Basically, due to new 'virtual' libraries & 'path' assets, we consider those 'virtualmy linked' IDs as:
* Local for editing purpose (i.e. they are editable).
* Linked for datablock management purposes (i.e. they can be made local, etc.).
Note: all this is more like a quick hack to test 'file-based' asset repositories (like cloud)
in comming weeks. I really do not think we should use that in the end, an full-featured
'overriding ID' system (as proposed in replacement of proxies for 2.8) would be much saner imho.
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.
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
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
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 :)
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
- 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.
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
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.
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
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).
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.
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!
* 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.
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.
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 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.
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
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
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
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.
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.
Since this file mostly defines/handles interactions with asset engines...
Asset handling itself is more done in library area (will split it too in next commit).
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).
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).
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.
Note that a few sub-data (like animdata) needs to keep option to unlink their
own ID usages themselves for now, this is not ideal, but need some more time
to mumble on this and find an elegant fix. Nothing blocking here anyway.
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.
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.
Idea is to avoid the need for asset engines to generate whole .blend files
with relevant datablocks in case they just need/want to provide files (images, videos, sounds...).
This commit introduces the concept of 'virtual' library, which in fact only exists to store
asset repository data. Datablocks from those virtual libs are then fully written in .blend file.
Much to do still, this is only preliminary/experimental work.
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.
Asset engines are now only selectable and used in link/append case. Makes no real sense to have
those in other cases (like regular file opening, or even worse when saving a file!).
Also allows to get back asset engine persistance accross filebrowser usage.
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.
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
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.
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.
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 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.
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)...
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
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]).
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
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.
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...).
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.
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.
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.
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.
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.
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
Dead dummy simplistic test runs (nearly) OK here (only have a memleak from versionning code...),
but obviously will need much serious tests to validate everything.
Also, many things in code will need cleanup/refactor. Not to mention UI/UX.
But we do have asset version check & reload on file opening now! :D
Process summary:
I) Open .blend file exactly as in master, load assets just like any other linked data (add placeholders in case not found).
II) Check all asset engines and ask them whether each asset is OK, needs to be updated/reloaded, is missing...
III) User then can decide to actually reload the asset, or not.
Step II) happens in an async job (since engines using remote storage/database could take some time to answer),
step III) is a locking task (just like initial linking). This should allow user to not be blocked at all by that
asset update process.
With current master this is not an issue, but with all the reload stuff,
a same main could end up going several time into 'do_version', which is absolutely
not desired (multi-allocations, repeating conversions over some values, etc.).
Whole 'database' is rebuilt later anyway, and in more complex 'asset reload' cases
(where we may actually be reloading same asset with same uuid from same lib file),
this creates confusion...
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.
Idea here is to both factorize it seriously (reload and relocate are very similar, let's
avoid as much as possible twice the same code...), and make new `lib_relocate_do()`
much more flexible - now it should be able to reload individual IDs as well
(plan is to use it to reload assets in relelvant branch too).
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.
So far, we were accepting the idea of several repositories per asset engine (e.g. different
directories in case of Amber). However, if this works OK during listing/initial linking,
it makes things way more complicated later in asset management, since asset uuid is no more
enough to unically indentify an asset!
Now an asset UUID is assumed to be really unique inside an asset engine.
For Amber, we 'salt' assets' uuids with new repo uuids (both being only 2 ints long now),
and keep a config 'cache' of mapping repo uuid -> path to repo.
- 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.
Fetching assets' uuids by engines is now in own func (will be used by update op itself too).
Also, each AssetRef has only *one* asset ID, the first one, avoids us a useless loop!
And early out of update_check op in case we have no asset in current Main, no need to start
job in this case.
The issue was caused by missing indirect dependencies from other scenes in the
current scene depsgraph, which was causing some threading conflicts.
Not sure what would be ideal solution here, for now use stupid but rather robust
approach to solve that. Maybe there's something better to do here tho.
Currently a lot of the nodes in the new dependency graph are empty placeholders
for organizational purposes. These nodes would, however, still be assigned a task
which gets scheduled and takes up some time for worker threads to pop from the
queue and run. This can be avoided by skipping these nodes during depsgraph
scheduling, and scheduling their childrent right away. Gives a few percent speedup
in BlenRig.
Some of the tagging functions would be called for every operation, and then
in turn tag their ID nodes with all their operations again. With extensive rigs
we get ID nodes with a lot (10,000+) operation nodes, which leads to millions
of unnecessary tagging calls.
Previously, adding a new vertex with Ctrl-LMB would move an existing
when there was an active spline but no active point.
This function is used as a fallback, extending an existing active point is already handled.
Request by Andy, should help him a lot doing weird and wonderful hair styles.
A bit experimental yet, details of behavior might be changed after some real
usage feedback.
Annoying bug caused by temp nature of looptri layer for CCGDM.
Fixed in a similar to CCG loops by using lock when allocating and
filling looptri arrays.
Real fix would be to make sure this array is allocated on object
evaluation using DAG's eval_flag, but that's more involved change
which we'll work on later.
Operator would call `WM_file_read()` directly whith G.main->name as filepath, which
gets freed whith main during new reading of file... Now use a local copy instead.
