This improves performance by **reducing** the amounts of threads used for tasks
which require a high memory bandwidth.
This works because the underlying hardware has a certain maximum memory
bandwidth. If that is used up by a few threads already, any additional threads
wanting to use a lot of memory will just cause more contention which actually
slows things down. By reducing the number of threads that can perform certain
tasks, the remaining threads are also not locked up doing work that they can't
do efficiently. It's best if there is enough scheduled work so that these tasks
can do more compute intensive tasks instead.
To use this new functionality, one has to put the parallel code in question into
a `threading::memory_bandwidth_bound_task(...)` block. Additionally, one also
has to provide a (very) rough approximation for how many bytes are accessed. If
the number is low, the number of threads shouldn't be reduced because it's
likely that all touched memory can be in L3 cache which generally has a much
higher bandwidth than main memory.
The exact number of threads that are allowed to do bandwidth bound tasks at the
same time is generally highly context and hardware dependent. It's also not
really possible to measure reliably because it depends on so many static and
dynamic factors. The thread count is now hardcoded to 8. It seems that this many
threads are easily capable of maxing out the bandwidth capacity.
With this technique I can measure surprisingly good performance improvements:
* Generating a 3000x3000 grid: 133ms -> 103ms.
* Generating a mesh line with 100'000'000 vertices: 212ms -> 189ms.
* Realize mesh instances resulting in ~27'000'000 vertices: 460ms -> 305ms.
In all of these cases, only 8 instead of 24 threads are used. The remaining
threads are idle in these cases, but they could do other work if available.
Pull Request: https://projects.blender.org/blender/blender/pulls/118939
The standard `threading::parallel_for` function tries to split the range into
uniformly sized subranges. This is great if each element takes approximately
the same amount of time to compute.
However, there are also situations where the time required to do the work for
a single index differs significantly between different indices. In such a case,
it's better to split the tasks into segments while taking the size of each task into
account.
This patch implements `threading::parallel_for_weighted` which allows passing
in an additional callback that returns the size of each task.
Pull Request: https://projects.blender.org/blender/blender/pulls/118348
Unless you're very familiar with `IndexRange`, it's often hard to know what
e.g. `IndexRange(10, 15)` means. Without more context, one could think
that it means `10-14`, `10-15` or `10-24`. This patch adds named constructors
to `IndexRange` to make the behavior more obvious when writing and when
reading the code. With those one can use `IndexRange::from_begin_end(10, 15)`,
`IndexRange::from_begin_end_inclusive(10, 15)` or `IndexRange::from_begin_size(10, 15)`
respectively. While being a bit more verbose, the explicitness makes code easier to
understand and also allows abstracting away some common index computations.
The old unnamed constructor that takes a begin and size is not removed by this patch,
as that would make the patch significantly bigger. I think it's reasonable to generally
use the named constructors going forward and to change the existing usages of the
old constructor over time.
Pull Request: https://projects.blender.org/blender/blender/pulls/118606
Listing the "Blender Foundation" as copyright holder implied the Blender
Foundation holds copyright to files which may include work from many
developers.
While keeping copyright on headers makes sense for isolated libraries,
Blender's own code may be refactored or moved between files in a way
that makes the per file copyright holders less meaningful.
Copyright references to the "Blender Foundation" have been replaced with
"Blender Authors", with the exception of `./extern/` since these this
contains libraries which are more isolated, any changed to license
headers there can be handled on a case-by-case basis.
Some directories in `./intern/` have also been excluded:
- `./intern/cycles/` it's own `AUTHORS` file is planned.
- `./intern/opensubdiv/`.
An "AUTHORS" file has been added, using the chromium projects authors
file as a template.
Design task: #110784
Ref !110783.
A lot of files were missing copyright field in the header and
the Blender Foundation contributed to them in a sense of bug
fixing and general maintenance.
This change makes it explicit that those files are at least
partially copyrighted by the Blender Foundation.
Note that this does not make it so the Blender Foundation is
the only holder of the copyright in those files, and developers
who do not have a signed contract with the foundation still
hold the copyright as well.
Another aspect of this change is using SPDX format for the
header. We already used it for the license specification,
and now we state it for the copyright as well, following the
FAQ:
https://reuse.software/faq/
Goals of this refactor:
* Reduce memory consumption of `IndexMask`. The old `IndexMask` uses an
`int64_t` for each index which is more than necessary in pretty much all
practical cases currently. Using `int32_t` might still become limiting
in the future in case we use this to index e.g. byte buffers larger than
a few gigabytes. We also don't want to template `IndexMask`, because
that would cause a split in the "ecosystem", or everything would have to
be implemented twice or templated.
