Goals:
* Better high level control over where devirtualization occurs. There is always
a trade-off between performance and compile-time/binary-size.
* Simplify using array devirtualization.
* Better performance for cases where devirtualization wasn't used before.
Many geometry nodes accept fields as inputs. Internally, that means that the
execution functions have to accept so called "virtual arrays" as inputs. Those
can be e.g. actual arrays, just single values, or lazily computed arrays.
Due to these different possible virtual arrays implementations, access to
individual elements is slower than it would be if everything was just a normal
array (access does through a virtual function call). For more complex execution
functions, this overhead does not matter, but for small functions (like a simple
addition) it very much does. The virtual function call also prevents the compiler
from doing some optimizations (e.g. loop unrolling and inserting simd instructions).
The solution is to "devirtualize" the virtual arrays for small functions where the
overhead is measurable. Essentially, the function is generated many times with
different array types as input. Then there is a run-time dispatch that calls the
best implementation. We have been doing devirtualization in e.g. math nodes
for a long time already. This patch just generalizes the concept and makes it
easier to control. It also makes it easier to investigate the different trade-offs
when it comes to devirtualization.
Nodes that we've optimized using devirtualization before didn't get a speedup.
However, a couple of nodes are using devirtualization now, that didn't before.
Those got a 2-4x speedup in common cases.
* Map Range
* Random Value
* Switch
* Combine XYZ
Differential Revision: https://developer.blender.org/D14628
Method which overrides a base class's virtual methods are expetced to
be marked with `override`. This also gives better idea to the developers
about what is going on.
This does two things:
* Introduce new `materialize_compressed` methods. Those are used
when the dst array should not have any gaps.
* Add materialize methods in various classes where they were missing
(and therefore caused overhead, because slower fallbacks had to be used).
Use a shorter/simpler license convention, stops the header taking so
much space.
Follow the SPDX license specification: https://spdx.org/licenses
- C/C++/objc/objc++
- Python
- Shell Scripts
- CMake, GNUmakefile
While most of the source tree has been included
- `./extern/` was left out.
- `./intern/cycles` & `./intern/atomic` are also excluded because they
use different header conventions.
doc/license/SPDX-license-identifiers.txt has been added to list SPDX all
used identifiers.
See P2788 for the script that automated these edits.
Reviewed By: brecht, mont29, sergey
Ref D14069
This adds `blender::is_same_any_v` which is the almost the same as
`std::is_same_v`. The difference is that it allows for checking multiple
types at the same time.
Differential Revision: https://developer.blender.org/D13673
This implements four optimizations in the Set Position node:
* Check whether the position input is the current position and ignore
it if it is. This results in a speedup when only the Offset input is used.
* Use multi-threading when copying to computed values to the
position attribute. All geometry types benefit from this.
* Use devirtualization for the offset and position input. This optimizes
the common case that they are either single values or computed
in the fly in a span.
* Write to `Mesh->mvert` directly instead of creating a temporary span.
This makes setting mesh vertex positions even more efficient.
In my simple benchmark I'm using a White Noise node to offset the
position of 1,000,000 vertices. The speed is `20 ms -> 4.5 ms` in the
multi-threaded case and `32 ms -> 22 ms` in the single-threaded case.
For some underlying data (e.g. spans) we had two virtual array
implementations. One for the mutable and one for the immutable
case. Now that most code does not deal with the virtual array
implementations directly anymore (since rBrBd4c868da9f97a),
we can get away with sharing one implementation for both cases.
This means that we have to do a `const_cast` in a few places, but
this is an implementation detail that does not leak into "user code"
(only when explicitly casting a `VArrayImpl` to a `VMutableArrayImpl`,
which should happen nowhere).
Goals of this refactor:
* Simplify creating virtual arrays.
* Simplify passing virtual arrays around.
* Simplify converting between typed and generic virtual arrays.
* Reduce memory allocations.
As a quick reminder, a virtual arrays is a data structure that behaves like an
array (i.e. it can be accessed using an index). However, it may not actually
be stored as array internally. The two most important implementations
of virtual arrays are those that correspond to an actual plain array and those
that have the same value for every index. However, many more
implementations exist for various reasons (interfacing with legacy attributes,
unified iterator over all points in multiple splines, ...).
With this refactor the core types (`VArray`, `GVArray`, `VMutableArray` and
`GVMutableArray`) can be used like "normal values". They typically live
on the stack. Before, they were usually inside a `std::unique_ptr`. This makes
passing them around much easier. Creation of new virtual arrays is also
much simpler now due to some constructors. Memory allocations are
reduced by making use of small object optimization inside the core types.