This is a follow-up fix for rBa10b2fe. The Action constraint uses a hackish
stub object and pose, which doesn't have a hash table for fast lookups.
This doesn't seem to be a big issue with the old depsgraph, but in the new
depsgraph it creates a large number of cache misses and significant slowdown,
possibly because of additional threading and less simple bone loops.
Normal Map node support for GLSL mode and the internal render (multiple tangents support).
The Normal Map node is a useful node which is present in the Cycles render.
It makes it possible to use normal mapping without additional material node in a node tree.
This patch implements Normal Map node for GLSL mode and the internal render.
Previously only the active UV layer was used to calculate tangents.
- break when object in hit-buffer.
- don't measure distance to object centers that can't be projected.
- take pixelsize into account for distance pixel distance limit.
The hash table is used to look up bone poses, particularly during constraint
evaluation. Without this the default BLI_findstring method on a plain ListBase
is used, which is really slow for extensive rigs.
Using 'shell-thickness' to offset UV's meant very sharp corners would offset far outside the image
causing project-paint to hang while collecting all pixels for each UV face.
Clamp the maximum offset to prevent this.
Couple of issues here:
- Was a bug in heap memory allocation when run out
of allowed stack memory.
- Debug MSVC was failing because it uses separate
allocator for some sort of internal proxy thing,
which seems to be unable to be using stack memory
because allocator is being created in non-persistent
stack location.
Part 2 of Fix for T48200. Without this, it's not possible to exit GPencil edit mode using tab key with pies enabled.
Menus are somehow similar to operators, which let the event pass through too after operator poll returns 0. So think we should handle them similar here.
Can't guarantee this is totally safe though ;)
This fixes a lot of things in NLA RNA update handling (which basically did not update anything previously).
There are more update issues with this editor though...
Issue was that before writing to disk grids are clipped against the
density field's tree to optimize for memory/disk space, which in the
case of simulations with no density field results in all grids having
empty trees.
For now avoid clipping against empty grids, but perhaps in the future it
can be interresting to have some UI parameters to let the user choose
the grid used for clipping.
This mainly touches extern libraries and few debug-only places in intern.
Some summary:
- External libraries are not strict at all about missing declarations,
so we can rather safely remove such warning together with other strict
flags.
- Bullet has some static functions which are not used.
Those were commented out.
- Carve now has some unused debug-only functions commented out as well.
While we're on the way of getting rid of Carve, it makes sense to make
things a bit cleaner for the time being.
- In LZMA we have some parts disabled which gives some set but unused
variables which is rather correct.
- Elbeem had quite some variables set and never used because their usage
is inside of debug-only code which is commented out.
Note about patching upstream libraries: surely one might say that we
have to make local patchset against this, but own experience says it
only gives extra work trying to merge such tweaks to a new upstream
version and usually it's just faster to re-apply such fixes again after
bundling new upstream library.
Mainly makes logging less verbose when doing progressive sampling in viewport.
Such kind of verbosity is not really possible to be filtered out with `grep`
so let's reshuffle few lines of code.
This introduces a snap-context that can be re-used for casting rays into the scene
(by operators such as walk-mode, ruler and transform code).
This can be used to cache data between calls too.
The script is updated for ArchLinux, since all dependencies are included in Arch's official repositories.
I've made a few changes, such as enabling OCIO and OSD without requiring locally installed lib-path, and a fix to ''get_package_version_ARCH()' so it ignores package epoch (as in the case of ffmpeg).
I intend to look at OpenVDB next.
Reviewers: mont29
Differential Revision: https://developer.blender.org/D1901
The 'precision' and 'dynamic' settings for binding are now always visible.
The settings that can not be edited after binding are disabled (not inactive).
I find it useful to see with what settings a mesh was bound, in case
the file is not mine or if I simply lost track of it.
Simple idea, use threads when dealing with "Copying Transformations to device"
scene update step. Only do it if there's enough objects in the scene.
Hopefully only brings less synchronization time and doesn't break anything.
From tests on my desktop this brings down transform update time from 58sec to
11sec on victor_cpu.blend scene from out benchmark.
This is an attempt to gracefully handle out-of-memory events
and stop rendering with an error message instead of a crash.
It uses bad_alloc exception, and usually i'm not really fond
of exceptions, but for such limited use for errors from which
we can't recover it should be fine.
Ideally we'll need to stop full Cycles Session, so viewport
render and persistent images frees all the memory, but that
we can support later, since it'll mainly related on telling
Blender what to do.
General rules are:
- Use as less exception handles as possible, try to find a
most geenric pace where to handle those.
For example, ccl::Session.
- Threads needs own handling, exception trap from one thread
will not catch exceptions from other threads.
That's why BVH build needs own thing.
Reviewers: brecht, juicyfruit, dingto, lukasstockner97
Differential Revision: https://developer.blender.org/D1898
This mimics behavior of default allocators in STL and allows all the routines
to catch out-of-memory exceptions and hopefully recover from that situation/
So now, jobs callback (async ones) of asset may return a specific job id value in case they actually
complete (or fail) on the first run.
