wisaac, with own edits and improvements.
This improves on existing tangent calculation functions too.
- BM_face_calc_tangent_auto: Chooses method based on number of sides, used by manipulator (not exposed to Python).
- BM_face_calc_tangent_edge: from longest edge.
- BM_face_calc_tangent_edge_pair: from longest edge-pair (most useful with quads).
- BM_face_calc_tangent_edge_diagonal: edge farthest from any vertex.
- BM_face_calc_tangent_vert_diagonal: vert farthest from any vertex.
Also optimize BM_vert_tri_calc_tangent_edge* functions to avoid sqrt.
This was originally supported, however relative links to examples & templates made it fail.
Now files in the source tree are copied to the build-dir, with ".." replaced with "__"
to avoid having to mirror Blender's source-layout in the Sphinx build-dir.
Also skip uploading the built docs when an SSH user-name isn't passed to sphinx_doc_gen.sh
instead of aborting (so people w/o SSH access to our servers can use the shell-script).
You can capture and stream video in the BGE using the DeckLink video
cards from Black Magic Design. You need a card and Desktop Video software
version 10.4 or above to use these features in the BGE.
Many thanks to Nuno Estanquiero who tested the patch extensively
on a variety of Decklink products, it wouldn't have been possible without
his help.
You can find a brief summary of the decklink features here: https://wiki.blender.org/index.php/Dev:Source/GameEngine/Decklink
The full API details and samples are in the Python API documentation.
bge.texture.VideoDeckLink(format, capture=0):
Use this object to capture a video stream. the format argument describes
the video and pixel formats and the capture argument the card number.
This object can be used as a source for bge.texture.Texture so that the frame
is sent to the GPU, or by itself using the new refresh method to get the video
frame in a buffer.
The frames are usually not in RGB but in YUV format (8bit or 10bit); they
require a shader to extract the RGB components in the GPU. Details and sample
shaders in the documentation.
3D video capture is supported: the frames are double height with left and right
eyes in top-bottom order. The 'eye' uniform (see setUniformEyef) can be used to
sample the 3D frame when the BGE is also in stereo mode. This allows to composite
a 3D video stream with a 3D scene and render it in stereo.
In Windows, and if you have a nVidia Quadro GPU, you can benefit of an additional
performance boost by using 'GPUDirect': a method to send a video frame to the GPU
without going through the OGL driver. The 'pinned memory' OGL extension is also
supported (only on high-end AMD GPU) with the same effect.
bge.texture.DeckLink(cardIdx=0, format=""):
Use this object to send video frame to a DeckLink card. Only the immediate mode
is supported, the scheduled mode is not implemented.
This object is similar to bge.texture.Texture: you need to attach a image source
and call refresh() to compute and send the frame to the card.
This object is best suited for video keying: a video stream (not captured) flows
through the card and the frame you send to the card are displayed above it (the
card does the compositing automatically based on the alpha channel).
At the time of this commit, 3D video keying is supported in the BGE but not in the
DeckLink card due to a color space issue.
defines a uniform that reflects the eye being rendered in stereo mode:
0.0 for the left eye, 0.5 for the right eye.
In non stereo mode, the value of the uniform is fixed to 0.0.
The typical use of this uniform is in stereo mode to sample stereo textures
containing the left and right eye images in a top-bottom order.
python:
shader = obj.meshes[0].materials[mat].getShader()
shader.setUniformEyef("eye")
shader:
uniform float eye;
uniform sampler2D tex;
void main(void)
{
vec4 color;
float ty, tx;
tx = gl_TexCoord[0].x;
ty = eye+gl_TexCoord[0].y*0.5;
// ty will be between 0 and 0.5 for the left eye render
// and 0.5 and 1.0 for the right eye render.
color = texture(tex, vec2(tx, ty));
...
}
The assembler version in Windows used to return the previous value
of the variable while all the other versions return the new value.
This is now fixed for consistency.
Note: this bug had no effect on blender because no part of the code
use the return value of these functions, but the future BGE DeckLink
module makes use of it to implement reference counter.
bge.logic.setRender(flag) to enable/disable render.
The render pass is enabled by default but it can be disabled with
bge.logic.setRender(False).
Once disabled, the render pass is skipped and a new logic frame starts
immediately. Note that VSync no longer limits the fps when render is off
but the 'Use Frame Rate' option in the Render Properties still does.
To run as many frames as possible, untick the option
This function is useful when you don't need the default render, e.g.
when doing offscreen render to an alternate device than the monitor.
Note that without VSync, you must limit the frame rate by other means.
fbo = bge.render.offScreenCreate(width,height,[,samples=0][,target=bge.render.RAS_OFS_RENDER_BUFFER])
Use this method to create an offscreen buffer of given size, with given MSAA
samples and targetting either a render buffer (bge.render.RAS_OFS_RENDER_BUFFER)
or a texture (bge.render.RAS_OFS_RENDER_TEXTURE). Use the former if you want to
retrieve the frame buffer on the host and the latter if you want to pass the render
to another context (texture are proper OGL object, render buffers aren't)
The object created by this function can only be used as a parameter of the
bge.texture.ImageRender() constructor to send the the render to the FBO rather
than to the frame buffer. This is best suited when you want to create a render
of specific size, or if you need an image with an alpha channel.
bge.texture.<imagetype>.refresh(buffer=None, format="RGBA", ts=-1.0)
Without arg, the refresh method of the image objects is pretty much a no-op, it
simply invalidates the image so that on next texture refresh, the image will
be recalculated.
It is now possible to pass an optional buffer object to transfer the image (and
recalculate it if it was invalid) to an external object. The object must implement
the 'buffer protocol'. The image will be transfered as "RGBA" or "BGRA" pixels
depending on format argument (only those 2 formats are supported) and ts is an
optional timestamp in the image depends on it (e.g. VideoFFmpeg playing a video file).
With this function you don't need anymore to link the image object to a Texture
object to use: the image object is self-sufficient.
bge.texture.ImageRender(scene, camera, fbo=None)
Render to buffer is possible by passing a FBO object (see offScreenCreate).
bge.texture.ImageRender.render()
Allows asynchronous render: call this method to render the scene but without
extracting the pixels yet. The function returns as soon as the render commands
have been send to the GPU. The render will proceed asynchronously in the GPU
while the host can perform other tasks.
To complete the render, you can either call refresh() directly of refresh the texture
to which this object is the source. Asynchronous render is useful to achieve optimal
performance: call render() on frame N and refresh() on frame N+1 to give as much as
time as possible to the GPU to render the frame while the game engine can perform other tasks.
Support negative scale on camera.
Camera scale was previously ignored in the BGE.
It is now injected in the modelview matrix as a vertical or horizontal flip
of the scene (respectively if scaleY<0 and scaleX<0).
Note that the actual value of the scale is not used, only the sign.
This allows to flip the image produced by ImageRender() without any performance
degradation: the flip is integrated in the render itself.
Optimized image transfer from ImageRender to buffer.
Previously, images that were transferred to the host were always going through
buffers in VideoTexture. It is now possible to transfer ImageRender
images to external buffer without intermediate copy (i.e. directly from OGL to buffer)
if the attributes of the ImageRender objects are set as follow:
flip=False, alpha=True, scale=False, depth=False, zbuff=False.
(if you need to flip the image, use camera negative scale)
Thai font is a complex script that assumes full featured unicode layout engine,
while Blender only knows about basic kerning (offset of a char based on the previous one).
So this commit edits Thai part of our i18n font to fix the very bad spacing of thai chars
we had in Blender so far.
Work done by Hồ Châu, many thanks!
A new option '-a' can be passed to the blenderplayer. It forces the
framebuffer to have an alpha channel.
This can be used in VideoTexture to return a image with alpha channel
with ImageViewport (provided alpha is set to True on the ImageViewport
object and that the background color alpha channel is 0, which is the
default).
Without the -a option, the frame buffer has no alpha channel and
ImageViewport always returns an opaque image, no matter what.
In Linux, the player window will be rendered transparently over
the desktop.
In Windows, the player window is still rendered opaque because
transparency of the window is only possible using the 'compositing'
functions of Windows. The code is there but not enabled (look for
WIN32_COMPOSITING) because 1) it doesn't work so well 2) it requires
a DLL that is only available on Vista and up.
give precedence to AA over Swap copy:
Certain GPU (intel) will not allow MSAA together with swap copy.
Previously, swap copy had priority over MSAA: fewer AA samples would be
chosen if it was the condition to get swap copy. This patch reverse the
logic: swap copy will be abandonned if another swap method (undefined or
exchange) will provide the number of AA samples requested. If no AA
samples is requested, swap copy still has the priority of course.
The purpose of the patch is to replace deprecated glShadeModel.
To decrease glShadeModel calls I've set GL_SMOOTH by default
Reviewers: merwin, brecht
Reviewed By: brecht
Subscribers: blueprintrandom, Evgeny_Rodygin, AlexKowel, yurikovelenov
Differential Revision: https://developer.blender.org/D1958
Regression from recent rB2c5dc66d5effd4072f438afb, if last item of last chunk of a mempool was valid,
it would not be returned by mempool iterator step, which would always return NULL in that case.
The basic idea is that a node represents a function F, without assuming anything about
how the input expressions are evaluated:
F = F()
In most common case this function can be decomposed and defined in terms of concrete values,
which are given by the input expressions:
F = F'(G1(), G2(), ...), where Gn are the respective inputs of the node
This is a first *very* naive implementation. It would rely heavily on common-subexpression-
elimination (CSE) to avoid generating the same code paths over and over again. There should
probably be some initial optimization on the BVM node level instead of relying just on LLVM passes.
The purpose of this generalization is to allow more complicated control flow inside node functions.
Nodes would then also be able to take additional variable arguments (e.g. an index), which can
be modified before being used as internal arguments for input expressions.
It's still not completely working - there are still some glitches - but far better than before.
To make buttons of the uiList work, you have to add a 'check' callback to the operator that invokes the menu. Only if it returns True, the uiList gets refreshed. To avoid this we have to make the region refresh tagging in the entire button handling a bit smarter.
Changes I had to do:
* Call uiList handling from menu/popup handling if needed.
* Make uiList handling use special popup refresh tag if placed in menu.
* Allow popups invoked from py operator to tag for refresh by using operator 'check' callback.
* Tag popup for refresh when resizing uiList.
