It appears that previously, assembly was straight up disabled for all ARM64 platforms in x264 - this re-enables it.
This also updates the version of FFmpeg to 7.1.1. I updated the patch files so that they would cleanly apply to 7.1.1.
For Windows ARM64, it also switches both of these libraries to use the copy of LLVM we build as part of the deps, instead of MSVC - this gives a small performance increase. To do this for x264, I added a small patch taken from VCPKG - this should be harmless to other platforms.
With all these changes, I was able to get a ~20-30% perf improvement in video transcoding in the sequence editor.
Pull Request: https://projects.blender.org/blender/blender/pulls/137670
`IMB_load_image_from_file_descriptor` use mmap to read images, and mmap
sometimes can have erroreous state llike accessing a file from network
drives. This patch tries to handle these exceptions like `BLI_mmap_read`
on Win32 (On unix the `sigbus_handler` will be called instead).
Pull Request: https://projects.blender.org/blender/blender/pulls/139472
During rendering when main thread is blocked or all screens are minimized the
garbage collection will not happen resulting in crashes as resources are not freed.
A better solution would be to do garbage collection in a separate thread but that requires a
ref counting system. The specifics of such a system is still unclear.
A possible solution for a Vulkan specific ref counting is to store the ref counts in the
resource tracker. That would only handle images and buffers, but it would solve the most
resource hungry issues.
Pull Request: https://projects.blender.org/blender/blender/pulls/139475
This patch uses 2D and 4D sockets for relevant node sockets in the
compositor. This only on the UI side, internally, some of the sockets
still store 3D data, which will be handled in a later patch in the 5.0
release.
Pull Request: https://projects.blender.org/blender/blender/pulls/139452
Meshes with millions of vertices could freeze when storing
edit-mode undo steps. This occurred when de-duplicating boolean arrays
(typically selection & hidden data).
This problem occurred after v3.3 because vert/edge/face flags were
moved into their own boolean custom-data layers.
Resolve using run-length encoding, used to encode all boolean
custom-data layers before adding them to the BArrayStore.
The overall speedup depends on the order of elements in the mesh
as well as the previous state of the mesh used for comparison.
Testing box selection with subdivided cubes (3.9 million vertices)
shows a speedup from an around between 8..16 seconds to ~0.025 seconds.
Even though this ran in a background thread, adding undo steps
waits for the previous step to finish which could freeze.
Tweaks to BLI_array_store such as increasing the block-size used by
edit-mesh undo and increasing length of hashed data helped but didn't
resolve the issue, see: !138975.
See code-comments for details.
Add RLE encoding/decoding functions for byte arrays for the purpose
of pre-processing arrays with large spans of (mostly) uniform values
before storing them in a BArrayState.
Part of a fix for #136737.
Previously, it was possible for the number of stored columns per table to grow
unbounded. While this likely isn't a problem in practice in most cases, one can
imagine cases where people temporarily have thousands of attributes when are
then never used again. We shouldn't have to store any data for these columns
forever.
This patch adds some simple garbage collection mechanism that keeps the number
of stored unavailable columns per spreadsheet table below a certain threshold
(50 currently). Least recently used columns are removed first.
Pull Request: https://projects.blender.org/blender/blender/pulls/139469
It was intended that pointers instancers would be skipped if they were
marked as invisible. However, we didn't account for the case of an
instancer using the "inherits" visibility but one of its ancestors in
the hierarchy being invisible.
Now these instancers will also be excluded. This is done by checking for
purpose and visibility earlier and halting the recursive traversal as
soon as we find a prim which doesn't meet the criteria.
Pull Request: https://projects.blender.org/blender/blender/pulls/139241
Videos that are higher bit depth than 8 bit are decoded into floating
point formats (so anything that uses 10/12 bits or HDR). Speedup:
- Multi-thread GBRA planar -> RGBA interleaved conversion.
- If video is rotated, do the rotation step while doing the above
conversion directly, instead of a separate rotation afterwards. Cuts
down on memory bandwidth usage.