This allows engines that do not need slow async stuff to perform an action to use simpler code
(e.g. imagine an engine able to list its assets from a cached DB, in most cases its list_dir
callback can execute instantaneously (from a user PoV), no need for a listing job then).
Reason is, even if paths returned by `list_dir` are OK, we will also use `load_pre`
during 'reload'/'update' scenarii, where we only have uuids, to retrieve final
.blend datablock paths.
Code only reports which datablocks/assets shall be reloaded etc. so far, actual
reloading needs id-remap code (and hence will be implemented in asset-experiment
branch later).
This commit:
* Fixes bad handling of 'stop iteration' (by adding a status flag, so that we can actually
stop in helper functions too, and jumping to a finalize label instead of raw return, to
allow propper clean up).
* Adds optional recursion into 'ID tree' - callback can also decide to exclude current id_pp
from recursion. Note that this implies 'readonly', modifying IDs while recursing is not
something we want to support!
* Changes callback signature/expected behavior: return behavior is now handled through flags,
and 'parent' ID of id_pp is also passed (since it may not always be root id anymore).
Used this new recursive behavior in ID preview generation as an example, would obviously be committed separately.
Reviewers: sergey, campbellbarton
Differential Revision: https://developer.blender.org/D1869
This commit:
* Fixes bad handling of 'stop iteration' (by adding a status flag, so that we can actually
stop in helper functions too, and jumping to a finalize label instead of raw return, to
allow propper clean up).
* Adds optional recursion into 'ID tree' - callback can also decide to exclude current id_pp
from recursion. Note that this implies 'readonly', modifying IDs while recursing is not
something we want to support!
* Changes callback signature/expected behavior: return behavior is now handled through flags,
and 'parent' ID of id_pp is also passed (since it may not always be root id anymore).
Used this new recursive behavior in ID preview generation as an example, still needs more testing!
We need that info (knowing which non-asset IDs are used by which asset-IDs),
otherwise managing reloading, updates etc. of assets would be impossible
(or rather, would leave a mess of unused IDs behind them).
Only partially done, non-functional yet.
This commit also adds asset repository references in libraries.
Note that this is only very first steps, I think we'll need to 'propagate'
uuid to all data used/linked by asset itself, otherwise most things won't get
updated/reloaded. Still have to think about this.
Note that this is only very first steps, I think we'll need to 'propagate'
uuid to all data used/linked by asset itself, otherwise most things won't get
updated/reloaded. Still have to think about this.
Issue here is again with NEVER_NULL usages: obdata (& co) would be dereferenced twice.
I'm not totally happy with this solution, ideally remap should never leave Main
in invalid state, but for now it will do...
* We need to clear the 'user-one' flags and potential usercount for old_id.
* We need to consider ID usages inside a same lib as **never** being indirect ones.
Also re-enabled proxy checking for now (not sure why I commented it, think testing stuff or so...).
Seems to work for simple cases, but deletion of complex libs from complex files
(Gooseberry ones) still generates lots of assert failures, and crashes in some cases.
This is hacky to make work on Linux, and seems to totally break on Windows.
So now, instead, we define a CollectionListBase (exact copy of ListBase), to be used inside RNA...
Seems to work nicely on Linux, lets see what win buildbot says.
issues with unlinking ParticleSettings.
We now have the option to still decrement user count of old_id, even if we could not
replace it with NULL new_id (aka unlinking), needed to keep proper user count when
unlinking and ID from datablocks it uses, to delete it.
Conflicts:
source/blender/blenkernel/BKE_library.h
source/blender/blenkernel/intern/library.c
Also clears now useless 'bool is_user_one' flag from editors' ID remap callback.
Idea is to add two new flags, one saying 'we need to ensure a real user exists',
the other 'we had to increment user count to ensure we have a real user'.
This allows us to easily control the extra user in release/delete/remap cases,
and also fixes the infamous 'add new image to texture, open image in ImageEditor,
delete image from texture, have a zero-user red image in Image Editor' issue.
There is still much to be done here, more places where we can use those flags,
also clear them when we force usercount to zero, etc.
All this allows us to fix unsolvable issues (like Group being ensure_user'ed
in loading code, but only if they do have objects in them), and to avoid returning
ugly bool from editors' callbacks (this is still to be cleaned up in the branch too).
Bad news - this means we cannot use short ID->flag anymore (not enough flags), for now
added a new int ID->flag2 (replacing pad int), not sure how to best manage change here,
maybe for 2.8 we can totally wipe ID->flag? But this would totally break forward compat.
So, idea is, since mat/tex/scene/etc. nodetrees are owned by their respective IDs
(those nodetree do not exist in Main, they are systematically freed with their ower IDs, etc.),
we should not treat them as IDs in IDlooper, but rather as mere sub-data, and hence directly
loop over the IDs of those nodetrees.
From quick check it seems to work, but this needs to be confirmed as a valid idea!
Heavily simplfies that freeing area, but again a rather risky and likely-to-break change.
At least, NodeTree is known to be a trouble-maker here, due to how it seems to be
'owned' by its mat/tex/sce/etc. Have to dig deeper here, this is still quite unclear
what exactly happens with those, and how to handle them correctly.
