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
Since {rBeae36be372a6b16ee3e76eff0485a47da4f3c230} the distinction
between float and byte colors is more explicit in the ui. So far, geometry
nodes couldn't really deal with byte colors in general. This patch fixes that.
There is still only one color socket, which contains float colors. Conversion
to and from byte colors is done when read from or writing to attributes.
* Support writing to byte color attributes in Store Named Attribute node.
* Support converting to/from byte color in attribute conversion operator.
* Support propagating byte color attributes.
* Add all the implicit conversions from byte colors to the other types.
* Display byte colors as integers in spreadsheet.
Differential Revision: https://developer.blender.org/D14705
- Missing star prefix.
- Unnecessary indentation.
- Blank line after dot-points
(otherwise doxygen merges with the previous dot-point).
- Use back-slash for doxygen commands.
- Correct spelling.
This implements two optimizations:
* Reduce virtual function call overhead when a non-standard virtual
array is used as input.
* Use a lambda in `type_conversion.cc`.
In my test setup, which creates a float attribute filled with the index,
the running time drops from `4.0 ms` to `2.0 ms`.
Differential Revision: https://developer.blender.org/D14585
This improves performance e.g. when creating an integer attribute
based on an index field. For 4 million vertices, I measured a speedup
from 3.5 ms to 1.2 ms.
This extracts the inner loops into a separate function.
There are two main reasons for this:
* Allows using `__restrict` to indicate that no other parameter
aliases with the output array. This allows for better optimization.
* Makes it easier to search for the generated assembly code,
especially with the `BLI_NOINLINE`.
This simplifies debugging, and can help improve performance
by making it easier for the compiler.
More optimization might still be possible by using `__restrict` in
a few places.
When boolean fields are evaluated and used as selections, we create
a vector of indices. This process is currently single-threaded, but
226f0c4fef added a more optimized multi-threaded version
of this process. It's simple to use this in the field evaluator.
I tested this with the set position node and a random
value node set to boolean mode on a Ryzen 2700x:
| | Before | After | Improvement |
| 10% Selected | 40.5 ms | 29.0 ms | 1.4x |
| 90% Selected | 115 ms | 45.3 ms | 2.5x |
In the future there could be a specialized version for non-span
virtual array selections that uses `materialize` to lower virtual
call overhead.
Differential Revision: https://developer.blender.org/D14436
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 commit adds infrastructure for 8 bit signed integer attributes.
This can be useful given the discussion in T94193, where we want to
store spline type, Bezier handle type, and other small enums as
attributes.
This is only exposed in the interface in the attribute lists, so it
shouldn't be an option in geometry nodes, at least for now.
I expect that this type won't be used directly very often, it
should mostly be cast to an enum type. However, with support
for 8 bit integers, it also makes sense to add things like mixing
implementations for consistency.
Differential Revision: https://developer.blender.org/D13721
Every translation unit that included the modified headers generated
some extra code, even though it was not used. This adds unnecessary
compile time overhead and is annoying when investigating the
generated assembly.
This patch implements the vector types (i.e:`float2`) by making heavy
usage of templating. All vector functions are now outside of the vector
classes (inside the `blender::math` namespace) and are not vector size
dependent for the most part.
In the ongoing effort to make shaders less GL centric, we are aiming
to share more code between GLSL and C++ to avoid code duplication.
####Motivations:
- We are aiming to share UBO and SSBO structures between GLSL and C++.
This means we will use many of the existing vector types and others
we currently don't have (uintX, intX). All these variations were
asking for many more code duplication.
- Deduplicate existing code which is duplicated for each vector size.
- We also want to share small functions. Which means that vector
functions should be static and not in the class namespace.
- Reduce friction to use these types in new projects due to their
incompleteness.
- The current state of the `BLI_(float|double|mpq)(2|3|4).hh` is a
bit of a let down. Most clases are incomplete, out of sync with each
others with different codestyles, and some functions that should be
static are not (i.e: `float3::reflect()`).
####Upsides:
- Still support `.x, .y, .z, .w` for readability.
- Compact, readable and easilly extendable.
