For a more detailed description of the issue see the commit
message for rB497cd3d7dd6e497be484eb78a8ddb23f53b20343
This change moves fftw to a shared library and reverts the bandaid
we did for 2.83.
Unsure what it is that upsets it so much, but when linking
both sndfile and fftw dynamically, the linker gets confused
and thinks that fftw is importing sf_close from the blender
binary (which makes *NO* sense) leading to a start-up error.
Generating the import library from the .def file using the
ms lib tool creates an import library that works fine.
This is not as much a fix as a work around, but given the real
involves replacing how we build fftw, it is not eligible for 2.83
which is in BCON3 already.
The root of the issue lies with (how we build) fftw3
The first issue is: fftw does not build with MSVC, there are other
dependencies that are not compatible with MSVC and for those we
build the libraries required with mingw64, same for fftw
The second issue is: for reasons unknown we really really really
liked all deps to link statically so wherever possible we did so.
Now during the building of the fftw it linked a few symbols from
libgcc (which we do not ship) like __chkstk_ms, for which we passed
some flags to stop generating calls to it. Problem solved! There
is no way this could possibly turn around and bite us in the rear.
fast forward to today mystery crashes that look like a race condition.
What is happening is, we tell the linker that each thread will require
a 2-megabyte stack, now if every thread immediately allocated 2 megs,
that be 'rough' on the memory usage. So, what happens is (for all apps
not just blender), 2 megs are reserved but not backed by any real memory
and the first page is allocated for use by the stack, now as the stack
grows, it will eventually grow out of that first page, and end up in
an area that has not been allocated yet, to deal with that the allocated
page is followed by a guard page, someone touches the guard page it's
time to grow the stack!
Meanwhile in FFTW is it's doing substantial allocation using alloca
(up to 64 kb) on the stack, jumping over the guard page, and ending
up in reserved but not yet committed memory, causing an access violation.
Now if you think, that doesn't sound right! something should have
protected us from that! You are correct! That thing was __chkstk_ms
which we disabled.
Given we do not want a dependency on libgcc while building with MSVC
the proper solution is to build fftw as a shared library which will
statically link any bits and pieces it needs, however that change
is a little bit too big to be doing in BCON3.
So as a work around, we change the size the stack grows from 8k to
68k which gives fftw a little bit more wiggle room to keep it out
of trouble most of the time.
Note this only sidesteps the issue, this may come up again if the
conditions are just right, and a proper solution will need to be
implemented for 2.90.
Search for all potential library names in each directory, otherwise e.g.
libImath-2_2.a from a system directory will be preferred over libImath.a even
if we specified a directory.
It was disabled in D7520 to keep the pdb's from growing out
of control however the increased link time is just not worth
it.
I'll keep an eye on the dailies and see if we have to come up
with a different solution.
The msvc linker had been warning about libcmt being dragged
in for a bit, finally tracked the issue down to the deps of
jpeg and pthreads which both ignored our cflags.
this diff changes them both to use the dynamic crt rather
than the static one so they'll be in line with all our
other libraries.
We now want to use fairly recent TBB (2018 at least?), so distro a bit
old will not have required package anymore...
Tested with Debian, other distro "should work" (c), but tests there are
much welcome! ;)
Previous code would forec built libs to always be rebuilt when one of
their deps was using a package. While this could be useful when said
package was actually updated, this generates way too much false
positives.
Now only rebuild a built lib when we actually switch from built to
package for one of its deps.
Ever since debug symbols were added for release builds the linker
has been on the chatty side about symbols being missing for our
binary libs.
There's currently no plans to supply those, so best for the linker
not to warn us about them.
Static tbb has always been frowned upon [1] sofar it has worked for us but
given our reliance on tbb is about to increase (D7475), I'd like to move the library
to more supported configuration. Which means moving it to be a dynamic library
The libs part of this change is in rBL62416
Reviewed By: Brecht
Differential Revision: https://developer.blender.org/D7570
This diff add supports for crash logs on windows for
release builds. This can be toggled on/off with the
`WITH_WINDOWS_PDB` cmake option. by default it is on.
Things to take into consideration:
Release builds are hightly optimized and the resulting
backtraces can be wrong/misleading, take the backtrace
as a general area where the problem resides rather than
an exact location.
By default we ship a minimized symbol file that can only
resolve the function names. This was chosen to strike
a balance between growth in size of the download vs
functionality gained. If more detailed information is
required such as source file + line number information
a full pdb can be shipped by setting `WITH_WINDOWS_STRIPPED_PDB`
to off.