Mostly fixes T48612.
Don't store maximum length of text per undo step,
or attempt to pack all data in a single array.
Was storing 32766 characters per undo step, irrespective of actual text length.
The way cloth is coded, structural springs are only effective when stretched, while bending springs act only when shrunk. However, when cloth is exactly in its rest shape, neither have any effect, and effectively don't exist for the implicit solver.
This creates a stability problem in the initial frames of the simulation, especially considering that gravity seems to act so precisely that it doesn't disturb the strict equality of lengths, so in parts of the cloth this springless state can continue for quite a while.
Here is an example of things going haywire because of this and some suspicious logic in collision code acting together: {F314558}
Changing the condition so that structural springs are active even at exactly rest length fixes this test case. The use of >= is also supported by the original paper that the cloth implementation in blender is based on.
Reviewers: lukastoenne
Reviewed By: lukastoenne
Projects: #bf_blender
Differential Revision: https://developer.blender.org/D2028
This is just the very basic version of the sparse convolution noise
with a Gabor kernel. It does not yet have derivatives or anisotropic
filtering. The only frequency spectrum implemented atm is isotropic
band-limited noise.
All text was displaying black.
BLF uses alpha-only textures which aren't supported by the basic-shader,
Workaround this by using texture swizzle so the RGB components of the texture are set to 1.
This hopefully fixes T48383 by avoiding two numerical problems that I found in the volume code.
Reviewers: sergey, dingto, brecht
Reviewed By: sergey, dingto, brecht
Maniphest Tasks: T48383
Differential Revision: https://developer.blender.org/D2051
It's not obvious that they use the mouse coordinate of the initial key-press event (behavior since rBf1f33ba7be2d), so added comment.
Also corrected other comments.
OpenCL seems to work fine here, and for some reason that comparison was
giving compilation error on OpenCL here.
Better to compile OpenCL kernel than to be fully robust to weird corner
cases.
Newly computed custom normals were forgotten during poly flipping, leading
to wrong custom normals being assigned to wrong loop...
Dead simple, but was tough to track down this one!
At the moment light shading in Blender is produced in viewspace. Apparently, that's why
shader nodes work with normals in camera space. But it is not convenient for artists.
The more convenient approach is implemented in Cycles where normals are represented in world space.
Blend4Web Team designed the engine keeping in mind shader parameters readability,
so normals are interpreted in world space as well. And now our users have to use some tweaks, like
empty node group with the name "Replace", which is replacing one input by another on the engine side
(replacing working configuration in Blender Viewport by the configuration that has the same behavior in the engine).
This patch adds the ability to switch to world space for normals and lamp vector in BI and Viewport.
This patch is very important to us and we crave to see this patch in Blender 2.7 because
it will significantly simplify Blend4Web material creation workflow.
{F315547}
{F315548}
Reviewers: campbellbarton, brecht
Reviewed By: brecht
Subscribers: homyachetser, Evgeny_Rodygin, AlexKowel, yurikovelenov
Differential Revision: https://developer.blender.org/D2046
Was using O(n^2) lookup on ID's with undo.
This caused undo to hang with 1000's of data-blocks
(especially with heavy scenes & outliner-space, which doesn't even need to be visible to cause a slow-down).
Internally this uses a ghash per id-type, which is lazy-initialized.
Each key uses the name and library since there may be name collisions between libraries.
Developer Notes:
- Adds small `BKE_main_idmap_*` API.
- Needed to change linking order for this to build.
The original quad intersection test works by just testing against the two triangles that define the quad.
However, in this case it's actually faster to use the same test that's also used for portals: Determining
the distance to the plane in which the quad lies, calculating the hitpoint and checking whether it's in the
quad by projecting onto the sides.
Reviewers: brecht, sergey, dingto
Reviewed By: dingto
Differential Revision: https://developer.blender.org/D2045
Being granular means we need to re-build depsgraph a bit more often..
The issue was caused by rigidbody requiring some special nodes to
handle physics which were not created with just tagging object for
update.
The issue was caused by non-watertight nature of intersection, which is now addressed.
Hopefully it doesn't cause any regression caused by uninitialized precalculated storage.
Makes behavior of proxy_from backlink working similar to the old dependency graph.
it's nasty, but needed here in the studio to get proxies fixes ASAP.
In certain types of animation it's possible to have some objects
scaling to zero. In this case we can save render times by avoid
traversing such instances.
Better to do ti ahead of a time, so traversal stays simple.
Reviewers: lukasstockner97, dingto, brecht
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D2048
Currently for windows only, this is an initial commit towards native
support of NUMA.
Current commit makes it so Cycles will use all logical processors on
Windows running on system with more than 64 threads.
Reviewers: juicyfruit, dingto, lukasstockner97, maiself, brecht
Subscribers: LazyDodo
Differential Revision: https://developer.blender.org/D2049
FCurve evaluation depended on FCurve.curval, which isn't threadsafe.
Now only use this value for debug display,
and pass the value instead of storing in the FCurve for all but debug-display.
This covers the following workflow. Animator will disable Update All 3D views
in order to get quick playback in a particular 3D editor. However, it also
handy to see playback in all editors and image editors to see comparison of
animation and reference footage.
So the idea here is to refresh reasonable editors when playback is invoked
from animation editors.
Commit to make Hjalti happy.
Still not sure how to properly solve the issue, needs some trickery to get
actual optimized values from intersection function (using printf() avoids
some optimization and makes stuff render correct).
For the time being let's just simplify check.
- the name of the enumerator `SNAP_NOT_OBEDIT` was changed to `SNAP_NOT_ACTIVE`.
- the parameter `snap_to_flag` was moved to outside `SnapObjectParams`.
- the member `use_object_edit` was renamed to `use_object_edit_cage`.
- added the arg `params` in `ED_transform_snap_object_project_ray`.
- simplifications in the loop of the function `snapObjectsRay`.
That was a nice and funny hunt, albeit rather time consumming!
To summarize, so far code was using a static global gpu_buffer for pbvh vbo drawing
of 'grid' types (multires mostly?).
There were two issues here:
1) Global gpu buffer was assigned to GPU_PBVH_Buffers->index_buf, but then nearly no
check was done when freeing that buffer, to ensure we were not freeing the global one
(not totally sure this one was actually causing any issue, but was bad and unsafe anyway).
Was solved by adding a flag to GPU_PBVH_Buffers to indicate when we are using some
'common' buffer here, which freeing is handled separately.
2) Main issue: if several multires objects in sculpt mode with different grid size
were present simultaneously, the global gpu buffer had to be resized for each object draw
(i.e., freed and re-allocated), but then the pbvh nodes from other objects storing freed reference
to that global buffer had no way to know that it had been freed, which was causing the segfault & crash.
Was solved by getting rid of that global buffer, and instead allocating one 'grid_commmon_gpu_buffer' per pbvh.
Told ya baby, globals are *PURE EVIL*!
This seems to be illegal and not productive anyway. Do it ahead of
a time now, which solves shading issues in edit mode and prevents
assert from happening.
Issue here is that for Blender, any pass which name starts with 'Diffuse' is considered
a diffuse pass - and it does not really support several passes of the same type in renderresult.
So for now, passtype_from_name() is now checking whether a pass of same type already exists
in render layers, and return 0 (uknown passtype) in this case.
Also standard functions for copying and zeroing values are now
implemented per type rather than relying on a simple load/store.
This will become more important with larger aggregate types such as matrix44.
Some adjustments to how bevel edge 'profiles' are adjusted in some
cases. For the bug fix, wanted to handle cases of middle of three
coplanar beveled cases to make profile curve rather than linear
interpolate.
Also undid an earlier decision to make profile plane be perpendicular
to beveled edge i the non-coplanar case.
- add_vertex() can be called from split_vertex() which does not guarantee
to have properly pre-allocate arrays.
- Need to check whether Cycles is compiled with OSL in XML reader.
A bit tricky, need to pass additional information about what the attribute
is and how to deal with it.
BI path stays unchanged, just to make things simplier for now.
Fixes T48555: Cycles GLSL- Incorrect Vertex Color results from Attribute node
Some of these values can get quite large and are hard to read, adding this
makes it easy to read them at a glance.
Reviewed By: sergey
Differential Revision: https://developer.blender.org/D2039
We had transparent shadows disabled for some time because they were causing
drivers to crash. Can't reproduce that issue anymore with current drivers,
so will enable them and see how it goes.
As long as the compiler class is WIP the two implementations go out of sync too
easily. Better to make a copy of the dual version and then reduce it when the
compiler class is reasonably stable.
This approach suffers from difficulty of generating machine-independent IR code
and is better generated using the C++ API directly. The llc tool can be used to suggest
implementations with C++ API usage, see for example
http://fdiv.net/2012/11/16/llvm-generates-code-that-generates-code
Although there is a "UI limit", it's necessary to have the hard limit on the
property too. I noticed this bug first hand just now, when, after accidentally
setting the size to 8188 (due to a combination of typos and sluggish UI refresh)
my machine locked up completely when trying to jump to that frame. It got so bad
that I had to do a hard reset to fix it - so, it's possible that even 1000 or 2000
as used now are actually way too large still.
It was possible to have issues in cases when several child dependencies
goes to IDs with different layers. In this case order of flushing was not
really well defined, which could lead to cases when indirect dependency
via invisible object wouldn't work.
Need some sort of barrier to prevent scheduling of parent nodes for until
all children are done, but that's becoming quite nasty thing to implement.
Added a temp field to component for now. maybe it's not so crazy actually
and we might use it for evaluation as well, so we wouldn't flush updates
to components which does not affect visible stuff.
This patch enables Tube, Sphere and Box projections in GLSL for the image texture node.
Reviewers: sergey
Projects: #nodes, #opengl_gfx
Differential Revision: https://developer.blender.org/D2036
Previously a whole mesh was stored between undo steps,
This commit uses BLI_array_store to de-duplicate memory use between undo steps.
Memory saving depends entirely on kinds of edits performed,
in own tests 5x-15x less memory use is common.
Compacting the memory does give some overhead however its done in a background thread
so its not blocking in most cases.
New behavior and threading can be ifdef'd out to check for regressions.
See D2026 for details.
Not replacing with some BLI_task_stuff here, tests show this is pointless
(in absolute best case - i.e. single huge mesh in scene - parallelizing here switches
from 0.8ms to 0.5ms for that piece of code - with something like 750ms per frame update...).