(instead of whole Main content).
Could replace maybe things like constraint, rigisbody world, etc. '_relink' func,
but for now it's only intended to replace custom 'id releasing' code in _free()
funcs of all IDs!
Previous situation was pretty much horrible, a few data types having their own coocking to
unlink, often doing more than only unlinking, often doing the same thing two times or more,
often messing with areas they should not have touched (like editors from withing BKE)...
Now we hope to have something generic, working the same way for all ID types
(we do have to add some specific handling for groups/objects/scene unfortunately,
but this remains reasonable).
Needless to say such a change is calling for troubles - I tried to follow and reproduce
as best as I could previous code, but most likely some areas will become buggy. Do not think
previous code was 100% correct anyway, things like Objects have a really big and complicated
usage of IDs...
Also, foreach_id tool has been enhanced, again for complex types like objects & co, we should
cover much better all possible IDs now (e.g. rigidbody & logic bricks were
totally missing from there).
And there is more to come...
Think core of the system is getting pretty much OK, now we'll likely have to deal
with tons of specific cases, given how Blender is totally inconsistent when it comes to
handling user counts (e.g. spaceimage...).
Also, still have to make generic ID_unlink and ID_release_datablocks func to replace
redundant code in BKE we have currently.
Still missing: at least modifiers (have to change all modifiers callbacks, yuck).
AFAIK constraints never refcount thier IDs so we can keep current code here.
Similar changed to those done for mesh/curve/mball.
Also, systematically nullify pointers in `_free()` functions, this does not cost
much and can help troubleshooting later.
And tag `BKE_object_unlink()` as an horrible piece of code...
Renamed their `_unlink()` functions to `_release_datablocks()`, since they do not
unlink anything (not in the sense `BKE_object_unklink()` does, at least)!
Also, added a `const bool do_id_user` to their `_free()`, for consistency, now
relevant `_release_datablocks()` func only gets called when this option is set.
They would be removed from bmain, but never added again nor freed - in this case
we can simply explicitely reload them in fact, non-linkable is only a user-related status...
We need to append to linklist, or our library indices would be reversed compared to lib list...
As for locking, we need reentrant one here if we want to use it. :(
Performances are not an issue at all here, but it makes code slightly simpler, avoids
a loop over ghash of libs, avoids some potential mem fragmentation, and will be easier
with lib relocate code too in future...
Reason is, on next fileread, proxy will be assigned with its target's data again...
Note that this may not be true about all its datablocks pointers - but proxy objects
are really brittle in current code anyway, they could use a serious work as well.
Basics seem to be working, proving idea is valid. There is much to do yet though,
handling correctly all possible scenarii (especially those including indirect libs
used by several libraries, or libs used both directly and indirectly, etc.)
is not going to be a piece of cake...
Making the code in WM area more generic (will be used by relocate libs feature too in future).
And multi-append/link shall no more close & re-open lib files for each and every idcode!
Also, serious cleanup of BLO_append_... API (renamed BLO_link_..., since it links and never appends!),
main real changes there are removing the bContext arg in favor of scene/v3d (a bit more verbose, but does
not requires a valid context anymore to be able to instanciate ob/groups!). And logic behind instanciating
or not has been simplified, previous code was really obscure and sometimes redondant, from quick tests
it still works as expected.
This is temp branch for until 2.76 is over, shall be merged in master asap then.
Note that this op will have to be moved to WM area in fact, it's way too generic for Outliner only...
and we need to re-use most of append/link operator code!
We have to take care of Object->data aside from main foreach loop, since we can
only replace that pointer if object is **not** in Edit mode!
Also, added a "Remap" action to ID context menu of outliner. This is **very far**
from nice and pretty code/feature, but it allows to quickly test the code.
Only did very quick tests with meshes so far...
This seems to be OK, but will obviously need much more testing - and is useless as-is, we need
the big ID remapping code itself still.
Also, Nodetrees' remapping of Scene IDs is still TODO, not sure yet how to do this.
Needed, since trying to (re)init data already set could lead to many issues,
and handling this is not in the scope of this function.
Note that definition of BLI_memcmp_null stuff is rather bad currently, no good idea where to place this. :|
Just having 'NULL' data as placeholder is not really possible, too much places
to check against that.
WIP commit, still much to be done!
See T43351 for details.
Differential Revision: https://developer.blender.org/D1394
Idea is, instead of ignoring completely missing linked datablocks, to
create void placeholders for them.
That way, you can work on your file, save it, and find again your missing data once
lib becomes available again.
Plans are also to be able to locate missing libs and reload them at runtime.
To support all that, we must be able to make Blender survive those missing data
(i.e. empty datablocks) all over the place. This commit contains some early work
in that direction, but this will need much much more work!
Indeed, though rather unlucky, we may end up handling same source file from
different preview threads, which could lead to conflicts and bugs.
So idea is to add a simple way of locking a given source file path, since
thumb handling of different paths shall never conflict.
Note that this adds some 'generic' stuff to GHash and Threads area, that are to
be committed separately of course.
The task was not so annoying (workers are just sleeping anyway), but having the timer
always checking for updates was not so nice. This remains quite minor optimization probably...
