Age | Commit message (Collapse) | Author |
|
- Re-arrange locks, so no actual memory allocation
(which is relatively slow) happens from inside
the lock. operation system will take care of locks
which might be needed there on it's own.
- Use spin lock instead of mutex, since it's just
list operations happens from inside lock, no need
in mutex here.
- Use atomic operations for memory in use and total
used blocks counters.
This makes guarded allocator almost the same speed
as non-guarded one in files from Tube project.
There're still MemHead/MemTail overhead which might
be bad for CPU cache utilization
|
|
There's one thing we didn't foresee from the beginning,
which is apparently TLS is only available in OSX starting
from version 10.7, and we still do support of 10.6.
After recent Brecht's changes about locked viewport
while initializing BI render this TLS is not needed
in trunk anymore. So reverting this chunk of base
iteration to use static variable. But leaving all the
other static variables warped into context still, it
should help a bit in the future refactor.
Real fix would be to have some kind of graph context
evaluation structure which would be passing to update
routines (which will solve threaded mballs update) and
making depsgraph responsible for getting a motherball.
But this is all for GSoC project.
|
|
This time issue was caused by static variables used in
BKE_scene_base_iter_next function.
Change is not so much ultimate actually, but didn't
find more clear solution for now. So the changes are:
- Wrap almost all the static variables into own context-
like structure, which is owned by the callee function
and getting passed to the iteration function.
- Recursion detection wasn't possible with such approach,
so recursion detection still uses static in_next_object
variable, but which is now stored in thread local
storage (TLS, or thread variable if this names are more
clear for you).
This makes code thread-safe, but for sure final solution
shall be completely different. Ideally, dependency graph
shall be possible to answer on question "which object is
a motherball for this metaball". This will avoid iterating
via all the bases, objects and duplis just to get needed
motherball.
Further, metaball evaluation ideally will use the same
kind of depsgraph filtering, which will get result for
question like "which objects belongs to this group of
metaballs".
But this ideal things are to be solved in Joshua's and
mind GSoC projects.
Tested on linux (gcc and clang) and windows (msvc2008),
hopefully no compilation error will happen.
Thanks to Brecht for reviewing the change and getting
feedback for other possible ways we've dicussed!
|
|
data.
Now the viewport rendering thread will lock the main thread while it is exporting
objects to render data. This is not ideal if you have big scenes that might block
the UI, but Cycles does the same, and it's fairly quick because the same evaluated
mesh can be used as for viewport drawing. It's the only way to get things stable
until the thread safe dependency graph is here.
This adds a mechanism to the job system for jobs to lock the main thread, using a
new 'ticket mutex lock' which is a mutex lock that gives priority to the first
thread that tries to lock the mutex.
Still to solve: undo/redo crashes.
|
|
Now it works for blender internal, cycles and other multithreading code in
Blender in both background and UI mode.
|
|
Added a mutex lock for smoke data access. The render was already working with a
copy of the volume data, so it's just a short lock to copy things and should not
block the UI much.
|
|
This commit makes BKE_image_acquire_ibuf referencing result, which means once
some area requested for image buffer, it'll be guaranteed this buffer wouldn't
be freed by image signal.
To de-reference buffer BKE_image_release_ibuf should now always be used.
To make referencing working correct we can not rely on result of
image_get_ibuf_threadsafe called outside from thread lock. This is so because
we need to guarantee getting image buffer from list of loaded buffers and it's
referencing happens atomic. Without lock here it is possible that between call
of image_get_ibuf_threadsafe and referencing the buffer IMA_SIGNAL_FREE would
be called. Image signal handling too is blocking now to prevent such a
situation.
Threads are locking by spinlock, which are faster than mutexes. There were some
slowdown reports in the past about render slowdown when using OSX on Xeon CPU.
It shouldn't happen with spin locks, but more tests on different hardware would
be really welcome. So far can not see speed regressions on own computers.
This commit also removes BKE_image_get_ibuf, because it was not so intuitive
when get_ibuf and acquire_ibuf should be used.
Thanks to Ton and Brecht for discussion/review :)
|
|
Replace old color pipeline which was supporting linear/sRGB color spaces
only with OpenColorIO-based pipeline.
This introduces two configurable color spaces:
- Input color space for images and movie clips. This space is used to convert
images/movies from color space in which file is saved to Blender's linear
space (for float images, byte images are not internally converted, only input
space is stored for such images and used later).
This setting could be found in image/clip data block settings.
- Display color space which defines space in which particular display is working.
This settings could be found in scene's Color Management panel.
When render result is being displayed on the screen, apart from converting image
to display space, some additional conversions could happen.