Sometimes the autokeying flags don't get set correctly (i.e. the "mode" flags
used for 'Add + Replace' vs 'Replace Only' aren't set), meaning that the old logic
would always fall through to the "replace only" case. When this happens, the resulting
behaviour can be quite strange and hard to debug. This fix prevents problems like
this from continuing to be an issue...
Not sure why, but ANIM_animdata_update() totally ignored NLAStrip type of bAnimListElem.
For now only added support for ANIM_UPDATE_DEPS type of update, don't know whether
others are needed in this case or not... time will say.
This is in preparation for specializing the value types of compiler implementations,
so that we can replace dual struct types with independent llvm values.
This is mainly a maintenance commit which was aimed to make work with
this module more pleasant and solve such issues as:
- Annoyance with looong files, which had craftload in them
- Usage of STL for the data structures we've got in BLI
- Possible symbol conflicts
- Not real clear layout of what is located where
So in this commit the following changes are done:
- STL is prohibited, it's not really predictable on various compilers,
with our BLI algorithms we can predict things much better.
There are still few usages of std::vector, but that we'll be
solving later once we've got similar thing in BLI.
- Simplify foreach loops, avoid using const_iterator all over the place.
- New directory layout, which is hopefully easier to follow.
- Some files were split, some of them will be split soon.
The idea of this is to split huge functions into own files with
good documentation and everything.
- Removed stuff which was planned for use in the future but was never
finished, tested or anything.
Let's wipe it out for now, and bring back once we really start using
it, so it'll be more clear if it solves our needs.
- All the internal routines were moved to DEG namespace to separate
them better from rest of blender.
Some places now annoyingly using DEG::foo, but that we can olve by
moving some utility functions inside of the namespace.
While working on this we've found some hotspot in updates flush, so
now playback of blenrig is few percent faster (something like 96fps
with previous master and around 99-100fps after this change).
Not saying it's something final, there is still room for cleanup and
API simplification, but those might happen as a regular development
now without doing any global changes.
Like node trees, textures are only optionally constructed in the depsgraph when something
uses them. When a modifier (or other potential future user of the "handle" system)
declares a texture dependency, the internal relations of that texture are not necessarily built -
this only happens when the texture is part of a material texture stack. So we need to also
construct the internal texture relations (lazy init) in addition to the external modifier->texture
dependency.
Gives over 50% faster scope update (from 4.5ms to 2.2ms here with SD shot)!
Probably mostly due to more clever usage of thread-local data (which avoids any lock,
when OMP code had a rather stupid critical section for minmax)...
Because pynode groups now use the bl_idname of tree types rather than
expecting a python class attribute, these tree types need to be registered
*before* generating node group classes. Messy stuff, needs some rewriting.
The updateDepsgraph callback for modifiers needs to be used in both passes of
depsgraph building, once for creating nodes and then for relations. Without the
initial node building pass some datablocks like textures can be missing, if they
are not incidentally built by other means such as a render texture.
This is mostly a debugging feature for getting visual feedback about
the derivatives. The function doesn't work in the regular scheme of
decomposing a function into the value and derivative parts, so it is
only implemented in the internal IR (making it directly inlineable).
Bevel had assumed that when rebuilding a face that touches
a vertex with beveled edges, the edges of the face at that vertex
would be adjacent in internal order. That is not necessarily true
if there are edges with more than two faces attached.
We could just prohibit beveling any edges that touch a vertex
where this happens (we already don't bevel non-manifold edges)
but the use case in the model of T47257 seems reasonable.
Also had to fix the edge-ordering code, and the face reconstruction
code to take care of cases where the face normal may not be as expected.
New code using loop/finalize model is about 45% faster (from 4.4ms tp 2.4ms per frame,
overall playback of single shot in sequencer in this preview mode goes from 40 to 45fps).
For now, just validate new pointcache names as if they were filnames.
Ideal solution would be rather to validate names in `ptcache_filename()`, but it would
likely break some existing caches - and we'd also have to ensure we still have
uniqueness of name after it has been cleaned up, wich would not be trivial at this stage.
So let's go for simple solution for now, especially since this part of code is to be
fully rewritten in 2.8...
See T48453 for details and test scenes
Reviewers: a.romanov, sergey
Reviewed By: a.romanov, sergey
Projects: #opengl_gfx, #nodes
Differential Revision: https://developer.blender.org/D2011
Currently Dual2 arguments are structs with the main value and 2 derivatives
in x and y. Node functions will generally be implemented using the base types,
so the value and derivatives have to be extracted and later combined using
the chain rule.
It's not really clear at this moment how we can detect cases
when attribute needs linearization. For now added a comment
so we don't forget about this, hopefully.
Some closures were missing from calculation, leading to an array
under-allocation, presumable causing memory corruption issues with
emission shaders on OpenCL and was causing issues with Volume 3D
textures with CUDA.
The issue was identified by Thomas Dinges, the patch is different
from the original D2006. See the brief discussion there. Current
approach is similar (or the same) as Brecht suggested.
In Cycle standalone, the xml import of uv's is done in a way that
a vertex could only have one unique uv coordinate. In practice it
is most of the time not the case.
Patch by Laurent Boiron, thanks!
This fixes a rare case where NaNs could exist inside Cycles.
When certain invalid meshes were passed in, Cycles would try too normalize
a zero length normal during its setup stage. While it does check against
division by zero, it still returns a zero length normal and passes it on to
the path tracing kernel. The kernel then operates under the assumption that
normals are valid, and in the case of such a zero length normal, would
eventually create NaNs that propagate through and result in black pixels.
Reviewers: #cycles
Subscribers: brecht, sergey
Projects: #cycles
Differential Revision: https://developer.blender.org/D2008
The title says it all actually.
Added special custom data type, because we don't know in advance
whether we're referencing UV or Color layer. Also made it so vertex
attributes are normalized.
TODO: Border render in viewport ignores the normalization of the
attribute array for some reason, will be looked into still.
Reviewers: mont29, brecht, campbellbarton
Reviewed By: brecht, campbellbarton
Differential Revision: https://developer.blender.org/D2022
At some point the idea was that we could have an optimization where we could
render multiple render layers without re-exporting the scene, by just updating
the layer bits. We are not doing this now and in practice with the available
render layer control like exclude layers it's not always possible anyway.
This makes it easier to support an arbitrary number of layers in the future
(hopefully this summer), and frees up some useful bits in the kernel.
Reviewed By: sergey, dingto
Differential Revision: https://developer.blender.org/D2020
This commit implements Bump node in GLSL, making it possible to
see previews of bump mapping in viewport without need to render.
Nothing really fancy going on here, just uses internal dFdx/dFdy
functions to get derivatives of the surface and map itself.
Quite basic but seems to behave correct-ish.
This commit also makes Displacement material output to affect
viewport shading by re-linking unconnected Normal input to a
node which was used for displacement output (via Bump node).
Intention of all this is to make it really easy to do bump map
painting with Cycles as an active render engine.
Reviewers: campbellbarton, mont29, brecht, psy-fi
Reviewed By: brecht
Subscribers: Blendify, eyecandy
Differential Revision: https://developer.blender.org/D2014
Since version 6 G++ switched to C++11 by default, which breaks some logic
around WITH_CXX11 checks in out CMake files, leading to compilation errors.
This is easy to solve by explicitly enabling older C++ standard when C++11
was not explicitly enabled by CMake options.
However, G++-6 will also use new ABI by default even if older standard was
specified in the compiler options. This is being addressed by a special
define flag.
This tricks made it possible to use new G++-6 without need to recompile
any of pre-compiled libraries.
However, this might break compilation with existing system libraries, which
might already be using new ABI. We can't address this automatically, so
now we simply default WITH_C11 and WITH_CXX11 options to whatever defaults
of the current compiler are. This means, for G++-6 we'll set WITH_CXX11 to
truth. This should make linking with system libraries working just fine,
but to make pre-compiled libraries we still might need to disable CXX11.
This should work fine work for a new environments with G++-6 and install_deps
script run from scratch there, because C++ standard will be the same for
both Blender dependencies and Blender itself.
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.
Usage:
* D+X - Works anytime, anywhere
* Shift-X - Works in EditMode only
* Via Delete Menu - EditMode only
Often doing video tutorials or perhaps during dailies/shot review you want to
quickly get rid of a quick scribble you made for making a point, without having
to undo (i.e. maybe you edited some objects in between) and/or without having
to use the eraser (i.e. it'd take too long to cover the entire area).
On light coloured backgrounds (especially on white), it was impossible to see
where the cursor was. This commit adds a second ring (black) to the cursor so
that on light backgrounds, even if the light ring isn't visible the black one
will be.
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...).
It gives some slight differences on the plane corners, but can't
really figure out source of the issue here yet.
It's still better than fully white texture for the previews anyway.
At this point we should perhaps ifdef chunks of the code in order
to have faster GLSL compilation and probably even faster compiled
code. Will look into this shortly.
Compile time per kernel increased alot after recent image commits, re-shuffle some code to fix this.
Patch by "LazyDodo".
Differential Revision: https://developer.blender.org/D2012
In addition to the original bug report, I've gone through cleaning up a range of
related bugs which only became clear when hunting around the code...
* Custom Handle References weren't getting cleared when the bones they used got
deleted. But, neither was the custom bone shape location/transform reference.
* Various places where posebone settings are copied around were also missing code
to handle the new Bendy Bone properties.
(WHY DO WE HAVE SO MANY VARIATIONS OF COPYING POSE DATA!?!?)
* If duplicating a Bendy Bone with custom references, and the custom references
are also selected/duplicated, the new Bendy Bones will use the corresponding
duplicated bones
This commit adds Peak Memory to the stamp options, the value is the same one that is already shown in the image viewer.
Requested by @nutel.
Reviewers: campbellbarton
Subscribers: campbellbarton, nutel
Differential Revision: https://developer.blender.org/D1989
Once and for all - I hope!
So, for the records: Xinput1 events' valuators have a first_axis and axis_count defining the 'range'
of valid axes for that event - but valid data in the valuator's array **always** start at zero
(which means that, if event reports first axis as being axis 2, axis 2's value is in axis_data[0]).
This means we can have differing implementations for every potential use
of nodes, such as plain value functions vs. dual values with derivatives.