* Allow for more multi-threading. The old `IndexMask` contains a single
array. This is generally good but has the problem that it is hard to fill
from multiple-threads when the final size is not known from the beginning.
This is commonly the case when e.g. converting an array of bool to an
index mask. Currently, this kind of code only runs on a single thread.
* Allow for efficient set operations like join, intersect and difference.
It should be possible to multi-thread those operations.
* It should be possible to iterate over an `IndexMask` very efficiently.
The most important part of that is to avoid all memory access when iterating
over continuous ranges. For some core nodes (e.g. math nodes), we generate
optimized code for the cases of irregular index masks and simple index ranges.
To achieve these goals, a few compromises had to made:
* Slicing of the mask (at specific indices) and random element access is
`O(log #indices)` now, but with a low constant factor. It should be possible
to split a mask into n approximately equally sized parts in `O(n)` though,
making the time per split `O(1)`.
* Using range-based for loops does not work well when iterating over a nested
data structure like the new `IndexMask`. Therefor, `foreach_*` functions with
callbacks have to be used. To avoid extra code complexity at the call site,
the `foreach_*` methods support multi-threading out of the box.
The new data structure splits an `IndexMask` into an arbitrary number of ordered
`IndexMaskSegment`. Each segment can contain at most `2^14 = 16384` indices. The
indices within a segment are stored as `int16_t`. Each segment has an additional
`int64_t` offset which allows storing arbitrary `int64_t` indices. This approach
has the main benefits that segments can be processed/constructed individually on
multiple threads without a serial bottleneck. Also it reduces the memory
requirements significantly.
For more details see comments in `BLI_index_mask.hh`.
I did a few tests to verify that the data structure generally improves
performance and does not cause regressions:
* Our field evaluation benchmarks take about as much as before. This is to be
expected because we already made sure that e.g. add node evaluation is
vectorized. The important thing here is to check that changes to the way we
iterate over the indices still allows for auto-vectorization.
* Memory usage by a mask is about 1/4 of what it was before in the average case.
That's mainly caused by the switch from `int64_t` to `int16_t` for indices.
In the worst case, the memory requirements can be larger when there are many
indices that are very far away. However, when they are far away from each other,
that indicates that there aren't many indices in total. In common cases, memory
usage can be way lower than 1/4 of before, because sub-ranges use static memory.
* For some more specific numbers I benchmarked `IndexMask::from_bools` in
`index_mask_from_selection` on 10.000.000 elements at various probabilities for
`true` at every index:
```
Probability Old New
0 4.6 ms 0.8 ms
0.001 5.1 ms 1.3 ms
0.2 8.4 ms 1.8 ms
0.5 15.3 ms 3.0 ms
0.8 20.1 ms 3.0 ms
0.999 25.1 ms 1.7 ms
1 13.5 ms 1.1 ms
```
Pull Request: https://projects.blender.org/blender/blender/pulls/104629
This is necessary for correctness of the code to avoid duplicate symbols.
In practice, this wasn't necessary yet, because usually we pass lambdas
into these functions which cause every instantiation to have a different
signature.
Alignment here means that the size of the range passed into callback
is a multiple of the alignment value (which has to be a power of two).
This can help with performance when loops in the callback are are
unrolled and/or vectorized. Otherwise, it can potentially reduce
performance by splitting work into more unequally sized chunks.
For example, chunk sizes might be 4 and 8 instead of 6 and 6 when
alignment is 4.
Previously, `tbb::parallel_for` was instantiated every time `threading::parallel_for`
is used. However, when actual parallelism is used, the overhead of a function
call is negilible. Therefor it is possible to move that part out of the header
without causing noticable performance regressions.
This reduces the size of the Blender binary from 308.2 to 303.5 MB, which is
a reduction of about 1.5%.
This can improve performance in some circumstances when there are
vectorized and/or unrolled loops. I especially noticed that this helps
a lot while working on D16970 (got a 10-20% speedup there by avoiding
running into the non-vectorized fallback loop too often).
This adds a new `Interpolate Curves` node. It allows generating new curves
between a set of existing guide curves. This is essential for procedural hair.
Usage:
- One has to provide a set of guide curves and a set of root positions for
the generated curves. New curves are created starting from these root
positions. The N closest guide curves are used for the interpolation.
- An additional up vector can be provided for every guide curve and
root position. This is typically a surface normal or nothing. This allows
generating child curves that are properly oriented based on the
surface orientation.