So now, we do not start anymore preview task when enabling previews. This is deferred
to the moment we really need it (i.e. when we cache block of entries currently visible).
Preview timer is started at the same time.
Further more, when preview task has been inactive for about one second (i.e. all
queued previews have been handled, and no new one have been queued), we free it
(switching to some kind of 'standby' mode).
This also allows to avoid stoping/restarting preview tasks on each redraw when user is scrolling.
Classical bool vs char value mismatch... Here it was leading to a constant re-filtering
of the list, which invalidates cached previews.
Strange thing is, it would look as if Windows CMake builds are using 'real' bool,
while scons ones are using fake char-based one?
Instead of computing MD5sum of paths, just use incremental UUIDs!
There meaning and lifetime for internal listing is very limited anyway,
they shall only be unique for a given listing.
Note about using atomic op here: did it because it's nice use case,
not sure it's worth adding atomic stuff to /editors area though?
This commit replaces `filelist_numfiles()` by `filelist_files_ensure()`. The later
also returns number of visible files (i.e. filtered ones), but also ensures filelist
is filetered and sorted.
Otherwise, during listing process, operators could be executed between an update of the listing job,
and the refresh from the file space - i.e. at a time filelist is virtually empty (from filebrowser PoV).
Also, got rid of 'need_refresh', this ended up being unused...
* Do not systematically clear selection_state's uuids GHash when updating filelist
during listing process - uuids used here should remain valid.
Allows to not lose selections during listing background job!
* Use new BLI_ghash_lookup_p to handle selection state setting, symbolic optimization but...
A file which does not have *any* preview is perfectly valid, error is only when there
are some previews, but not the same number as actual datablocks...
Conflicts:
source/blender/editors/space_file/file_draw.c
source/blender/editors/space_file/file_ops.c
source/blender/editors/space_file/space_file.c
NOTE: Tagged some parts in new code to be fixed, rather do that in a separate commit.
This means this commit is broken and won't compile.
Also adds some 'state' flags to AssetEngine, so that custom AE's setting changes
can be notified to filebrowser, as default internal ones are.
For now, we only have 'dirty filtering' and 'dirty sorting'.
With big screens and tiny drawing, we can show much more than 1k items at once...
Also allows to reduce number of cached items in case we show few things big!
Mismatch in 'owner' of job, not giving right one to stoping helper, sigh.
Note that this was also affecting any 'interrupting' task (like changing level of flat listing...).
We cannot keep 'stable' list of items during listing process (each time new
entries are added, we have to filter and sort the whole lot again, which
means rebuilding 'draw' data from scratch - previews included).
Just wait until our filelist is complete to generate/use previews.
* Let asset engine draw most of header bar itself!
* Pass whole FileSelectParams to AE's sort_filter function.
* Move 'use_library_browsing' to params level (this way AE's sort_filter can be ware of it too).
Amber:
* Now supports real basic sorting/filtering modes
* Also basic support of tags (needs a way to refresh though, currently).
this is mandatory, since we do not keep previews for all listed items anymore
(with 10k datablocks would start to take too much mem). So we now generate (read)
them from thumbnailing threaded task as any other image, has several advantages:
* Quicker initial listing of data from .blend files.
* Since thumbnails are cached on disk, re-generating them is rather quick.
* General consistency!
Note thumbnail generation of datablocks could (should!) be enhanced, right now
it reads all previews for a given datatype in a file each time it has to generate
a single datablock thumbnail, we should generate all of those previews in a single
run. But this means handling cache dirty things (TBH_FAIL, outdating, etc.) on another
level so need some more work.
Also, still some memleaks issues with thumbnails, have to track those dose!
Since with internal listing we have no choice but to list everything, store
those data in the smallest possible struct - note we do not even store preview
here, these are only handled in full FileDirEntry cache (re-loading them from
cached thumbnails is really quick anyway).
Still not complete, have to re-do ID previews now, among other things...
And some memleaks again with previews, ugh.
Those were eating quite a bit of bites, not neglectable with thousands of entries...
Also, useless to print those info most of the time, they are not used!
So now, we only store 'raw' stat data, and expose funcs to convert them to readable
strings on demand.
Since we do not store anymore ell entries, we cannot use them to hold selection status.
So we now use a ghash in FileList struct, with entry UUID's as keys.
Also cleanuped up / refactored selection handling, filelist.c now features a nice intern
api and all other code uses it instead of accessing directly entries' selflag!
Conflicts:
source/blender/editors/space_file/filelist.c
source/blender/editors/space_file/filelist.h
source/blender/makesrna/intern/rna_internal.h
Notes:
* asset-experiments branch still needs some work (e.g. entires selection is to be reworked too),
but was more than time to do a first merge...
* This is raw merge, code is hence broken and needs more fixes to take into account new changes!
Get rid of cache iterator code (not needed, and don't think we'll need it later,
easy to add back if needed). Also comment some debug prints.
New preview (thumbnails) handling seems to be working pretty nice now.
* Full path is no more static, so we need dynamic tooltips to use it (still an
issue with drag stuff here :| ).
* Remove timer would attempt to free timer's customdata, which is mere int-in-pointer
in case of new notifier timer, so we need a special tweak on remove here.