This conversions are:
- View, which defines tone curve applying before display transformation.
These are different ways to view the image on the same display device.
For example it could be used to emulate film view on sRGB display.
- Exposure affects on image exposure before tone map is applied.
- Gamma is post-display gamma correction, could be used to match particular
display gamma.
- RGB curves are user-defined curves which are applying before display
transformation, could be used for different purposes.
All this settings by default are only applying on render result and does not
affect on other images. If some particular image needs to be affected by this
transformation, "View as Render" setting of image data block should be set to
truth. Movie clips are always affected by all display transformations.
This commit also introduces configurable color space in which sequencer is
working. This setting could be found in scene's Color Management panel and
it should be used if such stuff as grading needs to be done in color space
different from sRGB (i.e. when Film view on sRGB display is use, using VD16
space as sequencer's internal space would make grading working in space
which is close to the space using for display).
Some technical notes:
- Image buffer's float buffer is now always in linear space, even if it was
created from 16bit byte images.
- Space of byte buffer is stored in image buffer's rect_colorspace property.
- Profile of image buffer was removed since it's not longer meaningful.
- OpenGL and GLSL is supposed to always work in sRGB space. It is possible
to support other spaces, but it's quite large project which isn't so
much important.
- Legacy Color Management option disabled is emulated by using None display.
It could have some regressions, but there's no clear way to avoid them.
- If OpenColorIO is disabled on build time, it should make blender behaving
in the same way as previous release with color management enabled.
More details could be found at this page (more details would be added soon):
http://wiki.blender.org/index.php/Dev:Ref/Release_Notes/2.64/Color_Management
--
Thanks to Xavier Thomas, Lukas Toene for initial work on OpenColorIO
integration and to Brecht van Lommel for some further development and code/
usecase review!
|
|
Issue was caused by threading conflict between compositor output node which
is freeing buffers used by render result image and image draw code which
could use buffers at the same time as compositor frees this buffers.
Solved by adding adding lock around viewer image invalidation and image
drawing.
Use renamed LOCK_PREVIEW mutex for this, which si not called LOCK_DRAW_IMAGE.
With new compositor locking for preview is not needed so it could be removed.
Added the same lock around viewer operation which also frees buffers used
by viewer image. It's actually quite difficult to check whether this is
indeed needed. This code seems to be using acquire/release technique, but
somehow acquiring ImBuf before invalidating it in compositor operation
doesn't resolve the issue, so probably it's not actually locking acquire
and things should be checked deeper.
|
|
|
|
waiting in scheduling and worker threads instead of continuous loops with sleep times. This should help reduce unnecessary wait times in Tile.
|
|
|
|
|
|
merge it didn't need to be) - now rendering uses its better if its threadsafe.
|
|
without this each face would get a solid color, this is the same method used in object mode.
also copy BLI_array.h fix from bmesh branch.
|
|
without the underscores these clogged up the namespace for autocompleation which was annoying.
|
|
|
|
===========================
Commiting camera tracking integration gsoc project into trunk.
This commit includes:
- Bundled version of libmv library (with some changes against official repo,
re-sync with libmv repo a bit later)
- New datatype ID called MovieClip which is optimized to work with movie
clips (both of movie files and image sequences) and doing camera/motion
tracking operations.
- New editor called Clip Editor which is currently used for motion/tracking
stuff only, but which can be easily extended to work with masks too.
This editor supports:
* Loading movie files/image sequences
* Build proxies with different size for loaded movie clip, also supports
building undistorted proxies to increase speed of playback in
undistorted mode.
* Manual lens distortion mode calibration using grid and grease pencil
* Supervised 2D tracking using two different algorithms KLT and SAD.
* Basic algorithm for feature detection
* Camera motion solving. scene orientation
- New constraints to "link" scene objects with solved motions from clip:
* Follow Track (make object follow 2D motion of track with given name
or parent object to reconstructed 3D position of track)
* Camera Solver to make camera moving in the same way as reconstructed camera
This commit NOT includes changes from tomato branch:
- New nodes (they'll be commited as separated patch)
- Automatic image offset guessing for image input node and image editor
(need to do more tests and gather more feedback)
- Code cleanup in libmv-capi. It's not so critical cleanup, just increasing
readability and understanadability of code. Better to make this chaneg when
Keir will finish his current patch.
More details about this project can be found on this page:
http://wiki.blender.org/index.php/User:Nazg-gul/GSoC-2011
Further development of small features would be done in trunk, bigger/experimental
features would first be implemented in tomato branch.
|
|
|
|
lazely created but this wasn't done in a thread safe way.
|
|
|
|
http://markmail.org/message/fp7ozcywxum3ar7n
|
|
* Some changes and cleanup pointed on the codereview.
|
|
|
|
- fixed bug in paste material, exposed by stricter warnings.
