Use more direct access to next/previous vertices.
- `BM_edge_other_vert(l_curr->e, l_curr->v)` -> `l_curr->next->v`.
- `BM_edge_other_vert(l_curr->prev->e, l_curr->v)` -> `l_curr->prev->v`.
Add asserts to keep the intention clear.
Using BMesh operators through the edit-mesh API created a full copy
of the mesh so it was possible to restore the mesh in case
one of the operators raised an error.
Remove support for automatic backup/restore from the EDBM_op_* API's
as it adds significant overhead and was rarely used.
Operators that need this can use the BMBackup API to backup & restore
the mesh in case of failure.
Add warning levels to BMO_error_raise so operators can report problems
without it being interpreted as a request to cancel the operation.
For high-poly meshes creating and freeing a full copy is an expensive
operation, removing this gives a speedup of ~1.77x for most operators
except for "connect_verts" / "connect_vert_pair"
which still uses this functionality.
Previously, any face groups that could not be merged into a face
caused the entire operation to report an error and do nothing.
Now these cases are skipped over, dissolving faces where possible.
- Only create arrays with groups of two or more faces.
- Remove raising exception for zero length arrays.
- Remove redundant exception check (assert there is no exception).
- Use a struct for face array & it's length instead of a NULL
terminated array (removes the need to count faces in a loop).
The error codes could be used to look up messages from a table
of messages however this wasn't especially useful.
Now all calls to BMO_error_raise must inclue a message.
Normals now includes many functions including normal splitting &
custom normal manipulation split this into it's own file
to centralize related functions.
Skip updating normals & tessellation for contiguous geometry regions
for operations such as translate & uniform scale.
This means when all geometry is selected, no updates are needed
as the relative locations of vertices aren't being modified.
Performance:
As this is skipping a multi-threaded operation,
larger improvements are noticeable on systems with fewer cores.
- ~1.15x to ~1.3x overall gain for 32 cores.
- ~1.7x to ~2.2x overall gain for 1 core (limited using `-t 1` argument).
Details:
- Rotate & non-uniform scale only skip tessellation.
- Proportional editing and axis-mirror have special handling
ensure geometry is properly grouped before considering
a face part of a single group that can be skipped.
- Loose vertices always need their normals to be recalculated
since they're calculated based on the location.
- Non-affine transform operations such as shrink-fatten & bend,
don't take advantage of this optimization.
- Snap projection also disables the optimization.
This patch fixes many minor spelling mistakes, all in comments or
console output. Mostly contractions like can't, won't, don't, its/it's,
etc.
Differential Revision: https://developer.blender.org/D11663
Reviewed by Harley Acheson
This patch uses threading to flush selection from verts to edges and
from edges to faces. The solution is lockless and should scale well on
modern CPU architectures.
Master:{F10185359}
Patch:{F10185361}
End user performance went from 3.9 to 4.6 FPS (Stanford Dragon) but that
was measured on a Intel Core i7-6700 CPU and AMD RX480 Gpu. The more
cores the better improvements you get.
Reviewed By: mano-wii
Differential Revision: https://developer.blender.org/D11644
During selections the total selection is refreshed at the end. This
process was done single threaded. This patch will do a parallel iter
approach.
Master: 0.043612s Threaded 0.017964s.
Master: {F10179586}
This patch: {F10179587}
Reviewed By: mano-wii
Differential Revision: https://developer.blender.org/D11622
This patch ensures that selection mode flushing updates total selection
counts internally. This reduces recounting when we are sure that the
input total selection counts were up to date.
For example for circle selection the total selection counts were
correct. But during flushing the selection could have been changed and
therefore the selection was always recounted.
This increased the performance on selected system from 6.90 FPS to 8.25
FPS during circle selection operations.
Before: {F10179981}
After: {F10179982}
Reviewed By: mano-wii
Differential Revision: https://developer.blender.org/D11647
This matches BMesh which also has tessellation in it's own file.
Using a separate file helps with organization when
extracting code into smaller functions.
Support calculating face normals when tessellating. When this is done
before updating vertex normals it gives ~20% performance improvement.
Now vertex normal calculation only needs to perform a single pass on the
mesh vertices when called after tessellation.
Extended versions of normal & looptri update functions have been added:
- BM_mesh_calc_tessellation_ex
- BM_mesh_normals_update_ex
Most callers don't need to be aware of this detail by using:
- BKE_editmesh_looptri_and_normals_calc
- BKE_editmesh_looptri_and_normals_calc_with_partial
- EDBM_update also takes advantage of this,
where calling EDBM_update with calc_looptri & calc_normals
enabled uses the faster normal updating logic.
Bypass stored edge-vectors for ~16% performance gains.
While this increases unit-length edge-vector calculations by around ~4x
the overhead of a parallel loop over all edges makes it worthwhile.
Note that caching edge-vectors per-vertex performs better and may be
worth investigating further, although in my tests this increases code
complexity with barley measurable benefits over not using cache at all.
Details about performance and possible optimizations are noted in
bm_vert_calc_normals_impl.
Use BM_iter_parallel for face tessellation, this gives around
6.5x speedup for BM_mesh_calc_tessellation on high poly meshes in my
tests, although exact speedup depends on available cores.
Support thread local storage for BLI_task_parallel_mempool,
as well as support for the reduce and free callbacks.
mempool_iter_threadsafe_* functions have been moved into a private
header thats only shared between task_iterator.c and BLI_mempool.c
so the TLS can be made part of the iterator array without having to
rely on passing in struct offsets.
Add test task.MempoolIterTLS that ensures reduce and free
are working as expected.
Reviewed By: mont29
Ref D11548
Simplify vertex normal calculation by moving the main normal
accumulation function to operate on vertices instead of faces.
Using faces had the down side that it needed to zero, accumulate and
normalize the vertex normals in 3 separate passes, accumulating also
needed a spin-lock for thread since the face would write it's normal
to all of it's vertices which could be shared with other faces.
Now a single loop over vertices is performed without locking.
This gives 5-6% speedup calculating all normals.
This also simplifies partial updates, fixing a problem where
all connected faces were being read from when calculating normals.
While this could have been resolved separately,
it's simpler to operate on vertices directly.
This patch exposes functionality for performing partial mesh updates
for normal calculation and face tessellation while transforming a mesh.
The partial update data only needs to be generated once,
afterwards the cached connectivity information can be reused
(with the exception of changing proportional editing radius).
Currently this is only used for transform, in the future it could be
used for other operators as well as the transform panel.
The best-case overall speedup while transforming geometry is about
1.45x since the time to update a small number of normals and faces is
negligible.
For an additional speedup partial face tessellation is multi-threaded,
this gives ~15x speedup on my system (timing tessellation alone).
Exact results depend on the number of CPU cores available.
Ref D11494
Reviewed By: mano-wii