Start to looks good... This commit adds a new timer helper, that only send notifiers
(kind of very restricted subset of wm_job stuff), since using jobs for previews
would be now counter-productive and way too much heavy.
Also fix/enhance how previews are generated (order), etc. etc.
Still WIP though, having some weird crashes and such lurking around.
Note this is rather stupid anyway, we need to rewrite this from scratch,
using a job for that makes no sense now, we should stick to a lighter
threading schema with a queue, or something like that!
Core part seems to be working, but this is still heavy WIP, much to do
to resume complete features from existing code. Mainly:
* Previews (those should only be ran on cached items now).
* Use block-caching to load in-display entries.
* Rework things like selection (cannot store selected states in entires anymore).
Also, internal listing is stupid currently (since it still stores everything
as FileDirEntry & co), not really crucial currently, but ultimately it'll
use its own, compact struct to keep full listing!
Also did some minor cleanup/renaming.
That code is currently pretty stupid, but we need to clearly separate
what is handled by filebrowser itself, and what is by listing code
(be it internal one or future asset engines).
Plan/idea is to have a dual cache system (one 'block' of entries - centered
on currently viewed ones -, and one 'random' FIFO for random access to any entry).
This implies filebrowser itself shall now only be aware of the number of (filtered)
entries, and it then requests entries as needed.
Also, made size of columns in filebrowser fixed (ultimately, user configurable),
since it has no more access to the whole list of entries!
Still heavy WIP.
We cannot do that correctly with current listdir system, have to rework again
in asset-experiments brnach first, so that listing, filtering and sorting is
completely "delocalized" to asset engine (will also try to solve scalability issues).
Also, make clear tags are only informative here.
Reason is, if the project is a success, we have to think big and
imagine asset engines managing catalogs of tens or hundreds of thousands
of entries (web ones e.g.), we cannot hold list of those in filebrowser!
So most of filetring will be 'deffered' to asset-engine.
We'll also probably have to implement a 'chunck listing' behavior
(to only list at most n entries in a row), for same reasons.
This will make things slightly more complicated :|
Think issue also exists in master code actually, where some op properties
could be left uninitialized in 'params_to_op' helper func... But it does not
backfire like that for sure.
This will allow asset engines to:
* Actually ensure requested data is available (downloading it, generating it, etc. as needed).
* Change entries generated for user by actual data paths (allows e.g. to present
a set of different paths to load as a single 'asset' to user).
As usual, had to change existing filebrowser code... Seems to be working fine
from quick tests (i.e. regular filebrowser not being broken).
Basically, filelisting code in space_file becomes default, internal 'asset engine',
and user can switch to any other available engine instead.
Note filelisting with asset engines is fully main-thread from C code PoV,
parallelization/asynchronism is responsability of the py code!
Also, added a basic (dummy currently) AssetEngine py implementation.
Not much there, but helps checking things works as expected before
writing whole interface!
All this is very basic early code yet, much much to do before we get anything really usable...
Remove everything not really related to OS representation of file entries.
FileBrowser will now use a complete decoupled version of file representation,
much easier to keep sane behavior this way, easy integration with RNA, etc.
This allows us to greatly simplifies sorting (and avoid recomputing UI name each and everytime!).
Also, correctly free entries/variants names/descriptions.
Was not sure about that (adds two pointers to entries, variants **and** revisions :/ ),
but with the incremental process of listing we are using now, it makes things sooooo
much easy to handle (and also avoids tons of (deep)copy and mem (re)alloc), that it's
definitively worth it.
And since most (if not all) of 'public' access to filelist is done through filtered ones,
we can easily keep this a mere array of pointers, and hence keep quick index lookup.
Also, quite noticiably simplified the whole listing code (less functions, less levels of subcalls...).
Note: for now we stick to mere alloc, we may benefit from either memarena or mempool,
but those are not threadsafe (which means we could not so easily pass mem from worker
thread to main one), so not quite sure it would be worth it.
Note: thumbnail stuff is to be reworked from scratch too, have the feeling it's not yet
100% thread safe, and using a job here is probably way overkill, since we do not show
any progress in UI.
Also, reduce revision's uuid to 64bytes, should be more than enough!
Thing is, we can expect tens (if not hundreds) of thousands of revisions from a
big assets repository with history storage (vcs). So we want that struct
to be as small as possible.
Also, we do not expect to handle several variants/revisions of a same asset within
a same 'context' (there is always only one active variant & rev). So storing
path & co at entry level shall be enough. And helps make code simpler/nicer.
We may suffer a little performance loss here, but would not expect it to be really important,
since rebuilding a full abspath is not needed that often.
Also, get rid of modes/owner crap, this is quite advanced UNIX fs info, not much to
do in our filebrowser imho, and takes time, UI space and Mem space!
Move typeflag and selflag back to FileDirEntry, and add blentype there too,
this will be needed by asset engines, and avoids rechecking lib ID types over and over!
There's much more to do here though. Also, FileDirEntryRevision is going to be
used quite higly, we want to keep it as light as possible (much lighter than
its current state for sure!). Think we can move back to FileDirEntry paths
(and try to get rid of abspath on the run, too), but also most likely
preview (we won't store a preview for all possible revisions anyway),
owner/permissions, etc.
