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
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
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
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.
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.
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/
Instead of treating Fermi GPU limits as default,
and overriding them for other devices,
we now nicely set them for each platform.
* Due to setting values for all platforms,
we don't have to offset the slot id for OpenCL anymore,
as the image manager wont add float images for OpenCL now.
* Bugfix: TEX_NUM_FLOAT_IMAGES was always 5, even for CPU,
so the code in svm_image.h clamped float textures with alpha on CPU after the 5th slot.
Reviewers: #cycles, brecht
Reviewed By: #cycles, brecht
Subscribers: brecht
Differential Revision: https://developer.blender.org/D1925
Seems particular CUDA implementations has some precision issues,
which made integer coordinate (which was expected to always be
positive) to go negative.
Couple of things:
- No need to use string streams to format the version string,
we can do it at compile time and don't bother with anything
at runtime.
- Function declaration was wring and would have caused linking
conflicts in cases when util_version.h was included from
multiple places.
We should have an utility function to get Cycles version so
applications which are linked to Cycles dynamically can query
the version, but that can't be done as an inlined function in
header and would need to be a function properly exported to a
global symbol table (aka, be implemented in a .cpp file).
Now Cycles has its own versioning, that is mainly interesting for external projects, which integrate the engine.
We start with version 1.7.0. Reasons for that:
* The engine is too mature for a 1.0 release.
* We assume that Cycles inside of Blender 2.61 was version 0.1. We count upwards in 0.1 steps, therefore Cycles inside of Blender 2.77 would be 1.7.
We use a common versioning scheme here, with 3 decimals for the major, minor and patch level.
At the moment cycles --version can be used to display the version, easy to parse for external projects. The info will be added to the UI later aswell.
When the Moon is full it was possible to have a dead-lock in task
scheduler's exit() method.
Similar problem was fixed in Blender's task scheduler 3 years ago
in bae2a2c.
Policy here is a bit more complicated, if tree becomes too deep we're
forced to create a leaf node and size of that leaf wouldn't be so well
predicted, which means it's quite tricky to use single stack array for
that.
Made it more official feature that StackAllocator will fall-back to
heap when running out of stack memory.
It's still much better than always using heap allocator.
We had per-tree statistics already, but it's a bit tricky to see overall
time because trees could be building in parallel.
In fact, we can now print statistics for any TaskPool.
Main use case of this ID will be to emulate TLS which otherwise
would require having some platform-specific implementations which
is not always really optimal.
See notes about the argument in util_task.h.
Currently unused, but will be handy for an upcoming changes.
It'll also be nice to be able to do scoped_lock() for both
Mutex and Spin, but currently it's not really easy to do,
need some changes in typedefs and such, will happen as a
separate commit.
This seems quite useful for the development, so you don't need to wait
all the kernels to be re-compiled when working on a new feature, which
speeds up re-iteration.
Marked as an advanced option, so if it doesn't work so well in practice
it's safe to revert anyway.
This is a mix of regression and old unsupported configuration.
Regression was caused by some checks added on Blender side which was
checking whether python function returned error or not. This made it
impossible to enable Cycles when running from a file path which can't
be encoded with MBCS codepage.
Non-regression issue was that it wasn't possible to use pre-compiled
CUDA kernels when running from a path with non-ascii multi-byte
characters.
This commit fixes regression and CUDA parts, but OSL still can't be
used from a non-ascii location because it uses non-widechar API to
work with file paths by the looks of it. Not sure we can solve this
just from our side by using some codepage trick (UTF-16?) since even
oslc fails to compile shader when there are non-ascii characters in
the path.
It's not really handy to silence something unused hoping for it'll be
used in the future. We can end up with quite some silencing then.
Also made this flag which i find rather useless to NOT cause -Werror
in Cycles code.
The issue was caused by static vectors allocating some internal
data using rebound element allocator for them, which was causing
access to a non-initialized statistics objects and was failing a
lot when switching Blender to a fully guarded allocation.
Additionally, we were not able to free that internal memory before
Blender exits, which was causing false-positive memory leak prints.
Now we're not using GuardedAllocator for those proxy containers.
Ideally this should be done as a GuardedAllocator::rebind, but
it didn't work for vector<bool> because it seems some internal
parts are converting bool to char32_t, which either makes it so
we can't use GuardedAllocator for those vectors or the compiler
get's confused when we're trying explicitly allow GuardedAllocator
for rebind<char32_t>.
This with current approach we should be fine for the release.
Was caused by some safety things of making sure we've for NULL
terminator for the buffer when doing mbs<->wcs conversion, but
it turns out this simply confuses str::string and it can no
longer have proper .size(). Let's assume behavior of string
allocation is same all over the std, and we can avoid having
that extra null-terminator allocated.
C++ requires specific alignment of the allocations which was not an
issue when using GCC but uncovered issue when using Clang on OSX.
Perhaps some versions of Clang might show errors on other platforms
as well.