- All of the vector functions are available for all the vectors types
and can be restricted to certain types. Also template specialization
let us define exception for special class (like mpq).
- With optimization ON, the compiler unroll the loops and performance
is the same.
####Downsides:
- Might impact debugability. Though I would arge that the bugs are
rarelly caused by the vector class itself (since the operations are
quite trivial) but by the type conversions.
- Might impact compile time. I did not saw a significant impact since
the usage is not really widespread.
- Functions needs to be rewritten to support arbitrary vector length.
For instance, one can't call `len_squared_v3v3` in
`math::length_squared()` and call it a day.
- Type cast does not work with the template version of the `math::`
vector functions. Meaning you need to manually cast `float *` and
`(float *)[3]` to `float3` for the function calls.
i.e: `math::distance_squared(float3(nearest.co), positions[i]);`
- Some parts might loose in readability:
`float3::dot(v1.normalized(), v2.normalized())`
becoming
`math::dot(math::normalize(v1), math::normalize(v2))`
But I propose, when appropriate, to use
`using namespace blender::math;` on function local or file scope to
increase readability.
`dot(normalize(v1), normalize(v2))`
####Consideration:
- Include back `.length()` method. It is quite handy and is more C++
oriented.
- I considered the GLM library as a candidate for replacement. It felt
like too much for what we need and would be difficult to extend / modify
to our needs.
- I used Macros to reduce code in operators declaration and potential
copy paste bugs. This could reduce debugability and could be reverted.
- This touches `delaunay_2d.cc` and the intersection code. I would like
to know @howardt opinion on the matter.
- The `noexcept` on the copy constructor of `mpq(2|3)` is being removed.
But according to @JacquesLucke it is not a real problem for now.
I would like to give a huge thanks to @JacquesLucke who helped during this
and pushed me to reduce the duplication further.
Reviewed By: brecht, sergey, JacquesLucke
Differential Revision: https://developer.blender.org/D13791
This patch implements the vector types (i.e:`float2`) by making heavy
usage of templating. All vector functions are now outside of the vector
classes (inside the `blender::math` namespace) and are not vector size
dependent for the most part.
In the ongoing effort to make shaders less GL centric, we are aiming
to share more code between GLSL and C++ to avoid code duplication.
####Motivations:
- We are aiming to share UBO and SSBO structures between GLSL and C++.
This means we will use many of the existing vector types and others
we currently don't have (uintX, intX). All these variations were
asking for many more code duplication.
- Deduplicate existing code which is duplicated for each vector size.
- We also want to share small functions. Which means that vector
functions should be static and not in the class namespace.
- Reduce friction to use these types in new projects due to their
incompleteness.
- The current state of the `BLI_(float|double|mpq)(2|3|4).hh` is a
bit of a let down. Most clases are incomplete, out of sync with each
others with different codestyles, and some functions that should be
static are not (i.e: `float3::reflect()`).
####Upsides:
- Still support `.x, .y, .z, .w` for readability.
- Compact, readable and easilly extendable.
- All of the vector functions are available for all the vectors types
and can be restricted to certain types. Also template specialization
let us define exception for special class (like mpq).
- With optimization ON, the compiler unroll the loops and performance
is the same.
####Downsides:
- Might impact debugability. Though I would arge that the bugs are
rarelly caused by the vector class itself (since the operations are
quite trivial) but by the type conversions.
- Might impact compile time. I did not saw a significant impact since
the usage is not really widespread.
- Functions needs to be rewritten to support arbitrary vector length.
For instance, one can't call `len_squared_v3v3` in
`math::length_squared()` and call it a day.
- Type cast does not work with the template version of the `math::`
vector functions. Meaning you need to manually cast `float *` and
`(float *)[3]` to `float3` for the function calls.
i.e: `math::distance_squared(float3(nearest.co), positions[i]);`
- Some parts might loose in readability:
`float3::dot(v1.normalized(), v2.normalized())`
becoming
`math::dot(math::normalize(v1), math::normalize(v2))`
But I propose, when appropriate, to use
`using namespace blender::math;` on function local or file scope to
increase readability.
`dot(normalize(v1), normalize(v2))`
####Consideration:
- Include back `.length()` method. It is quite handy and is more C++
oriented.