Rev should probably only have its uuid, and date/time.
Also, simplified filetype generation (from extensions for real files, and BLO helpers
for libdata), now it's nicely encapsulated into each reading callback.
Main idea is that direntry had quite a bit of things already absolutely useless
in the mere 'file listing' context (image pointer, void 'data' pointer, etc.).
On the other end, to handle future asset stuff, we need *more* 'meta data' for our 'files'
(tags, variants, revisions...).
Further more, we need all this to be easily usable through RNA, i.e. we need it to
be defined in DNA.
So instead of trying to use direntry in DNA, mixing low-level system stuff like stat
where we do not want to see it, we define our own 'file' representation,
with only what we really need (who cares about chars or block or dev files in filebrowser?
we only need to know whether it's a regular file or a directory, period).
That way, low-level dirlisting remains clean, and we can easily extend our own
data as needed.
Note this new code works, does not crash (so far...) nor memleak, but it
needs quite a bit of work still, we are still storing useless stuff in direntry
(all those strings representing modes, size etc.!), and mem handling
during dir/lib listing is confusing at best (would like to use a memarena
or mempool here, but it's not that easy).
Passing around ID's themselves is not safe - they may change their mem addresses.
Yet, I had no issue on linux, but on win build this is still not enough,
have to use a super strange hack (add an 'import sys' line just before where error happens)
to get things running...
Batch preview generation can now also generate 'internal' previews (materials, textures etc.),
as well as scene ones. Also, some basic Cycles handling is there now.
You can now chose which kind of data to generate previews for, and whether the files are
trusted or not (for py autorun and drivers).
And rendering happens in background again, issue was actually a bug in creator.c, fixed in master.
Main TODOs reamining here: OpenGL rendering, and find a way to prevent '.blend1' when
saving over (either as a new option of save op, or simply by removing those files in script?).
Added a batch cleanup of previews too.
Also, added path of files as their tooltip 'image' button in main filebrowser view,
quite useful with the 'recursive' option...
Now, recursion 0 = previous behavior,
recursion 1 = only show content of immediate .blend files, do not recurse in actual dir hierarchy,
recursion >1 = complete recursion!
This will allow us to keep op panel independent, and have tabs for future assets tools!
Note 'conversion' from previous file area (channels) is not perfect, you have to hit 'T'
twice currently if you had a filebrowser in your .blend file...
Consider blendfiles and 'categories' inside them as regular dirs, but with this hierachy:
real dir > blende file lib > category in a blend lib.
This gives a much better sorting (among other things, gives more importance to
blend files!).
Got completely rid of C code for those panels.
Added full path in ttips of each 'bookmark' entry, as in current master.
Note this forced me to make tip in uiBut dynamic (like label). Not so sure
about this, but seems to me forcing tooltips to be static is actually quite
bad and limitating? we'll see...
Only remaining TODO (new feature) is reordering of bookmarks, should be easy now.
Also fixed some stupid (again!) mistake with RNA access to paths of bookmark entries.
* When overlaying icon, we have to also ensure background pict is always the same size
(else you get usgly icon size changes). Still not perfect though, suspect some more issues
in filrdraw code, probably should systematiacally run the 'ensure size' code during thumbnail
generation?
* Sharing of img ImBuf between thumbnail job and main thread could still happen,
leading to random crashes.
Note: unlike in current master, path is checked immediately when setting it
into filelist (simpler now that filelist building itself is threaded).
`filelist_setdir()` modifies in place given path if this one is not valid.
`BLI_fnmatch_strncpy_add_endswildcards()` will copy and add '*' wildcards on both ends of given string, if needed.
`BLI_fnmatch_strcmp_ignore_endswildcards()` is same as strcmp, but ignores any '*' wildcards on both ends of given strings.
New code shall be more easy to maintain and extend.
Sorting is now handled quite the same as filtering, and all filtering parameters
are now packed into a sub-struct to help extending it later.
Also done some optimizations in filelist refresh, and sorting/filtering area.
Now we should avoid re-sorting and re-filtering too often, also removed
calls to those in read_xxx funcs.
Note thumbnail job is still started basically on each call to `file_refresh()`,
will be addressed in next commit.
Doing it mostly because I need some update work done here so far to get py uilists/bookmarks working.
This exposes/moves part of bookmark code to ED_ area, so that it becomes available from e.g. RNA.
And adds RNA struct and such to represent a bookmark entry.
Also, Bookmark format is ammended to store optional custom names.
All this is still *very* crappy, should probably take some time soon to cleanup the whole branch!
See also T42777.
* Do not render cameras, lamps etc. for now.
* Generate bbox in camera local space, gives much better results.
* Add an op to clear previews from files.
Now for any object, we have an rna func to call with a set of (flatten :/)
coordinates, and we get back location this object should be to 'see'
all given points.
Would need more refinements, but will do for now.
First expose a method in Object RNA to get camera projection matrix (done).