Node functions for each compiler type are stored in a static map and are
lazy-initialized on the first node graph compilation.
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.
This adds support for CUDA Texture objects (also known as Bindless textures) for Kepler GPUs (Geforce 6xx and above).
This is used for all 2D/3D textures, data still uses arrays as before.
User benefits:
* No more limits of image textures on Kepler.
We had 5 float4 and 145 byte4 slots there before, now we have 1024 float4 and 1024 byte4.
This can be extended further if we need to (just change the define).
* Single channel textures slots (byte and float) are now supported on Kepler as well (1024 slots for each type).
ToDo / Issues:
* 3D textures don't work yet, at least don't show up during render. I have no idea whats wrong yet.
* Dynamically allocate bindless_mapping array?
I hope Fermi still works fine, but that should be tested on a Fermi card before pushing to master.
Part of my GSoC 2016.
Reviewers: sergey, #cycles, brecht
Subscribers: swerner, jtheninja, brecht, sergey
Differential Revision: https://developer.blender.org/D1999
Currently only the value part is actually accessed. Using dual values
allows the node graph to handle input/output arguments are single a
single llvm::Value instance.
It was totally useless to multiply diffuse color with the vertex color
when doing texture painting. It was masking actual texture and only was
forcing artists to create an empty vertex color layer to work this around.
Uses some virtual functions of the node compiler to allow different
implementation of values. This will be used for calculating values
and partial derivatives in texture nodes.
The idea of pole merge is to fade interocular distance after a certain
altitude to zero when altitude goes closer to a pole. This should prevent
annoyances looking up in the sky or down to the bottom.
Works for both panorama and perspective cameras when Spherical Stereo
is enabled.
Reviewers: dfelinto, brecht
Reviewed By: brecht
Subscribers: sebastian_k
Differential Revision: https://developer.blender.org/D1998
This commit makes it so malloc() is only happening once per volume and
once per transparent shadow query (per thread), improving scalability of
the code to multiple CPU cores.
Hard to measure this with a low-bottom i7 here currently, but from quick
tests seems volume sampling gave about 3-5% speedup.
The idea is to store allocated memory in kernel globals, which are per
thread on CPU already.
Reviewers: dingto, juicyfruit, lukasstockner97, maiself, brecht
Reviewed By: brecht
Subscribers: Blendify, nutel
Differential Revision: https://developer.blender.org/D1996
When using Left Click select, it wasn't possible to sculpt using E+LMB.
I've changed the order of things in the keymap so that the select operator won't
end up catching and blocking all these events.
Change the order of the bending controls ("Curve XY Offsets") so the user can activate both InX and OutX by holding down the left mouse button. This way, it's easy to bend symmetrically on X or Y.
This commit/patch/branch brings a bunch of powerful new options for B-Bones and
for working with B-Bones, making it easier for animators to create their own
rigs, using fewer bones (which also means hopefully lighter + faster rigs ;)
This functionality was first demoed by Daniel at BConf15
Some highlights from this patch include:
* You can now directly control the shape of B-Bones using a series of properties
instead of being restricted to trying to indirectly control them through the
neighbouring bones. See the "Bendy Bones" panel...
* B-Bones can be shaped in EditMode to define a "curved rest pose" for the bone.
This is useful for things like eyebrows and mouths/eyelids
* You can now make B-Bones use custom bones as their reference bone handles,
instead of only using the parent/child bones. To do so, enable the
"Use Custom Reference Handles" toggle. If none are specified, then the BBone will
only use the Bendy Bone properties.
* Constraints Head/Tail option can now slide along the B-Bone shape, instead of
just linearly interpolating between the endpoints of the bone.
For more details, see:
* http://aligorith.blogspot.co.nz/2016/05/bendy-bones-dev-update.html
* http://aligorith.blogspot.co.nz/2016/05/an-in-depth-look-at-how-b-bones-work.html
-- Credits --
Original Idea: Daniel M Lara (pepeland)
Original Patch/Research: Jose Molina
Additional Development + Polish: Joshua Leung (aligorith)
Testing/Feedback: Daniel M Lara (pepeland), Juan Pablo Bouza (jpbouza)
Together with the extended loop callback and userdata_chunk, this allows to perform
cumulative tasks (like aggregation) in a lockfree way using local userdata_chunk to store temp data,
and once all workers have finished, to merge those userdata_chunks in the finalize callback
(from calling thread, so no need to lock here either).
Note that this changes how userdata_chunk is handled (now fully from 'main' thread,
which means a given worker thread will always get the same userdata_chunk, without
being re-initialized anymore to init value at start of each iter chunk).
New code is actually much, much better than first version, using 'fetch_and_add' atomic op
here allows us to get rid of the loop etc.
The broken CAS issue remains on windows, to be investigated...
Some variants of gcc compilation were reporting 'control reaching end of non-void function' error
in this switch/case maze. Either use break everywhere or not at all (which is simpler, since we
only always return anyway...).
The blenvm system depends on the new depsgraph, so it doesn't make sense
to have to enable it every time.
The legacy depsgraph can still be switched back on by the --enable-legacy-depsgraph option.
There are some serious issues under windows, causing deadlocks somehow (not reproducible under linux so far).
Until further investigation over why this happens, better to revert to previous
spin-locked behavior.
This reverts commits a83bc4f597 and 98123ae916.
This currently includes just a few 'value' nodes, but should be extended as much as feasible.
In particular many vector functions can be split component-wise.
Reading the shared state->iter value after storing it in the 'reference' var could in theory
lead to a race condition setting state->iter value above state->stop, which would be 'deadly'.
This **may** be the cause of T48422, though I was not able to reproduce that issue so far.
The LLVM backend also supports PASS nodes, even though these are usually
removed from the node graph in a preprocessing step for clarity.
ARG nodes have their output values mapped in advance, so no extra code
needs to be generated for them.
Outputs are currently identified in the nodes via the DO_OUTPUT flag. This is not
nicely accessible for pynodes and needs to be reimplemented in a proper way.
Until then this optimization for node recalculation should just be ignored for texnodes,
so nodes now have a callback for checking usage, and will always be considered used by default.
When writing temp blenbuffer file, libraries of linked datablocks where not tagged correctly, which
means they were not put in the temp Main used to write the buffer, resulting in implicit localization
of linked data.
Previous to 2.77, this used to be default behavior, was changed in rB591f4549c958b.
However, in most append cases, you do want a full localization of your data, so this new behavior
is kept by default, but there is now an option in append operator to only localize the 'first level'
of data (i.e. datablocks from linked library itself, and not those from other 'sub-libraries').
On big and complex rigs like blendrig or koro, it can give up to ~10% more FPS in best cases.
Hard to tackle all cases in tests though, so please report any unexpected slowdown
in armature animation playback!
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.
Previously if image only had single channel only z buffer value was displaying.
This isn't handy for cases when you've got single channel buffer which is not
a z buffer.
Also fixed possible read past the array.
Previously the "parameters" depsnode was the POST operation for node trees, but has now
been supplanted by the cache invalidation node. Making an explicit dependency on the component
thus creates a cycle.
This also removes the static bNodeTree type definition C, because
BVM requires some functions that are best defined in python (and it
would have been switched eventually anyway).
install_deps can fail due to conflict between gcc (referenced by base-devel) and
gcc-multilib if the latter is installed. This avoids the conflict by filtering
the contents of base-devel when needed.
Reviewers: mont29
Differential Revision: https://developer.blender.org/D1944
This way, we also save 3/4th of memory for single channel byte textures (e.g. Bump Maps).
Note: In order for this to work, the texture *must* have 1 channel only.
In Gimp you can e.g. do that via the menu: Image -> Mode -> Grayscale
Until now, single channel textures were packed into a float4, wasting 3 floats per pixel. Memory usage of such textures is now reduced by 3/4.
Voxel Attributes such as density, flame and heat benefit from this, but also Bumpmaps with one channel.
This commit also includes some cleanup and code deduplication for image loading.
Example Smoke render from Cosmos Laundromat: http://www.pasteall.org/pic/show.php?id=102972
Memory here went down from ~600MB to ~300MB.
Reviewers: #cycles, brecht
Differential Revision: https://developer.blender.org/D1981
Take advantage of the efficiency provided by the snap_context.
Also fixes errors:
- volume snap fails based on view angle (T48394).
- multiple instances of dupli-objects break volume calculation.
Pass distance argument so its possible to limit the range we get all hits from.
Other changes:
- Use boundbox test before calling callback, avoids redundant calls.
- Remove meaningless return value.
- Add doc string, explaining purpose of this function.
This commit makes use of new taskpool feature (instead of allocating own tasks),
and removes the spinlock used to generate chunks (using atomic ops instead).
In best cases (dynamic scheduled loop with light processing func callback), we
get a few percents of speedup, in most cases there is no sensible enhancement.
This is what old dependency graph was doing and apparently this avoids some updates,
however it's not really clear why those nodes are considering done. Maybe just because
of the way how relations are working. But needs a closer investigation.
Appears mutex was guarateeing number of tasks is not modified at moments
when it's not expected. Removing those mutexes resulted in some hard-to-catch
locks where worker thread were waiting for work by all the tasks were already
done.
This reverts commit a1d8fe052c.
Brain melt here, intention was to reduce number of tasks in case we have not much chunks of data to loop over,
not to increase it!
Note that this only affected dynamic scheduling.
It uses some additional compute power and the evaluation priority is
not even used.
This brings fps 88.2 with blenrig_for_debugging.blend on this desktop.
Simple thing, and apparently fps goes up to 80 with the demo file from jpbouza.
Not sure why at this point fps is so much higher than the old dependency graph
here now. And it's definitely something what others should verify as well.
The general signature layout for node functions is
void foobar(out1_t *out1, out2_t *out2, ..., in1_t in1, in2_t in2, ...);
All outputs are passed as pointers or references (using references in C code for clarity).
Inputs can be passed by-value or as const pointers/references (for structs and arrays).
In the future we may also support single value return types for convenience.
This reduces stress on the task scheduler and avoids some unwanted overhead
caused by all the threading business in the cases when there's only one
children node. We try to immediately switch to it's evaluation now, keeping
active thread up and running.
This bumps FPS from 58 to 64 on the blenrig test file from jpbouza.