- I considered the GLM library as a candidate for replacement. It felt
like too much for what we need and would be difficult to extend / modify
to our needs.
- I used Macros to reduce code in operators declaration and potential
copy paste bugs. This could reduce debugability and could be reverted.
- This touches `delaunay_2d.cc` and the intersection code. I would like
to know @howardt opinion on the matter.
- The `noexcept` on the copy constructor of `mpq(2|3)` is being removed.
But according to @JacquesLucke it is not a real problem for now.
I would like to give a huge thanks to @JacquesLucke who helped during this
and pushed me to reduce the duplication further.
Reviewed By: brecht, sergey, JacquesLucke
Differential Revision: https://developer.blender.org/D13791
This patch implements the vector types (i.e:float2) by making heavy
usage of templating. All vector functions are now outside of the vector
classes (inside the blender::math namespace) and are not vector size
dependent for the most part.
In the ongoing effort to make shaders less GL centric, we are aiming
to share more code between GLSL and C++ to avoid code duplication.
Motivations:
- We are aiming to share UBO and SSBO structures between GLSL and C++.
This means we will use many of the existing vector types and others we
currently don't have (uintX, intX). All these variations were asking
for many more code duplication.
- Deduplicate existing code which is duplicated for each vector size.
- We also want to share small functions. Which means that vector functions
should be static and not in the class namespace.
- Reduce friction to use these types in new projects due to their
incompleteness.
- The current state of the BLI_(float|double|mpq)(2|3|4).hh is a bit of a
let down. Most clases are incomplete, out of sync with each others with
different codestyles, and some functions that should be static are not
(i.e: float3::reflect()).
Upsides:
- Still support .x, .y, .z, .w for readability.
- Compact, readable and easilly extendable.
- All of the vector functions are available for all the vectors types and
can be restricted to certain types. Also template specialization let us
define exception for special class (like mpq).
- With optimization ON, the compiler unroll the loops and performance is
the same.
Downsides:
- Might impact debugability. Though I would arge that the bugs are rarelly
caused by the vector class itself (since the operations are quite trivial)
but by the type conversions.
- Might impact compile time. I did not saw a significant impact since the
usage is not really widespread.
- Functions needs to be rewritten to support arbitrary vector length. For
instance, one can't call len_squared_v3v3 in math::length_squared() and
call it a day.
- Type cast does not work with the template version of the math:: vector
functions. Meaning you need to manually cast float * and (float *)[3] to
float3 for the function calls.
i.e: math::distance_squared(float3(nearest.co), positions[i]);
- Some parts might loose in readability:
float3::dot(v1.normalized(), v2.normalized())
becoming
math::dot(math::normalize(v1), math::normalize(v2))
But I propose, when appropriate, to use
using namespace blender::math; on function local or file scope to
increase readability. dot(normalize(v1), normalize(v2))
Consideration:
- Include back .length() method. It is quite handy and is more C++
oriented.
- I considered the GLM library as a candidate for replacement.
It felt like too much for what we need and would be difficult to
extend / modify to our needs.
- I used Macros to reduce code in operators declaration and potential
copy paste bugs. This could reduce debugability and could be reverted.
- This touches delaunay_2d.cc and the intersection code. I would like to
know @Howard Trickey (howardt) opinion on the matter.
- The noexcept on the copy constructor of mpq(2|3) is being removed.
But according to @Jacques Lucke (JacquesLucke) it is not a real problem
for now.
I would like to give a huge thanks to @Jacques Lucke (JacquesLucke) who
helped during this and pushed me to reduce the duplication further.
Reviewed By: brecht, sergey, JacquesLucke
Differential Revision: http://developer.blender.org/D13791
It is common to have fields that contain a constant value. Before this
commit, such constants were represented by operation nodes which
don't have inputs. Having a special node type for constants makes
working with them a bit cheaper.
It also allows skipping some unnecessary processing when evaluating
fields, because constant fields can be detected more easily.
This commit also generalizes the concept of field node types a bit.