Then, expose a method to get camera position (and scale for ortho) so that
it fits a given set of coordinates (still unfinished).
this among other things implied rewriting `BKE_camera_view_frame_fit_to_scene()`
underlying logic so that it uses `CameraParams` system, instead of `BKE_camera_view_frame()` one.
Note since this gives python possibility to change those previews,
we add flags to PreviewImage to tag as user_defined.
TODO: replace all those shorts-used-as-bools by bitflags!
Previuosly, we would only write existing data - and big previews were
nearly never generated by default.
Now, if saving previews is enabled, `wm_file_write()` will ensure
all 'common' previewable datablocks do have up-to-date previews
(brushes remain TODO here for now, they have both a preview **and**
an icon_imbuf, sigh...).
This implied adding an 'immediate' non-job-background version of
`ED_preview_icon_job()`.
Note this does not include objects/groups/etc. It's not simple to generate
previews for those, will probably go for a user-launched py operator here...
Also, cleanup some debug prints!
Previous implementation was uglyly relying on button using a specific
icon (VIEWZOOM) to enable this feature. Now we have a dedicated button flag
(UI_BUT_TEXTEDIT_UPDATE) and a dedicated RNA subtype (PROP_SEARCH) for that.
Also, if no icon is specified, it automatically set VIEWZOOM for it,
which saves us explicitely defining it in UI py code!
NOTE: this is to be committed asap in master, but has to wait for
after 2.73, so storing it here for now.
Also, try to make brushes also save their previews.
Anyway, something fishy is going on here, in theory we should get
previews for brushes, worlds, lamps... for free, but so far looks like
lib 'peek' code only handles mat/texture previews correctly?
Doing nothing useful yet, though!
Note: ultimately we may want to move that preview stuff to ID struct,
so that we get a real basic generic handling of it, but for now
simpler to keep it as is.
Also, definitively removed 'G.main.name protection' from lib reading
code. I do absolutely not see how listing a separate blend file handle
could possibly affect that anyway - and we do not want to access
G from non-main thread I'd bet...
Now it works reasonably OK for filesystem, and also for .blend libs.
Code is even quite simpler than previous one. Now it needs some serious cleanup,
and find where (new) memleak comes from too.
Only very partially working in this state, still needs some work.
Proof of concept seems rather interesting though.
Also, only pure filesystem stuff is implemented for now.
Also:
* Changed BLO_library_path_explode, now r_group and r_name must be pointers of char pointer,
if found they will point to right place in r_dir 'memspace'.
* Select everything in IDTypes filtering by default.
Nothing really magic here, just grouping the imports by libfile/grouptype.
We could probably get better code (loop less on given list of 'files' to import)
but not sure it'd be worth it, this is not supposed to be a performance-critical point.
* Delete filelist when changing flat setting (so that we get immediate update when changing that setting).
* Disable flat when going up in tree (we ;ight actually want to do this everytime).
Not much to add, we'll likely add more filtering options (tags, categories...) later,
maybe for now we just need an 'exclude' option for name filter...
Main goal is to see whether we can use existing filelist to handle that
(by default it is 100% tree-oriented).
Also, getting some strange issues with ID name currently...
2014-12-03 15:29:18 +01:00
1382 changed files with 93517 additions and 38609 deletions
set(PLATFORM_LINKFLAGS"${PLATFORM_LINKFLAGS} /Library/Frameworks/Python.framework/Versions/${PYTHON_VERSION}/Python")# force cmake to link right framework
# This is a workaround for our helperbinaries ( datatoc, masgfmt, ... ),
# They are linked also to omp lib, so we need it in builddir for runtime exexcution, TODO: remove all unneeded dependencies from these
execute_process(COMMANDditto-arch${CMAKE_OSX_ARCHITECTURES}${LIBDIR}/openmp/lib/libiomp5.dylib${CMAKE_BINARY_DIR}/Resources/lib/libiomp5.dylib)# for intermediate binaries, in respect to lib ID
# They are linked also to omp lib, so we need it in builddir for runtime exexcution,
# TODO: remove all unneeded dependencies from these
message(FATAL_ERROR"To compile WITH_GL_PROFILE_ES20 you need to set OPENGLES_DLL to the file path of an OpenGL ES 2.0 runtime dynamic link library (DLL).")
message(FATAL_ERROR
"To compile WITH_GL_PROFILE_ES20 you need to set OPENGLES_DLL to the file "
"path of an OpenGL ES 2.0 runtime dynamic link library (DLL)."
message(FATAL_ERROR"To compile WITH_GL_ANGLE you need to set D3DCOMPILER_DLL to the file path of a copy of the DirectX redistributable DLL file: D3DCompiler_46.dll")
message(FATAL_ERROR
"To compile WITH_GL_ANGLE you need to set D3DCOMPILER_DLL to the file "
"path of a copy of the DirectX redistributable DLL file: D3DCompiler_46.dll"
)
endif()
endif()
@@ -2483,7 +2609,10 @@ if(WITH_GL_EGL)
if(WITH_SYSTEM_GLES)
if(NOTOPENGLES_EGL_LIBRARY)
message(FATAL_ERROR"Unable to find OpenGL ES libraries. Install them or disable WITH_SYSTEM_GLES.")
@@ -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.