This commit implements new function BLI_task_pool_push_from_thread()
who's main goal is to have less parasitic load on the CPU bu avoiding
memory allocations as much as possible, making taks pushing cheaper.
This function expects thread ID, which must be 0 for the thread from
which pool is created from (and from which wait_work() is called) and
for other threads it mush be the ID which was sent to the thread working
function.
This reduces allocations quite a bit in the new dependency graph,
hopefully gaining some visible speedup on a fewzillion core machines
(on my own machine can only see benefit in profiler, which shows
significant reduce of time wasted in the memory allocation).
Many keyboard layouts (italian, spanish, german...) have direct access to '+' key on main
keyboard area (not the numpad one), ans x11 has own define for this key, so use it instead
of generating an unkown key event.
Note that we most likely have much more missing 'specific' keycodes for non-US keyboard layout,
but think since we already had a 'minus' keyevent, supporting 'plus' one is totally consistent.
And we had a spare space in our defined values just for it even!
This keyevent is only supported/generated by x11 and cocoa Ghost backends for now,
neither SDL nor win32 seem to have matching key events...
Title says it all, this adds OpenCL float4 texture support.
There is a bug in the code still, I get a "Out of ressources error" on nvidia hardware here, not sure whats wrong yet.
Will investigate further, but maybe someone else has an idea. :)
Reviewers: #cycles, brecht
Subscribers: brecht, candreacchio
Differential Revision: https://developer.blender.org/D1983
This commit:
* Removes most of all dirty internal details from public atomi_ops.h file, and move them into /intern private subdir.
* Removes unused 'architectures' (__apple__ and jemalloc).
* Split each implementation into its own file.
* Makes use of C99's limits.h system header to determine pointer and int size, instead of using fix hardcoded list of architectures.
* Introduces new 'faked' atomics ops for floats.
Note that we may add a lot more real and 'faked' atomic operations over integers and floats
(multiplication, division, bitshift, bitwise booleans, etc.), as needs arise.
Reviewers: sergey, campbellbarton
Differential Revision: https://developer.blender.org/D1982
This was originally done for T46320 in order to re-store depsgraph state
after all the constraint modifications.
However, it relations were updated if there's any IK chain the the pose,
which is a bit too annoying.
Set is much slower to iterate through (due to cache misses and such) and
the only advantage of using set is faster removal of link. However, we are
iterating links much much more often than removing them, and even when we
are removing links we don't really need to remove link from nodes which it
connects -- we don't support partial depsgraph updates, so removing links
from nodes on destruction is a waste of time.
If we ever want to support partial updates we can have dedicated function
to remove link from nodes it connects.
This gives a surprising increase of fps from 42 to 56 with test file from
Mr. J.P.Bouza (blenrig_for_debugging.blend). Surprising because old DEG is
actually slower here (52 fps). Didn't see any regressions (and don't see
why they will happen), so let's ask our riggers and animators to perform
further speed tests ;)
Use atomic operations instead, should in theory improve timing of
scheduling. However, probably not so visible yet because actual
task scheduling still have some locks and memory allocations.
Baby steps, what would i say.
This was causing alignment issues which were only visible on a platform
of particular bitness, so it was easy to break stuff for 32bit when
working on 64bit platform and vice versa.
* Fix "Attempt to free NULL pointer" when copying strokes for the first time
* Fix poll callback on "paste" operator, so that it is possible to paste
strokes when there are no editable strokes visible.
Commit 86abddc9, caused an error when the face-region included boundary edges.
Since removing the faces first, caused the edges along the boundaries to be removed.
Add support for deleting faces and internal edges, that keeps boundaries.
Base modules for nodes will instead be declared as external functions with regular
linkage through C code. This means some potential for low-level optimization might be lost,
but this shouldn't be a big problem as long as node functions don't get too complex
(which means they should be split up in the first place).
IR module loading is not the real issue, but generating the IR code from C/C++ code
in a generic way is not so easy. In the future we may actually generate most code for
basic operators and types directly inside blenvm and only use precompiled modules
for more complex functions.
Now, when creating new keyframes in the graph editor by ctrl-clicking, only
the newly created keyframes will be selected. This is a little workflow tweak
to make it faster to work, as you no longer have to deselect all, and then
re-select the newly added keys in order to manipulate them.
The old behaviour (not modifying the selection status of the old keys) has been
kept, but is now available via Shift-Ctrl-Click.
Feature request from @Shhlife
This commit makes remove_table skip the freeing if the offset is
already set to invalid - or, if it wasn't, set it to invalid after freeing.
That's what the current code was already doing in the Manager classes,
this change allows them to just call remove without the additional code.
Also, two potential memory leaks where new tables were always allocated
without freeing the old ones are fixed.
Reviewers: sergey, dingto, brecht
Differential Revision: https://developer.blender.org/D1974
This is an experimental option that I found would have been useful to have,
when making my earlier anim test. The only contentious point is whether it's
better to use a separate boolean flag (perhaps bound to the labels) instead
of overloading this setting (and describing the usage in the tooltip).
In the nodes editor, Ctrl-LMB-drag is taken by "Cut", and it seems impossible
for any other keymap to override that. Instead, the default lasso select hotkey
there is Ctrl-Alt-LMB-drag instead.
To get Grease Pencil lasso select (in GP Editmode) to work in the Nodes Editor too,
this commit adds the Ctrl-Alt-LMB-drag binding here too. However, to make things easier,
this extra binding will be part of the Grease Pencil keymap everywhere. There doesn't
seem to be any conflicts with having this in place - until we find them, this should
be ok to have.
I'm still not happy with this layout as it is now, but it seems a bit less unbalanced
than what I'd been trying before. So, let's leave this as-is for now.
This commit adds some of the initial support for a properties region in the
Action Editor. There are currently no panels to display, as there is still
a lot of work required to port over the required internal architecture to
support the panels seen in the Graph Editor.
After some test, a new iteration parameter has been added in order to
apply repetitive smoothing to the stroke. By default 1 iteration is applied,
but can used any number between 1 and 3.
The repetition uses different levels of intensity from 100% of the defined smooth
factor for the first loop, 50% for the second and 25% for the third. We use in each
loop a smaller value in order to avoid deform too much the stroke.
Now, stroke-related things (thickness, volumetric, points) and fill-related things
(HQ fill) are in the relevant columns, instead of having some in each column.
There is still some instability in how the triangulations are happening,
where the triangle count of the resulting triangulation fluctuates resulting
in weird artifacts sometimes.
To reproduce, try drawing some U-shapes, and keep reloading the file.
While trying to track down why I still keep getting some random "extra" triangles
showing up when reloading files with fills, I noticed that there's an index mismatch
here that may cause problems in some rare cases:
There are "gps->totpoints" items in tmp_triangles, but there should only be
"gps->tot_triangles" triangles in the gps->triangles array. If for whatever reason
some of the triangles in tmp_triangles are invalid, this could have meant that
while a tmp_triangles candidate was skipped, a corresponding slot for a valid
triangle also got skipped.
Improve filling for concave shapes using a triangulation of the stroke.
The triangulation information is saved in an internal cache and only is
recalculated if the stroke changes.
The triangulation is not saved in .blend file.
Reviewers: aligorith
Maniphest Tasks: T47102
Differential Revision: https://developer.blender.org/D1705
`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...
Swap the for-loops in the RenderBuffer reset code to follow the convention
of looping over y in the outer loop.
The improved cache performance won't really be noticable here, but it's nicer
if it follows the usual style.
Take curvature into account when calculating handle length.
Gives significantly better results for curve dissolve and 10-20% more efficient freehand drawing.
* When Baking wasn't used we got an error.
* On top of Volume Nodes (NODES_FEATURE_VOLUME), we now also check if we need volume sampling code,
so we can disable that as well and save some further compilation time.
If the CUDA Toolkit is installed and the user is on Linux,
adaptive, feature based CUDA runtime compile is now possible to enable via:
* Environment flag CYCLES_CUDA_ADAPTIVE_COMPILE or
* Debug menu (Debug value 256) in the Cycles UI.
This commit avoids draw of invisible parts of image in image editor,
making it faster to re-draw the image.
Especially handy when painting on a high-res image when zoomed-in.
This commit simplifies the code for the image arrays. Instead of having 2 arrays for float and byte textures,
we now use an array here. This simplifies the code (avoids code duplication), and makes it possible to easily extend it
with float1 and half-float types in the future.
Only tested with CPU yet, plus some cleanup / code de-duplication is still possible here.
Reviewers: #cycles, sergey
Reviewed By: #cycles, sergey
Subscribers: jesterking, sergey
Differential Revision: https://developer.blender.org/D1969
Originally the value was only needed when building the springs,
so a pointer to the input data was somewhat ok. However, recalculating
spring length dynamically requires keeping the actual value around.
The average length is used to adjust self-collision and stiffness
to match mesh density, and in either case considering the always zero
restlen sewing springs doesn't actually make much sense.
The struct_springs variable is also used for other things like total
spring count, so a new one needed for the true struct springs.
This comes out of considering a one-dimensional transition in weight
on a rectangular cloth grid. At the transition face loop, one side
of each rectangular face would be scaled by k1, and the opposite one
by k2, thus turning the rectangle into a trapezoid. Averaging would
be used to choose the scale factor for the remaining two sides.
If Geometrical Mean, i.e. sqrt(k1*k2) is used, it so happens that the
diagonals of the trapezoid also end up scaled by sqrt(k1*k2) compared
to the original rectangle. This means that the same scale factor is
correct for both structural and shear springs, which is not the case
with simple average.
Previously it was all over the place: without a vertex group it was
always enabled, and with it it depended on the Self Collision option
due to a likely editing mistake. Now it should be in sync with what
disables the fields in the UI.
The verts->shrink_factor field is changed to just cache the vertex
weight, with the shrink_min/shrink_max interpolation done later.
This is because cloth_apply_vgroup only runs if there are vertex
groups, and thus the factor may not update after property changes.
In addition, bending springs are now also affected by the shrink
factor to avoid visible distortion in object shape at high shrink.
From description, Use Coordinates evaluates the texture using
target coordinates in the local space of the force field object.
2D is supposed to ignore the Z coordinate. Thus one would assume
that if both are enabled, the force field effect would move with
the force field object, and Z would be 0.
However, instead first the 2D option projects points onto a plane
passing through the global zero and orthogonal to the local Z,
and only then the resulting point is transformed into local space.