Previously, `VArray` was a class with virtual methods that had to be overridden
to change the behavior of a the virtual array. Now,`VArray` has a fixed size
and has no virtual methods. Instead it contains a `VArrayImpl` that is
similar to the old `VArray`. `VArrayImpl` should rarely ever be used directly,
unless a new virtual array implementation is added.
To support the small object optimization for many `VArrayImpl` classes,
a new `blender::Any` type is added. It is similar to `std::any` with two
additional features. It has an adjustable inline buffer size and alignment.
The inline buffer size of `std::any` can't be relied on and is usually too
small for our use case here. Furthermore, `blender::Any` can store
additional user-defined type information without increasing the
stack size.
Differential Revision: https://developer.blender.org/D12986
Sometimes functions expect a span instead of a virtual array.
If the virtual array is a span internally already, great. But if it is
not (e.g. the position attribute on a mesh), the elements have
to be copied over to a span.
This patch makes the copying process more efficient by giving
the compiler more opportunity for optimization.
A virtual array is a data structure that is similar to a normal array
in that its elements can be accessed by an index. However, a virtual
array does not have to be a contiguous array internally. Instead, its
elements can be layed out arbitrarily while element access happens
through a virtual function call. However, the virtual array data
structures are designed so that the virtual function call can be avoided
in cases where it could become a bottleneck.
Most commonly, a virtual array is backed by an actual array/span or
is a single value internally, that is the same for every index.
Besides those, there are many more specialized virtual arrays like the
ones that provides vertex positions based on the `MVert` struct or
vertex group weights.
Not all attributes used by geometry nodes are stored in simple contiguous
arrays. To provide uniform access to all kinds of attributes, the attribute
API has to provide virtual array functionality that hides the implementation
details of attributes.
Before this refactor, the attribute API provided its own virtual array
implementation as part of the `ReadAttribute` and `WriteAttribute` types.
That resulted in unnecessary code duplication with the virtual array system.
Even worse, it bound many algorithms used by geometry nodes to the specifics
of the attribute API, even though they could also use different data sources
(such as data from sockets, default values, later results of expressions, ...).
This refactor removes the `ReadAttribute` and `WriteAttribute` types and
replaces them with `GVArray` and `GVMutableArray` respectively. The `GV`
stands for "generic virtual". The "generic" means that the data type contained
in those virtual arrays is only known at run-time. There are the corresponding
statically typed types `VArray<T>` and `VMutableArray<T>` as well.
No regressions are expected from this refactor. It does come with one
improvement for users. The attribute API can convert the data type
on write now. This is especially useful when writing to builtin attributes
like `material_index` with e.g. the Attribute Math node (which usually
just writes to float attributes, while `material_index` is an integer attribute).
Differential Revision: https://developer.blender.org/D10994
This adds support for mutable virtual arrays and provides many utilities
for creating virtual arrays for various kinds of data. This commit is
preparation for D10994.
In some multi-functions (such as a simple add function), the virtual method
call overhead to access array elements adds significant overhead. For these
simple functions it makes sense to generate optimized versions for different
types of virtual arrays. This is done by giving the compiler all the information
it needs to devirtualize virtual arrays.
In my benchmark this speeds up processing a lot of data with small function 2-3x.
This devirtualization should not be done for larger functions, because it increases
compile time and binary size, while providing a negilible performance benefit.
When a function is executed for many elements (e.g. per point) it is often the case
that some parameters are different for every element and other parameters are
the same (there are some more less common cases). To simplify writing such
functions one can use a "virtual array". This is a data structure that has a value
for every index, but might not be stored as an actual array internally. Instead, it
might be just a single value or is computed on the fly. There are various tradeoffs
involved when using this data structure which are mentioned in `BLI_virtual_array.hh`.
It is called "virtual", because it uses inheritance and virtual methods.
Furthermore, there is a new virtual vector array data structure, which is an array
of vectors. Both these types have corresponding generic variants, which can be used
when the data type is not known at compile time. This is typically the case when
building a somewhat generic execution system. The function system used these virtual
data structures before, but now they are more versatile.
I've done this refactor in preparation for the attribute processor and other features of
geometry nodes. I moved the typed virtual arrays to blenlib, so that they can be used
independent of the function system.
One open question for me is whether all the generic data structures (and `CPPType`)
should be moved to blenlib as well. They are well isolated and don't really contain
any business logic. That can be done later if necessary.
