TBB's ReduceByKey was using the generic DeviceAdapterGeneral
implementation and was about 50x slower than the serial implementation,
which is very efficient.
This patch improves TBB's RBK implementation significantly, though it still
does not scale well. On a quad core processor, this implementation performs
comparably or slightly worse than the highly efficient serial algorithm.
More than 4 cores may be needed to see sufficient parallel speedup that
would overcome the TBB overhead, and grain size does not seem to affect the
performance significantly.
In case where the number of steps taken by each particle is explicitly
provided, code to initialize the arrays for steps and statuses for the
particles was missing.
Added an overloaded version of vtkm::worklet::Threshold::Run that takes
a dynamic cell set instead of a concrete cell set type. This is
implemented by basically moving the CastAndCall from the Threshold
filter to the worklet class. This allows other filters (that might have
similar operations with different predicates) to use the same code. (It
also happens to grately simplify an example for the User's Guide I am
working on.)
Once cell sets use virtual methods, we should be able to collapse code
like this quite a bit anyway.
e9f9a3d8 remove setting of DeveiceAdapter from cosmotools worklet
cdf84ccb Add sample input
6ca2683f Remove the data file for examples
f3766449 Cosmology halo finder
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !919
The filter classes have an internal CellSetIndex member that tracks on
which cell set to operate on. The get accessor is called
GetActiveCellSetIndex (note the descriptive "Index" at the end of the
function name). However, the set accessor was called SetActiveCellSet
(sans "Index"). This discrepancy does not make a lot of sense.
This commit changes SetActiveCellSet to SetActiveCellSetIndex. Not only
do I like the extra descriptor (in case we later want to set cells by
name), it is also used much less than the get method so is less
disruptive.
74c38356 Merge branch 'master' of https://gitlab.kitware.com/vtk/vtk-m into streamline_returntype
e5843882 Merge branch 'master' of https://gitlab.kitware.com/vtk/vtk-m into streamline_returntype
40098181 Remove warnings.
7bbc1417 Merge branch 'streamline_returntype' of gitlab.kitware.com:dpugmire/vtk-m into streamline_returntype
48b3f64c Merge branch 'streamline_returntype' of gitlab.kitware.com:dpugmire/vtk-m into streamline_returntype
b2de8045 Cleanup. Add some additional tests for streamlines.
408765c9 Merge branch 'streamline_returntype' of gitlab.kitware.com:dpugmire/vtk-m into streamline_returntype
7e69b960 Fix merges with master.
...
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !849
Previously, ConvertNumComponentsToOffsets always used TryCompile on the
global set of runtime devices. That is still the default behavior, but
now you are able to specify your own runtime tracker. Also, there are
now versions of ConvertNumComponentsToOffsets that take a device adapter
tag.
Running friend of friends algorithm and then NxN most bound particle
after to find halo center. Cosmology center finder running NxN MBP
algorithm followed by a estimator reducing the problem to MxN MBP
to speed up run.
This allows you to defer its construction. The default constructor
will set up the scatter to have 0 inputs and 0 outputs, so using
it will likely quickly reveal an error.