Z is not locked at 0, so procedural textures like Spherical Blend
don't work as expected.
To fix this, apply local transform first, and then just clear Z if 2D.
Separate the creation of trees from EditMesh from the creation of trees from DerivedMesh.
This was meant to simplify the API, but didn't work out so well.
`bvhtree_from_mesh_*` actually is working as `bvhtree_from_derivedmesh_*`.
This is inconsistent with the trees created from EditMesh. Since for create them does not use the DerivedMesh.
In such cases the dm is being used only to cache the tree in the struct DerivedMesh. What is immediately released once
bvhtree is being used in functions that change(tag) the DM cleaning the cache.
- Use a filter function so users of SnapObjectContext can define how edit-mesh elements are handled.
- Remove em_evil.
- bvhtree of EditMesh is now really cached in the snap functions.
- Code becomes organized and easier to maintain.
This is an important patch for future improvements in snapping functions.
Using SSE2 intrinsics when available for this kind of conversions.
It's not totally accurate, but accurate enough for the purposes where
we're using direct colorspace conversion by-passing OCIO.
Partially based on code from Cycles, partially based on other online
articles:
https://stackoverflow.com/questions/6475373/optimizations-for-pow-with-const-non-integer-exponent
Makes projection painting on hi-res float textures smoother.
This commit also enables global SSE2 in Blender. It shouldn't
bring any regressions in supported hardware (we require SSE2 since
2.64 now), but should keep an eye on because compilers might have
some bugs with that (unlikely, but possible).
FSAA sample files in EXR format are no longer always updated (after some
changes between 2.73 and 2.74 releases), and the reported bug was caused
by old samples from a previous frame that were being merged by mistake.
The present revision addresses the documented issue by entirely skipping
the rendering of auto-generated scenes when there are no Freestyle strokes
to render, which suppresses sample merging of the render layers that were
not rendered.
Just avoid some unneeded initialization functions when the threaded
processor is simple enough to only depend on current chunk start
scanline and number of scanlines.
Caused by wrong alpha value in rB34b30f847429ac (used 0.01, should've been 0.1). Also made sure all editors initialize same default theme values for panels and cleaned up code a bit.
Works fine with old userpref.blend files.
Basically, issue is that outliner stores editbones for an edited armature, which are only valid during EditMode.
When more than one object use the same armature, selecting ebones from same armature but under non-edited object
would first select that object (and hence switch out of editmode), and then try to select editbone while editdata
no more exist.
Solution for now is to not change active object in this case - it's not totally ideal (not consistent),
but other solutions (like switching edited object without leaving editmode, or rebuilding (part of)
outliner tree in-between) are horribly hackish and most likely prone to epic failure anyway.
Shift-space was reserved for IME support, however IME will only
be enabled on certain languages. We can avoid any IME-related
exceptions form other languages without too much trouble.
There's one weak point around ui_ime_is_lang_supported() tho,
which is it might be doing string comparison a bit too much
often now, this we can avoid by handling those checks from BLT.
Not very efficient solution -- it'll update mapping node on init ntree exec and will
not work for viewport GLSL shading perhaps, but it's as good as it could be within
current dependency graph.
The issue here is that manual edit of values will cause cached matrix re-evaluation.
but using animation does not use rna update callbacks hence no matrix update was
happening.
- scripts that execute directly need to include their basedir in the sys.path
- modules which are in a directory without an __init__.py weren't importing.
__GCC_HAVE_SYNC_COMPARE_AND_SWAP_n are not defined on this architecture
for some reason, however those functions are available.
Adding a workaround for newly used __sync_fetch_and_and() function.
Own regression from rB019ce363b01bba0af, not totally sure why removing the heading
zero value in element_sum broke binary search distribution in random branch, but...
New code makes thing sooo much simpler, by simply getting rid of all zero-weight
items in element_sum (at the cost of adding a temporary extra index mapping array),
that way neither random nor 'even ordered' distribution code has to wory anymore
about skipping zero-weighted items!
In some cases, we can have nodegroups without attached nodetree, this case is checked
in top `node_group_has_output()` entry point, but not in recursive `node_group_has_output_dfs()`.
Issue appeared after rB42824f4403db3a35, presumably because check on node flag itself
was skipping that case, which is no more possible when using ID tags.
Reported by sebastian_k over IRC, thanks.
This commit mostly:
* removes some (very) old messages or hacks (like those for ubuntu 10.4...).
* Moves lib-specific dependencies installations to upper level
(simpler to manage, and those are small enough packages).
* Uses new dnf package manager for fedora (sigh, we now have tree different commands for fed/suse/rehl).
* fixes or enhances some options for commandline tools.
* Force anew building of oiio package, due to insane dependencies of this one in recent distro.
Part of D1933.
Now CD_SHAPEKEY_INDEX customdata is stored in edit-mode when hooks and vertex parents are used.
This also fixes a bug where undo would loose key-index data.
Move to structs for BM_mesh_bm_to/from_me to avoid passing many argument, which mostly aren't used.
This adds the possibility to edit a regular mesh like strand data.
It's more of a test for independence from particles than a real use case,
but could come in handy anyway.
This requires a somewhat hackish stub function to force linking of inlined
functions in the modules, so that clang doesn't just omit them when
producing IR code. Not very nice, but also not directly related to the
actual Blender code.
The code uses a pointer-based signature convention for node functions,
which helps to avoid issues with type coercion through clang.
In the future we could either find a more reliable way to produce llvm IR
for the nodes, or use node functions as external functions without inlining.
Names in C++ code tend to get mangled by the clang compiler according to the
respective target ABI. This is unfortunate because it makes it difficult to
map functions to nodes by name. For now we use a function attribute in the module
C++ code to pass on the "true" name of the function and map them internally.
Note that this is currently not quite working: The clang compiler call used
tentatively for generating the LLVM IR code for modules uses a lowering pass to
coerce struct data types (such as float3) according to the target ABI. This should
be avoided and is unnecessary for our purposes, since we stitch together and optimize
these code fragments ourselves. In principle we could avoid using a C compiler altogether
and write this stuff by hand, but the point of using the compiler is to avoid this tedious
work. More info:
http://stackoverflow.com/questions/22776391/why-does-clang-coerce-struct-parameters-to-ints
The previous way of acquiring functions from the cache was actually not threadsafe.
To do this efficiently the cache locking, lookup and potential update is now done
inside the API functions to prevent race conditions on the cache.
This code is not very elegant atm, but it does the job. Have to see how internal code
management of LLVM (modules) plays into this, the function cache then could become
redundant.
The awkward double shorthand "BVM" has two meanings here:
* BVM*** is short prefix for "BlenVM" (to be changed later?).
* ***FunctionBVM means a function in the bvm backend implementation (see FunctionLLVM etc.)
This should be cleaned up later, by chosing a more sensible name for the overall code module.
The idea is to use modules code for all the various backends to avoid lots of duplicate code.
This is similar to how the same code is used in Cycles for CPU and CUDA backends.
These were not needed by Blender yet, because LLVM is only used indirectly
through dependencies, which use their own build settings for LLVM include paths.
The naming here is a bit confusing: BlenVM is the name used for the overall code module. BVM can be a shorthand
for BlenVM, but is also the name of the backend for SVM-like code interpreter. General purpose code (such as
node graph and util files) should avoid the bvm prefix. BlenVM as the overall name for the code module is kept
for now, but may be replaced at a later point.
Note that currently one bad-level-include from non-bvm to bvm code remains, for the Value class used in nodes.
Values should ultimately be a backend-agnostic class for storing literal values (node inputs).
Note that kernel nodes now don't have explicit 'function' inputs
any more. Any expression input can be a separate block. Eventually
this should be handled by creating true loops, branches, etc. in the
generated code, rather than reading a jump address from the instructions
and passing it to a blackbox method as a "function pointer".
For the time being the kernel nodes simply prepend a jump address before
*every* input, using BVM_JMP_INVALID in case the input node is not in a
separate block.
The internal bvm nodes now can contain (implicit) cyclic connections, owing to the
connection of variable inputs to local variable outputs of kernel nodes. These link
cycles are never created explicitly by the node parser (pynodes), but are a result
of the semi-functional programming nature of the nodes: each node constitutes a
function of the local variables defined inside a kernel block. When the nodes are first
constructed from the user-level nodes these inputs are left unbound, and are then
resolved backwards as the individual node functions are inlined.
The use of "expressions" for node inputs must be generalized to support array-like
data and a level of functional programming. Inputs are now expressions *by default*.
A node can also have "variables" as inputs, which are implicitly defined in the base node
graph and should not be bounded in the first compiler step (i.e. are not defined by
pynodes). Variables instead are defined as outputs of other nodes (currently still called
"kernel" nodes), which would mostly constitute loops over such variables.
The concepts being worked on here are loosely based on those found in the Halide language
http://halide-lang.org/index.html
The rationale for this change is that copying variables on the bvm stack
will become necessary for compound variables (structs). In order to
efficiently keep pointers on the stack without making deep copies every
time such a variable is copied, the pointer must be a kind of shared_ptr
concept. In addition to this the stack variable itself must also be counted,
so that we have a way of detecting when it "goes out of scope", i.e. all
nodes using it have finished.
The node_counted_ptr is not a full C++ shared_ptr, because the user counting
is handled explicitly by the compiled bvm instructions. Using a std::shared_ptr
would be difficult due to the mixed levels of evaluation (compile-time constants,
storing shared_ptr on the opaque stack), and because of difficult interfacing
with C code (returning persistent plain pointers).
The stack size is calculated during codegen by an int-ceil division,
currently padding to int size (probably later to be changed to int4 for
better SSE support).
This is so the POINTER type can be used for general struct pointers and arrays.
RNA pointers should eventually be replaced with more specific types for
data that can be accessed in the bvm system.
The point of this is to keep type definitions out of the core bvm eval code.
TypeDesc is only used for the outermost layer of bvm::Function in order to
copy arguments to and results from the eval stack. OpCodes themselves are
agnostic to higher type definitions and just work with raw data.
Eventually this could allow actual user typedefs (through the API),
but primarily this is to make complex type definitions possible without
having to do a full definition every time, and simply use the type name.
The returned mesh_ptr result needs to be cleared to ensure the user
counter is also freed. To avoid deleting the mesh data as well a new
function 'reset' was added, which frees the user counter but only resets
the data pointer without freeing it.