- removed/renamed various shadowed vars.
- removed BGE lamp.colour, only allow lamp.color attribute.
|
|
how removing opengl textures from outside main thread is done so it
happens as part of the main loop.
|
|
* Division by zero fix for TNT SVD code.
* Sound fix, in case ffmpeg decode fails, don't use the samples.
* Fix for incorrect bounds of transformed objects in new raytracing code.
* Gave memory arena's a name used for allocations for easier memory
usage debugging.
* Dupligroup no_draw option was using layers but not restrict view/render
setting. (not a bugfix exactly but would do display list context switching
while drawing for no reason).
* Fix objects instanced on hair particles not giving consistent results
when the object is transformed.
* New math functions: madd_v4_v4fl, len_squared_v3v3, interp_v4_v4v4v4,
mul_v4_m4v4, SH and form factor functions, box_minmax_bounds_m4.
* mul_m4_m4m4 and mul_m3_m3m3 now accept the same pointers for multiple
arguments.
* endjob callback for WM jobs system.
* Geometry node uv/color layer now has search list/autocomplete.
* Various small buildsystem tweaks, not strictly needed yet in trunk.
|
|
|
|
instead they are queued to be freed the next time GPU_image_free() is run from the main thread.
|
|
|
|
* Viewer node could free image while it is being redrawn, viewer image
buffers now need acquire/release to be accessed as was already the
case for render results.
* The Composite node could free the image buffers outside of a lock,
also causing simultaneous redraw to crash.
* Especially on Windows, re-rendering could crash when drawing an image
that was freed. When RE_RenderInProgress was true it would access the
image buffer and simply return it while it could still contain a pointer
to a render result buffer that was already freed. I don't understand
why this case was there in the first place, so I've removed it.
Possibly fixes bugs #20174, #21418, #21391, #21394.
|
|
|
|
- reload fix didnt work last commit.
|
|
to wait for an item to be put in the queue and then pop immediately without,
this makes it possible to avoid sleep() while waiting for the results of a
thread.
|
|
* Rendering twice or more could crash layer/pass buttons.
* Compositing would crash while drawing the image.
* Rendering animations could also crash drawing the image.
* Compositing could crash
* Starting to rendering while preview render / compo was
still running could crash.
* Exiting while rendering an animation would not abort the
renderer properly, making Blender seemingly freeze.
* Fixes theoretically possible issue with setting malloc
lock with nested threads.
* Drawing previews inside nodes could crash when those nodes
were being rendered at the same time.
There's more crashes, manipulating the scene data or undo can
still crash, this commit only focuses on making sure the image
buffer and render result access is thread safe.
Implementation:
* Rather than assuming the render result does not get freed
during render, which seems to be quite difficult to do given
that e.g. the compositor is allowed to change the size of
the buffer or output different passes, the render result is
now protected with a read/write mutex.
* The read/write mutex allows multiple readers (and pixel
writers) at the same time, but only allows one writer to
manipulate the data structure.
* Added BKE_image_acquire_ibuf/BKE_image_release_ibuf to access
images being rendered, cases where this is not needed (most
code) can still use BKE_image_get_ibuf.
* The job manager now allows only one rendering job at the same
time, rather than the G.rendering check which was not reliable.
|
|
Preview renders:
- Added proper button type (BUT_EXTRA) for preview buttons, to
handle drawing better. It now first draws an alpha mask, to
ensure the preview is correctly fitting inside the widget style.
It then draws the outline.
- Added protection for executing preview renders while regular
rendering, that's not going to work...
|
|
Will have to do a second pass tomorrow to fix some leftovers.
|
|
New functions to easily dispatch work to a limited number of thread, transparently.
NOTE: Could be merged in trunk, if needed.
|
|
This can be safely merged in trunk, in case anyone needs something like that.
|
|
whatever the system has, useful in the studio with 2,4,8 core systems when sharing files.
|
|
by yesterdays commit.
Now a designater LOCK_IMAGE is used for all image write/read.
|
|
Removed all limitations from render code for maximum threads. The only
define for this now is in BLI_threads.h, and currently set to 8.
Note that each thread renders an entire tile, and also allocates the
buffers for the tiles, so; more threads might work better with smaller
tiles.
IMPORTANT: node system won't work yet with more than 2 threads! So, don't
try material nodes or compositing with over 2 threads. That I'll commit
later today.
What does work (should work :) is AO and soft shadow now.
|
|
module itself, replacing the special MEM_mallocT/MEM_freeT functions.
Mutex locking is only enabled when threads are running.