- Sometimes a point should only be interpolated using a subset of the
guides. This can be achieved using the `Guide Group ID` and
`Point Group ID` inputs. The curve generated at a specific point will
only take the guides with the same id into account. This allows e.g.
for hair parting.
- The `Max Neighbors` input limits how many guide curves are taken
into account for every interpolated curve.
Differential Revision: https://developer.blender.org/D16642
As described in the comment on `BLI_task_isolate`, deadlocks can happen
when isolation is used with threading primitives that separate spawning tasks
from executing them. All threads are waiting the tasks to complete but no
thread is able to continue working due to task isolation.
The fix is to not pass lazy-threading hints through task isolations. This way
isolated regions can't create new tasks in a scheduler further up the call stack.
This may lead to minor slowdowns because less threading may be used.
It's generally possible to get rid of the slowdown again by sending the
lazy-threading hint before entering the isolated region.
In large node setup the threading overhead was sometimes very significant.
That's especially true when most nodes do very little work.
This commit improves the scheduling by not using multi-threading in many
cases unless it's likely that it will be worth it. For more details see the comments
in `BLI_lazy_threading.hh`.
Differential Revision: https://developer.blender.org/D15976
`parallel_invoke` allows executing functions on separate threads.
However, creating tasks in tbb has a measurable amount of overhead.
Therefore, it can be benefitial to disable parallelization when
the amount of work done per function is small.
See D15539 for some benchmark results.
Differential Revision: https://developer.blender.org/D15539
Apply a change similar to e130903060 for
`parallel_reduce`, just like `parallel_for`. I measured a performance
improvement in viewport FPS of at least 10% with 1 million small
instances (one bottleneck was computing many small bounding boxes).
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
We often call `parallel_for` in places with very variable
sized workloads. When many elements are processed,
using multi-threading is great, but when processing
few elements (possibly many times) using `parallel_for`
can result in significant overhead.
I measured that this improves performance by >20% in
the refactored realize instances code I'm working on
separately. The change might also help with debugging
sometimes, because the stack trace is smaller and contains
fewer irrevelant symbols.
From patch D11780 from Erik Abrahamsson.
It parallelizes making the vertices, destruction of map entries,
finding if the result is PWN, finding triangle adjacencies,
and finding the ambient cell.
The latter needs a parallel_reduce from tbb, so added one into
BLI_task.hh so that if WITH_TBB is false, the code will still work.
On Erik's 6-core machine, the elapsed time went from 17.5s to 11.8s
(33% faster) on an intersection of two spheres with 3.1M faces.
On Howard's 24-core machine, the elapsed time went from 18.7s to 10.8s
for the same test.
This makes it easier to use task isolation in c++ code.
Previously, one either had to check `WITH_TBB` (possibly indirectly
through `WITH_OPENVDB`) or one had to use the C function which
is less convenient.
This namespace groups threading related functions/classes. This avoids
adding more threading related stuff to the blender namespace. Also it
makes naming a bit easier, e.g. the c++ version of BLI_task_isolate could
become blender::threading::isolate_task or something similar.
Differential Revision: https://developer.blender.org/D11624
* USD and OpenVDB headers use deprecated TBB headers, suppress all deprecation
warnings there since we have no control over them.
* For our own TBB includes, use the individual headers rather than the tbb.h that
includes everything to avoid warnings, rather than suppressing all.
This is in anticipation of the TBB 2020 upgrade in D10359. Ref D10361.
If a define of NOMINMAX was made before BLI_task.hh was included,
the compiler would emit a
warning C4005: 'NOMINMAX': macro redefinition
warning, to work around this only define it if it is not already
defined, and only undefine it if we were the ones that made the
define earlier.
TBB includes Windows.h which defines a min/max macro
leading to issues when you want to use std::min and
std::max.
This change prevents Windows.h from defining them
sidestepping the issue.
`BKE_mesh_calc_edges` was the main performance bottleneck in D9141.
While openvdb only needed ~115ms, calculating the edges afterwards
took ~960ms. Now with some parallelization this is reduced to ~210ms.
Parallelizing `BKE_mesh_calc_edges` is not entirely trivial, because it
has to perform deduplication and some other things that have to happen
in a certain order. Even though the multithreading improves performance
with more threads, there are diminishing returns when too many threads
are used in this function.
The speedup is mainly achieved by having multiple hash tables that are
filled in parallel. The distribution of the edges to hash tables is based on
a hash (that is different from the hash used in the actual hash tables).
I moved the function to C++, because that made it easier for me to
optimize it. Furthermore, I added `BLI_task.hh` which contains some
light tbb wrappers for parallelization.
Reviewers: campbellbarton
Differential Revision: https://developer.blender.org/D9151