Modifiers should be able to work on meshes directly as much as possible,
without direct reference to an object. The space transform should
ultimately be passed along with the DM (TODO, currently requires explicit
inputs).
This was formerly part of an addon, but is useful enough to include directly in the branch
for debugging purposes. It uses the existing graphviz dump functions to write into a
temp dot file, then runs graphviz on it and opens the resulting image in the standard
system image viewer.
This is tailored for linux systems, but that shouldn't be a big issue since most developers
(who are the target audience) use this platform anyway.
Nodes can have expression inputs and override argument nodes with local
variables. This means that a local block must be created for each expression
input. The boundaries of such blocks are defined implicitly by upstream nodes
which define their own local arguments (or ultimately by the graph arguments).
The current version is still not quite generic enough to allow arbitrary
nesting of blocks, but it should ensure that expressions don't include all
the parent nodes which don't depend on local arguments and therefore should
be calculated in the parent block (a form of constant folding).
This simplifies code generation because we can use a set with efficient
lookup which also keeps a valid ordering of nodes as dictated by links.
Since any generated code uses only subsets of nodes, the ordering still
is valid for any BasicBlock.
This will allow different implementations of the compiler depending
on the intended target. This could be other bytecode formats (LLVM, SPIR),
but most imminently would be used for debugging.
In fact there is no per-output data in node instances at all, so the
OutputInstance data was removed entirely. All the connectivity and
value information is in the inputs.
These values are heap-allocated virtual classes, because general node code is
largely agnostic to data types. After adding values to a node (either as
default input values or per instance), care must be taken to free them when
the node type or instance is destroyed.
There is a redundancy in the nodetree/node relationship, since each nodes
is part of exactly one node tree. This means we can test for either trees
or nodes to detect cycles, and because we usually poll for node trees
rather than nodes that would be the natural choice.
This allows creating unoptimized node graph output dumps without
causing crashes, because the codegen step is not performed when
just producing debug output.
The use of notifiers from node trees is very unsatisfactory atm.
Notifiers should be integrated into the depsgraph, so that changes
to data automatically trigger appropriate UI redrawing.
This works just like the MESH type, but wraps a ListBase instead of
a DerivedMesh. Using this type the duplis can be generated internally
and returned as a result.
All such data types should be handled in a generalized way eventually.
These are an extension of the type description to support contiguous
arrays of base values. They effectively form a second set of types,
mirroring the 'single' base types. Socket connections will have to
strictly match between these types as well, i.e. any dereferencing and
array access will have to be explicit.
So far any change in a node tree would cause both a recompile of bvm functions
as well as a re-evaluation of any users of that function. This is quite inefficient
and should be separated to allow caching of the bvm functions independent from
their users.
To this end the nodes now have 2 callbacks: one for getting compile deps, the other
for getting eval deps.
* Compile dependencies are typically other node trees (node groups). In the future
this could also be Text datablocks, similar to how OSL script nodes work.
* Eval dependencies are all the blend_data blocks that a function might access.
For modifier nodes these could typically be meshes, curves; for shaders could
be images, textures; and so forth.
Separating these dependencies means the bvm functions will have to be recompiled
only when the actual bNodes change (add/remove, linking, socket values).
Conversely when some modifier input data changes (editing, cfra) the modifier
function will be re-evaluated.
This makes it possible to create generic dependency functions to
handle dependencies in various ways, using the same reporting callback.
Currently used for depsgraph building in 2 passes (nodes & relations).
This allows node groups and similar nodes with a node tree ID pointer
to register in the depsgraph.
In addition to defining relations, the builder must also make sure
that the node tree depsnodes are created first, because, unlike objects,
node trees depsnodes have no direct scene connection and are not generated
in advance. Since the build process has two distinct steps for creating
nodes and relations, the same API is used for both to simplify the code.
The function cache allows callers to store functions in a cache (hash table).
The functions also have a reference count, so replacing the function
in the cache (when the node tree changes) does not immediately delete
the current instance. This allows background threads to keep using the
old function until they can cleanly return (tasks should be restarted
when the nodes become outdated, but this is up to the caller system).
The const-ness is very important: strings are only stored as const char*
on the stack, which point to memory inside the instructions list.
The instructions contain the actual char array, with a null terminator.
This does not allow allocating new strings during eval, but it makes
it possible to use static strings as identifiers and keys for global data.
Eventually we could allow string operations through an allocation system
similar to DerivedMesh handling, with init/release nodes to manage
alloc/free.
This affects modifiers which use an external target object (currently
boolean and curve path). The modified object is a graph input argument,
which the target object is looked up via its key. The relative transform
means the values calculated from external geometry are always in the
modified object's local space.
Note that transform and its inverse are provided explicitly to nodes.
This will allow optimization for constants later on.
This should also work through the API! However, it requires building the nodetree
depsnodes before making relations which is not well supported yet.
The forced relations are currently needed to make updates inside the component
node trees trigger object recalc.
This simply reflects the existing C API for the depsgraph, as already used
by modifiers. The object nodes gain two callbacks, one for depsgraph
relations and one for registering globals. Since both of these should
require the same objects and other ID blocks to be registered, they share
a common `relations_update` function on the pynode level for convenience.
This introduces an improved way of handling ID relations inside the
node tree. Instead of storing simple indices to an object array,
the nodes now generate a 'key' (currently just ID name + lib name).
At eval time this key is then looked up from a hash table in globals,
to produce the actual pointer to the object.
Currently the node will not work, because the curve is not yet registered
in the globals map. This should happen in conjunction with depsgraph
updates.
Global data is used primarily for blend_data pointers (objects primarily).
These should not be stored directly inside the function instructions,
but should be stored by a key (hash value) and resolved at eval time.
These are generated automatically as terminators in pass node chains.
No pointer/mesh values will be stored in instructions though, the
sockets are just dummies to ensure a consistent value node signature.
Pointers in the instruction lists are bound to be invalid most of the time. They should only be
supplied externally via graph input arguments.
Note that opcodes still exist for POINTER/MESH, but these always store a NULL/empty DM
default value, rather than reading from instructions. They are only used by the code generator
as fallbacks if such sockets are not connected.
In combination with implicit float->vector conversion these can also
be used for simple scalar multiplication. This is not quite ideal
because of the conversion, but math nodes have to be redesigned at
some point anyway.
These are like value nodes in that an output socket stack entry is reserved for them.
However, like 'pass' nodes they don't actually perform an operation. Their purpose is
to provide a place on the stack for input arguments that other nodes can then read from.
Transforms can be created as translations, rotations (euler and axis/angle)
and scale.
The UI nodes include and automatic multiplication node for convenience.
This is not particularly nice, because it relies on every interface variable to call the
update explicitly. Eventually such things should work via the depsgraph, but this is
currently not easy to do.
The node group instance prevents recursive groups by disallowing nested
node trees, both when selecting a different group tree for an existing
node instance, and when adding an existing node to a tree (copy/paste).
Keymap entries have to use explicit values for properties if they
are modified in any of the variants. Otherwise the value is not
reset to the default in subsequent calls once it gets changed.
The previous implementation was based on recreating links and did not work
very well when reordering connections, because the insert_link callback does
not allow changing the link target.
This is primarily a python implementation for UI convenience.
It makes the list of input sockets behave dynamically for typical list operations:
* adding an element at the end
* inserting an element
* removing an element
Each of these keeps the list compact and shifts remaining sockets, so users don't
have to move each individual connection every time.
Internally it still uses the simple binary combine operation, adding as many
nodes as necessary.
These update calls can have bad consequences such as removing node sockets,
which should not happen until after the operator is finished. This was already
questionable in the past. At some point all update handling for nodes should go
through the depsgraph rather than the current fragile update flag/callback system.
The internal bvm nodes require strict type matching. Any implicit
automatic type conversions should be added on the level of DNA->bvm
conversion. This keeps the bvm nodes simple, and also allows for
semantics on the DNA node level that could not be properly implemented
in bvm.
This will be useful for separately generating kernel functions, which
are to be stored as own code blocks so kernels can call them repeatedly
as needed.
Only kernel nodes will be allowed to have function inputs. This is
necessary because any kernel node will define the boundaries of
node subtrees used for function inputs.
VALUE_CONSTANT inputs can not have connections, their values are known at compile time.
bvm stores these values directly in the instruction list.
VALUE_VARIABLE inputs store values at evaluation time. bvm stores a stack index for them
which is used to load and store from/to the evaluation stack.
VALUE_FUNCTION inputs are passed to kernels to be called internally when needed.
bvm stores an instruction index and allows a jump to this instruction (TODO).
This is currently always 0, but will be used in the future to allow
storing per-element functions along with the main code. These can then
be called within kernels.
This creates a (minor) mem leak when statically creating the mesh. Instead it
should be constructed in a init/free pair of functions.
Because this variable is used in all different compile units, each of them has
to create their own global 'empty_mesh' variable.
For constant inputs no stack index is needed. Previous to node pruning
this did not trigger the assert because there were still pass nodes
allocating the stack indices.
This is really awful, but we don't have a good generic way of defining node tree relations
in the dependency graph yet. More importantly, the notifier system is not up to the task -
putting the responsibility for notifying into operators requires anticipating their effects,
which is not good at all for node operators which affect geometry through complex processing.
The depsgraph really should be handling this.
This is a hack! The nodes are defined in python and the notifier is currently
triggered far to broadly, but currently there is no nicer way of doing it.
This generally triggers a mesh data recalc when *anything* inside the
object node tree is changed. Eventually it should be more specific to
changes in a geometry node tree, but the current design of deps nodes
does not reflect such internal node semantics nicely (although new
depsgraph will allow us to do it properly in the future).
Any extensive data stored as a pointer on the stack should be ref-counted,
so that it stays alive only as long as used by some node.
The way this works is by enclosing instructions that access the stack
entry with 'init'/'release' opcodes. These are added automatically based
on output socket types (nodes don't have to be created explicitly for them).
These nodes don't have opcodes any more, they are merely placeholders
to simplify node graph construction. Ultimately they will be removed
in the 'finalize' function, along with any other unreachable nodes.
This is necessary unfortunately because of the unreliable nature of dm->needsFree.
Without it the BKE_object_free_derived_caches function can destroy our static
empty DM at any point.