Playing back two 1920x1080 video tracks, where both are rotated and
H.265 10 bit HDR, on my PC (Ryzen 5950X), average playback frame time
goes 47ms -> 32ms (so 21FPS -> 31FPS)
Pull Request: https://projects.blender.org/blender/blender/pulls/139439
Previously, columns were removed from a table in the spreadsheet if they are not
available anymore. This also meant that their position and width was lost. When
the same column became available again, it was inserted at the end.
This patch makes it so that each table also remembers the columns that are not
available currently (and flags them accordingly). This way, the position and
width can be restored once the data becomes available again.
Pull Request: https://projects.blender.org/blender/blender/pulls/139440
On selected platforms there were some validation errors. It was caused by
platforms that returned a different number of swapchain images then were
requested. In that case the semaphores can get out of sync.
Current mechanism isn't future proof as the max number of images are
statically defined.
For this change the present semaphores is also separated from the frames
to better support out of order swapchain images.
Pull Request: https://projects.blender.org/blender/blender/pulls/139446
This patch turns the Blur node options to inputs.
Size is now a 2D vector and replaces the Size X and Y option. Bokeh was
renamed to Separable to reflect its actual function. Relative was
removed in favor of the newly added Relative To Pixel node workflow.
There is a slight difference in variable size blurring due to float vs
integer computations, so two tests were updated.
Reference #137223.
Pull Request: https://projects.blender.org/blender/blender/pulls/139329
This removes redundant labels from various input nodes like the Value, Integer
and Object node.
Design wise, this is mostly straight forward except for two aspects:
* Some input nodes some have a gizmo icon. In this case I just added the gizmo
icon on the same row.
* The checkbox in the Boolean input node should probably still have a label, so
I kept that.
Implementation wise this adds a new function to socket declarations that allows
us to override the draw behavior of individual sockets per node.
Pull Request: https://projects.blender.org/blender/blender/pulls/139432
GHOST backend didn't use logging. This PR adds an initial ghost.vulkan
logging and improves the reporting of logging in vulkan.
logging can be enabled by `blender --log "gpu.vulkan,ghost.vulkan" --log-level 2`
it shows the optional extensions that are enabled and information about swap chain
events.
Pull Request: https://projects.blender.org/blender/blender/pulls/139437
Looks like newer versions of c++ doesn't want to implicitely do type
promotions from int to float. The patch fixes them.
Also fixed a typo `outter` -> `outer`.
Pull Request: https://projects.blender.org/blender/blender/pulls/139435
This patch adds support for 2D and 4D vector sockets. The default value
structure for vectors was extended to include a new dimensions input,
which can be 2, 3, or 4. The default value was also extended to be a
float4, but only some of its components might be used depending on the
dimensions members.
Each vector subtype now has three variants ending with 2D or 4D as a
prefix depending on its dimensions, and the 2D/4D prefix was taken into
account for the socket type RNA enum functions.
All node systems currently always treat the vectors as 3D, but support
for it in the compositor will shortly follow in another patch.
Depends on #138805.
Pull Request: https://projects.blender.org/blender/blender/pulls/138824
This is adding a hard error for backends do that do not supports it.
Even without this, the backend would return a nullptr. But the error
message in this case was just "see console for more details" without
any additional detail. Which was confusing.
Unlike OpenGL and Metal, this handle is not shared, but rather Cycles
has to take ownership of it. This required a fair amount of refactoring
to ensure the handle is closed, ownership is properly transferred, and
the handle is recreated once when the pixel buffer is modified.
The cryptomatte textures where optimized for samplers, but could trigger
UB as they are also used with load/store. This PR will change the
dynamic data type of the cryptomatte to always use RGBA32F.
Detected with Vulkan validation layers.
Pull Request: https://projects.blender.org/blender/blender/pulls/139427
The `immVertex2i` calls in `drawWalkPixel` will now cause an assert in
`immAttr2i` after 617858e453 because the component types are more
strictly enforced. Now switch to using floats to avoid asserts and
potentially drawing garbage in the viewport.