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
Previously we used a `CellValue` class to hold the data for a cell,
and called a function to fill it whenever necessary. This is an
unnecessary complication when we have virtual generic arrays
and most data is already easily accessible that way anyway.
This patch removes `CellValue` and uses `fn::GVArray` to provide
access to data instead.
In the future, if rows have different types within a single column,
we can use a `GVArray` of `blender::Any` to interface with the drawing.
Along with that, the use of virtual arrays made it easy to do a
few other cleanups:
- Use selection domain interpolations from rB5841f8656d95
for the mesh selection filter.
- Change the row filter to only calculate for necessary indices.
Differential Revision: https://developer.blender.org/D13478
We often had to use two `FieldEvaluator` instances to first evaluate
the selection and then the remaining fields. Now both can be done
with a single `FieldEvaluator`. This results in less boilerplate code in
many cases.
Performance is not affected by this change. In a separate patch we
could improve performance by reusing evaluated sub-fields that are
used by the selection and the other fields.
Differential Revision: https://developer.blender.org/D13571
This implements an optimization pass for multi-function procedures.
It optimizes memory reuse by moving destruct instructions up.
For more details see the in-code comment.
In very large fields with many short lived intermediate values, this change
can improve performance 3-4x. Furthermore, in such cases, peak memory
consumption is reduced significantly (e.g. 100x lower peak memory usage).
Differential Revision: https://developer.blender.org/D13548
This reduces the number of separate memory allocations done
by the multi-function procedure executor (which is used by the
field evaluation).
Now a linear memory allocator is used to allocate all intermediate
values. Furthermore, more buffers are reused when possible. This
reduces the total amount of allocated memory and improves
cache efficiency because the values are more likely to be in cache
already.
The performance improvement of this patch are most noticable
when few elements are processed by many functions. The situation
will improve even more with D13548, because then buffers can actually
be reused in practice. I measured up to 20% faster field evaluation
in extreme cases with this change.
Calling `foreach_field_input` on a highly nested field (we do that
often) has an exponential running time in the number of nodes.
That is because the same node may be visited many times.
This made Blender freeze on some setups that should work just fine.
Now every field keeps track of its inputs all the time. That replaces
the exponential algorithm with constant time access.
Most of our field inputs are currently specific to geometry. This patch introduces
a new `GeometryFieldInput` that reduces the overhead of adding new geometry
field input.
Differential Revision: https://developer.blender.org/D13489
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).
Previously, there was a fixed grain size for all multi-functions. That was
not sufficient because some functions could benefit a lot from smaller
grain sizes.
This refactors adds a new `MultiFunction::call_auto` method which has the
same effect as just calling `MultiFunction::call` but additionally figures
out how to execute the specific multi-function efficiently. It determines
a good grain size and decides whether the mask indices should be shifted
or not.
Most multi-function evaluations benefit from this, but medium sized work
loads (1000 - 50000 elements) benefit from it the most. Especially when
expensive multi-functions (e.g. noise) is involved. This is because for
smaller work loads, threading is rarely used and for larger work loads
threading worked fine before already.
With this patch, multi-functions can specify execution hints, that allow
the caller to execute it most efficiently. These execution hints still
have to be added to more functions.
Some performance measurements of a field evaluation involving noise and
math nodes, ordered by the number of elements being evaluated:
```
1,000,000: 133 ms -> 120 ms
100,000: 30 ms -> 18 ms
10,000: 20 ms -> 2.7 ms
1,000: 4 ms -> 0.5 ms
100: 0.5 ms -> 0.4 ms
```
Under some circumstances that can lead to more than a 2x
performance increase, because math nodes can better optimize
for the case when the slice is a single value or span.
Currently the geometry nodes evaluator always stores a field for every
type that supports it, even if it is just a single value. This results in a lot
of overhead when there are many sockets that just contain a single
value, which is often the case.
This introduces a new `ValueOrField<T>` type that is used by the geometry
nodes evaluator. Now a field will only be created when it is actually
necessary. See D13307 for more details. In extrem cases this can speed
up the evaluation 2-3x (those cases are probably never hit in practice
though, but it's good to get rid of unnecessary overhead nevertheless).
Differential Revision: https://developer.blender.org/D13307