There was no good reason to have these separate, it just led to ugly
hacks when calling functions with non-threadsafe malloc from threads.
|
|
- Renderwin still used a thread-unsafe malloc, in the header text print
- Setting clipping flags in vertices for parts required a mutex lock after
all... I thought it would go fine, but noticed on renders with small
amounts of faces that sometimes faces disappear from a render.
(was doing movie credits, so all faces are visible! Otherwise it would
have hardly been noticable...)
|
|
For some reason I thought SDL thread handling would be much simpler... but
the migration to posix pthread went very smooth and painless. Less code
even, and I even notice a slight performance increase!
All threading code is still wrapped in blenlib/intern/threads.c
Only real change was making the callback functions to return void pointer,
instead of an int.
The mutex handling is also different... there's no test anymore if a
mutex was initialized, which is a bit confusing. But it appears to run
all fine still. :)
Nathan Letwory has been signalled already to provide the Windows pthread
library and make/scons linking. For MSVC we might need help from someone
else later though.
|
|
In Orange we've been fighting the past weeks with memory usage a lot...
at the moment incredible huge scenes are being rendered, with multiple
layers and all compositing, stressing limits of memory a lot.
I had hoped that less frequently used blocks would be swapped away
nicely, so fragmented memory could survive. Unfortunately (in OSX) the
malloc range is limited to 2 GB only (upped half of address space).
Other OS's have a limit too, but typically larger afaik.
Now here's mmap to the rescue! It has a very nice feature to map to
a virtual (non existing) file, allowing to allocate disk-mapped memory
on the fly. For as long there's real memory it works nearly as fast as
a regular malloc, and when you go to the swap boundary, it knows nicely
what to swap first.
The upcoming commit will use mmap for all large memory blocks, like
the composit stack, render layers, lamp buffers and images. Tested here
on my 1 GB system, and compositing huge images with a total of 2.5 gig
still works acceptable here. :)
http://www.blender.org/bf/memory.jpg
This is a silly composit test, using 64 MB images with a load of nodes.
Check the header print... the (2323.33M) is the mmap disk-cache in use.
BTW: note that is still limited to the virtual address space of 4 GB.
The new call is:
MEM_mapalloc()
Per definition, mmap() returns zero'ed memory, so a calloc isn't required.
For Windows there's no mmap() available, but I'm pretty sure there's an
equivalent. Windows gurus here are invited to insert that here in code! At
the moment it's nicely ifdeffed, so for Windows the mmap defaults to a
regular alloc.
|
|
I noticed still several cases where the Imbuf library was called within a
thread... and that whilst the Imbuf itself isn't threadsafe. Also the
thread lock I added in rendering for loading images actually didn't
work, because then it was still possible both threads were accessing the
MEM_malloc function at same time.
This commit nearly fully replaces ImBuf calls in compositor (giving another
nice speedup btw, the way preview images in Nodes were calculated used
clumsy imbuf scaling code).
I've also centralized the 'mutex' locking for threading, which now only
resides in BLI_threads.h. This is used to secure the last ImBuf calls
I cannot replace, which is loading images and creating mipmaps.
Really hope we get something more stable now!
|
|
for compositing code.
Officially malloc/calloc/free is threadsafe, but our secure malloc system
requires all memory blocks to be stored in a single list, so when two
threads write in this list you get conflicts.
|
|
- Compositor now is threaded
Enable it with the Scene buttons "Threads". This will handle over nodes to
individual threads to be calculated. However, if nodes depend on others
they have to wait. The current system only threads per entire node, not for
calculating results in parts.
I've reshuffled the node execution code to evaluate 'changed' events, and
prepare the entire tree to become simply parsed for open jobs with a call
to node = getExecutableNode()
By default, even without 'thread' option active, all node execution is
done within a separate thread.
Also fixed issues in yesterdays commit for 'event based' calculations, it
didn't do animated images, or execute (on rendering) the correct nodes
when you don't have Render-Result nodes included.
- Added generic Thread support in blenlib/ module
The renderer and the node system now both use same code for controlling the
threads. This has been moved to a new C file in blenlib/intern/threads.c.
Check this c file for an extensive doc and example how to use it.
The current implementation for Compositing allows unlimited amount of
threads. For rendering it is still tied to two threads, although it is
pretty easy to extend to 4 already. People with giant amounts of cpus can
poke me once for tests. :)
- Bugfix in creating group nodes
Group node definitions demand a clear separation of 'internal sockets' and
'external sockets'. The first are sockets being linked internally, the latter
are sockets exposed as sockets for the group itself.
When sockets were linked both internal and external, Blender crashed. It is
solved now by removing the external link(s).
|