Pull Request: https://projects.blender.org/blender/blender/pulls/139429
This patch adds a new Dimensions output to the Image Info node. Contrary
to the Resolution output, the Dimensions output takes transformations
into account. So an image of resolution 512x512 that is scaled up by 2
would have a Resolution of 512x512 but a Dimensions of 1024x1024.
Depends on #138935.
Pull Request: https://projects.blender.org/blender/blender/pulls/139211
This is a follow-up to 9ce0a2d1d5 and uses the same logic. It allows
translation of node panel labels inside the material properties, using
their specified translation context. Not in use currently, but may be
needed in the future.
Pull Request: https://projects.blender.org/blender/blender/pulls/139367
This patch adds a new Image Coordinates node for the compositor. The
Texture Coordinates and Pixel Coordinates outputs were removed from the
Image Info node and were added to the new node instead. Additionally, a
new Normalized Coordinates output was added.
The Pixel Coordinates output now no longer includes half pixel offsets.
Pull Request: https://projects.blender.org/blender/blender/pulls/138935
Usage of conditional to fix threading and performance delays when
waiting for the submission fence to become valid. The previous
(faulty) implementation didn't work well on WoA devices.
Previously, the node group operator only had fairly basic flat drawing for the
inputs. Due to previous refactors, it's now possible to reuse the drawing code
of the Geometry Nodes modifier. That way the redo panel now has all the features
that also exist in the modifier. Also, future improvements will benefit both
systems and potentially more in the future.
Pull Request: https://projects.blender.org/blender/blender/pulls/139389
With #128051, the default timeline tool changed to
a `sequencer.select` on press with no associated
properties (before, it had `deselect_all`).
This meant that retiming key selection broke (as
reported in #129892) and strips would only deselect
on release if clicking into empty space (due to a separate
keymap item in the global sequencer keymap).
This patch reverts 97e44901b4 while keeping its fix,
in favor of a more robust solution (just add `deselect_all`
to the keymap item properties in the box select tool).
This simplifies extending `blender::get_default_hash` to more than 4 parameters
by just adding extra factors to the list. The behavior should be the same as
before.
Pull Request: https://projects.blender.org/blender/blender/pulls/139392
Readfile code would not always properly set the main file version number
in 'split' Mains created for a library, and would also not detect
'future versions' blendfiles.
This info is currently not used by linking code, so this should be a
no-op change in current main.
The goal is to simplify reusing the same code to draw the inputs of the node
tool operator.
This patch simply extracted all the modifier-specific code into callbacks which
are passed in from a higher level. The operator drawing code would basically
just need to provide different callback functions. Although some more additional
cleanups may become necessary to make that fully work.
Pull Request: https://projects.blender.org/blender/blender/pulls/139384
This function has not python equivalent, using the
returned pointer to write properties seems enough
as equivalent as how is done in python.
Also, this removes the unused `uiItemFloatO` API.
Pull Request: https://projects.blender.org/blender/blender/pulls/139355
The goal is to be able to reuse parts of this code in all the places that call
Geometry Nodes. Currently, that's only a modifier and an operator. More places
may follow in the future though.
This patch does not implement any of the code-reuse yet, it just extracts the
code into a separate file. Follow-up patches will work towards better
code-reuse.
Note, this extracts the entire modifier drawing, including error messages,
baking etc. I found this to be easier, because they still often share a common
core (like getting the logged data).
Pull Request: https://projects.blender.org/blender/blender/pulls/139379
Nowadays, Geometry Nodes is not only evaluated by a modifier, but also by an
operator and in the future potentially by brushes. Therefore, the old name was
misleading because it sounded like it was specific to the modifier.
Pull Request: https://projects.blender.org/blender/blender/pulls/139378
This is used as base class for the compute contexts for group and evaluate
closure nodes. Furthermore, in the future this can be used for the compute
context that is passed into field evaluation.
Pull Request: https://projects.blender.org/blender/blender/pulls/139377
Qualcomm devices don't support external memory, but there an external
memory pool was still being constructed. This PR skips the creation and
asserts when using external memory on those devices.
Pull Request: https://projects.blender.org/blender/blender/pulls/139326