Differential Revision: https://developer.blender.org/D7520
Reviewed by: brecht
This diff add supports for crash logs on windows for
release builds. This can be toggled on/off with the
`WITH_WINDOWS_PDB` cmake option. by default it is on.
Things to take into consideration:
Release builds are hightly optimized and the resulting
backtraces can be wrong/misleading, take the backtrace
as a general area where the problem resides rather than
an exact location.
By default we ship a minimized symbol file that can only
resolve the function names. This was chosen to strike
a balance between growth in size of the download vs
functionality gained. If more detailed information is
required such as source file + line number information
a full pdb can be shipped by setting `WITH_WINDOWS_STRIPPED_PDB`
to off.
The Release in the title of this diff refers to the
release build type, not the official blender releases.
Initially this will only be enabled for nightly build
bot versions of blender, official releases as of now
will not ship with symbols.
Differential Revision: https://developer.blender.org/D7520
Reviewed by: brecht
This patch enables TBB as the default task scheduler. TBB stands for Threading Building Blocks and is developed by Intel. The library contains several threading patters. This patch maps blenders BLI_task_* function to their counterpart. After this patch we can add more patterns. A promising one is TBB:graph that can be used for depsgraph, draw manager and compositor.
Performance changes depends on the actual hardware. It was tested on different hardwares from laptops to workstations and we didn't detected any downgrade of the performance.
* Linux Xeon E5-2699 v4 got FPS boost from 12 to 17 using Spring's 04_010_A.anim.blend.
* AMD Ryzen Threadripper 2990WX 32-Core Animation playback goes from 9.5-10.5 FPS to 13.0-14.0 FPS on Agent 327 , 10_03_B.anim.blend.
Reviewed By: brecht, sergey
Differential Revision: https://developer.blender.org/D7475
Usable with the CMake option 'WITH_GHOST_WAYLAND'
The following functionality is working:
- Building with X11 and Wayland at the same time,
wayland is used when available.
- Keyboard, pointer handling.
- Cursor handling.
- Dedicated off-screen windows.
- Drag & drop.
- Copy & paste.
- Pointer grabbing.
See D6567 for further details.
Add root usd library directory to build arguments, same as other libraries.
Also fix error/typo in compile_USD regarding _is_building.
Reviewed By: mont29
Differential Revision: https://developer.blender.org/D7563
sccache [1] is one of the few ccache like solutions that will
work on windows.
sccache support can be enabled with the `WITH_WINDOWS_SCCACHE`
cmake option however it will only will work with ninja as the
build system, msbuild is not supported currently.
Advanced option, developes are expected to obtain and configure
sccache on their own.
```
Full build no cache 1428.90s (100.00%)
Full build cached 434.34s ( 30.40%)
```
[1] https://github.com/mozilla/sccache
Reviewed By: nicholas_rishel, Brecht
Differential Revision: https://developer.blender.org/D7466
Otherwise file browsers do not recognize the Blender executable. This is
already done for official releases.
We leave it off for non-portable builds, since that's how Linux distribution
packages will typically build Blender and we can continue to follow the OS
default there. Using a file browser to launch executables from e.g. /usr/bin
would be rare as wel.
Differential Revision: https://developer.blender.org/D7363
Following up to b555b8d.
Building Blender with hidden symbols but using libraries with visible symbols
was giving linker warnings, specifically for USD. So revert that for now, as
it was not needed for the bugfix.
Hide USD symbols (some of which are not in the USD namespace) to avoid potential
conflicts. May potentially help with AMD OpenCL issues in T74262.
This means symbols from Blender itself and most external libraries. We can't
just hide all because that breaks some libraries. The better solution would
be to rebuild all library dependencies with hidden visibility.
Fixes T75223: Luxrender add-on failing to load on macOS
Happens on macOS 10.15.4 and Xcode 11.4.
The reason of failure is caused by following factors:
- xcodebuild reports full semantic macOS SDK version 10.15.4
- The actual SDK file path will only include major and minor part
of the version (10.15, MacOSX10.15.sdk)
- Previous CMake code of ours expected direct match between SDK
version and file path.
The solution is to make our detection code a bit more flexible and
additionally check for major.minor macOS SDK version in the path.
The specific goal of this change is to get rid of separate code paths
for older and newer Xcode versions.
The version 8.2 is picked since it's the latest version which runs on
macOS 10.11 (which is our current deployment target). If that turns
out too new for some reason the alternative would be to require Xcode
version 5.
The main CMakeLists already requires CMake 3.5, so there is no point of
requiring "newer" CMake on macOS.
This was a code from a while back where CMake 3 was not required on all
platforms.