This only applies to a handful of common types that can be unambiguously converted.
In any other case the link is considered invalid (proper error handling still needed).
This has several advantages over doing the same in C/C++ code:
- The two parts of the code are clearly separate (parsing DNA settings vs. building and evaluating)
- UI constructs with modes, special cases, etc. can be done in py, heavy lifting still happens in C
- Scripters can easily add own nodes without having to deal with internals
lookup rather than a plain ListBase.
This paves the way for future inclusion of evaluation context info in
the EffectorContext struct, and hides some of the annoying details of
the ListBase double dereferencing.
This library provides a way to register modules and functions. Functions are simple
opcode-based descriptions of functionality (programs). They can be evaluated with
a simple interpreter using monolithic kernels, or could be compiled into more
optimized languages via LLVM or OpenCL.
The recursive registration for functions uses a simple int array for
flagging found functions. This is shared with the parent, so any child
functions will overwrite parent function flags for the same index
(this is not overloading).
The patch now requires using a larger array that can accomodate all
registerable functions from parents and children. It puts flags for
parents first, then children.
This includes a generic 'id' pointer for a single ID datablock pointer
in the node. The ID type can be defined using the `bl_id_property_type`
registerable attribute.
Based on the ray casting technique described here: http://www.joesfer.com/?p=84
Volume samples currently are not bound to mesh deformation in any way, they are
simple object-space locations. Potentially a harmonic weight approach as in
mesh cages could be used.
Currently strand edit mode tool use the object data 'x mirror' flag.
This flag is only available on mesh objects, but the strand editing
can be used for dupli override caches etc. as well. Eventually strand
editing should get it's own independent flag to avoid this problem.
separate file.
This may seem a bit like overkill, but it helps ensure that no particle
depedencies messes up the strand editing code. The same will be done for
other use cases of the strand editing code in the future.
Note that currently this has virtually no effect, because the mirror
option relies on exact positions of vertices, which does not happen with
random hair placement (the add brush has no mirror option yet).
Eventually topological mirroring should help with this case, but is not
implemented for either old or new strand edit yet.
These are a modified version of their BMEditMesh counterparts.
use_topology is not yet implemented for strands. Native strand topology
is not very useful for this. Instead, the topology of the scalp mesh
should be used for finding mirrored strand roots, then the arc- or
parametric length of a vertex from the root to find mirrored verts.
Conflicts:
source/blender/blenkernel/BKE_editstrands.h
When copying mesh data to bmesh the MVERT and similar customdata types
have to be omitted. Otherwise the bmesh instance ends up with NULL
pointers in customdata layers, but entries in the typemap != -1. The
effect was that when storing new steps after one or more undo, the
resulting original data would be copied, and subsequent undo steps
are ignored.
data masks explicitly.
A dummy mesh is used for strand edit undo storage like in mesh edit
to prevent unnecessary code duplication. However, when copying from/to
BMesh only the mesh data layers are copied by default, omitting the new
data layers for strands (currently only MSurfaceSample hair root data).
This uses the generalized undo stack system which is also used for
object data edit. An extension is necessary to tell the undo functions
which object is specifically used for generating undo data and
identifying the stack, since strand editing does not edit the obdata
itself.
consistent segment lengths when transforming vertices.
Warning: The implementation is not correct yet, but all the steps should
be there.
The main idea is to treat strands as a sequence of joints that are
displaced out of their original locations by a transform or other tool.
The solver then tries to find a global per-strand solution that keeps
the segment lengths unmodified, with a minimum change in angles from
the original starting shape. Such a solution is much more usable and
efficient than the current O(n^2) attempt of "spreading the error"
across the strand.
The inverse kinematics method is very flexible. It can also include
stretching, which would be very welcome for features like the length
tool. Different parts of the strand could be weighted separately using
scaling factors for the angle/stretch parameters.
Conflicts:
source/blender/physics/intern/implicit.h
mode.
This method is simple, but not really very usable. It works by
successively relaxing segments that are too long or too short, moving
both vertices along the edge between them. This is repeated N^2 times
(N: number of vertices on the strand).
A true IK solver could give a lot better results, as well as providing
many opportunities to apply weighting for targets (e.g. preferring to
move non-selected over selected vertices). Many different methods for
simple IK solvers exist, so there should be one that works well for
large number of simple strands. See e.g.
http://www.math.ucsd.edu/~sbuss/ResearchWeb/ikmethods/iksurvey.pdf
vertex positions.
This works a lot better with strokes perpendicular to the general hair
strand direction. With the previous comb tool such strokes would only
make a small dent in the hair curve and then vertices would slip out
of the tool circle. The edge combing affects the local direction of
strands, which acts as a kind of grabbing functionality by moving
further vertices in front of the tool circle. The result is that
drawing a curvy hair shape with the comb becomes much easier.
In addition, the new tool also uses edge filtering and weighting rather
than vertices. This means that small brushes also work well, instead
of having no effect when hitting the edge segments between vertices.
Further improvement could be achieved later by using a global strand
solver, which adjusts vertex positions based on an error metric along
the whole of the strand.
direction vector for the hair stroke tool.
This is necessary for directional tools such as combing, where the
stroke direction can lead to unwanted results if it changes too
abruptly.
The rationale behind this is that the BKE code could be used for
modeling hair and fur as well as a number of other features such as
grass. The primary addition to BMesh is the limitation to strand-like
topology (simple vertex chains, optionally rooted on a mesh surface).
The editor code OTOH is quite hair specific, since the result should
be suitable for hair simulation, and the workflow should mimick actual
hair grooming intuitively. Eventually the hair edit mode could become
a generalized strand edit mode with slightly different tool sets for
various purposes, but for now it is quite specifically built for hair.
Conflicts:
source/blender/blenkernel/CMakeLists.txt
source/blender/blenkernel/intern/particle.c
drawing code for edit mode.
The lighting uses the same method as the particle hair, interpreting
strand direction as the normal. This is nowhere near realistic hair
shading, but doing this with line shading alone in the fixed-function
pipeline is probably not possible. A GLSL shader could be used instead
for a more realistic alternative as part of the viewport project. At
least this simple shading mode gives some orientation feedback while
editing hair.
Currently uses the same approach as old particle edit mode (rescale
hair segments from the root on). A more sophisticated approach using
least-square error minimization of the displacement could yield better
results.
Stroke tools will be categorized by the hair elements they operate on:
vertices, segments or strands (roots). In addition to that a filter
function defines the influence of the brush. This should be defined by
the other brush settings and be largely independent of the main tool
mode.
This requires converting the old messy particle num/num_dmcache/fuv/foffset
data into the new mesh samples, which can potentially introduce floating
point errors and inaccuracies due to lack of face index mapping in the
new system. However, in any well-constructed particle system the hair
roots should be nearest to their num face, so mapping would be accurate
enough. If necessary face index data could be added to samples as a
legacy code hack, but probably it's best to eventually replace the
hair system as a whole anyway.
Currently particles only have a single weight value, controlling
pinning in the simulation (for root and virtual root verts) and goal
force scaling. This will be replaced by a full vertex group weight
system eventually.
This is more for testing purposes, since currently there is only a
single mass property for the psys as a whole. This should change in the
future though, to allow variable mass per strand or vertex.
Conflicts:
source/blender/bmesh/intern/bmesh_interp.c
Hair/Strand editing will only use a subset of the bmesh topology and
expect a specific topology that needs to be verified and enforced.
However, this extra requirement is much less work than reimplementing a
whole edit data system with the same feature set as bmesh and avoids
much redundant code.
Conflicts:
source/blender/blenkernel/intern/customdata.c
source/blender/makesdna/DNA_customdata_types.h
This means using mempools to store curve and vertex data, which allows
arbitrary addition and removal of data more easily. Also this includes
an iterator system similar to bmesh iterators (although the simpler
topology makes it a lot less complex).
Conflicts:
source/blender/blenkernel/intern/customdata.c
source/blender/makesdna/DNA_customdata_types.h
particle hairs remain visible.
Note that currently the hair root location (num/num_dmcache, fuv, foffset)
is not stored from edit data, so all hairs end up in a default location.
This makes it work more like editmesh, and avoid the awkward and
basically bad-level approach in particles, where the edit data is an
anonymous pointer in particle systems together with a callback for
freeing.
Conflicts:
source/blender/blenkernel/CMakeLists.txt
sampling system to keep the code simple.
Now there is a MSurfaceSampleStorage struct that encodes the storage
details, which the algorithms don't have to care about.
This code is not yet used by any feature in master, but has proved very useful in the Gooseberry branch and several WIP patches. Inclusion into master was requested to avoid diverging code in branches and get part of the Gooseberry code merged early on.
The core idea is to have a common way of identifying points on mesh surfaces. Efficient evaluation of such points after deformation and mappable modifiers (e.g. subdivision) tracks positions, normals and other surface attributes. This is used extensively in the particle/hair systems and could be useful for certain tools, duplicators, etc.. Currently the particle systems in particular are very fragile and succeptible to mesh changes because of the unreliable way mesh surface mapping is stored. While the new system at this point uses essentially the same method, it at least provides a centralized DNA type that can be improved easily without too much effort and compatiblity issues.
In addition to evaluating surface samples the system also provides a central place to implement surface sampling algorithms, such as random sampling, ray casting, paint-like tools, etc.
Differential Revision: https://developer.blender.org/D922
sampling system to keep the code simple.
Now there is a MSurfaceSampleStorage struct that encodes the storage
details, which the algorithms don't have to care about.
@@ -12,13 +12,13 @@ base class --- :class:`PyObjectPlus`
..attribute:: name
The name assigned to the joystick by the operating system. (read-only)
:type:string
..attribute:: activeButtons
A list of active button values. (read-only)
:type:list
..attribute:: axisValues
@@ -27,8 +27,10 @@ base class --- :class:`PyObjectPlus`
:type:list of ints.
Each specifying the value of an axis between -1.0 and 1.0 depending on how far the axis is pushed, 0 for nothing.
The first 2 values are used by most joysticks and gamepads for directional control. 3rd and 4th values are only on some joysticks and can be used for arbitary controls.
Each specifying the value of an axis between -1.0 and 1.0
depending on how far the axis is pushed, 0 for nothing.
The first 2 values are used by most joysticks and gamepads for directional control.
3rd and 4th values are only on some joysticks and can be used for arbitary controls.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
