Scatter in worklets
Add the functionality to perform a scatter operation from input to output in a worklet invocation. This allows you to, for example, specify a variable amount of outputs generated for each input.
See merge request !221
The original workaround for inclusive_scan bugs in thrust 1.8 only solved the
issue for basic arithmetic types such as int, float, double. Now we go one
step further and fix the problem for all types.
The solution is to provide a proper implementation of destructive_accumulate_n
and make sure it exists before any includes of thrust occur.
Previously, there was a declaration ConstArrayPortalFromThrust<const T>
in ArrayManagerExecutionThrustDevice. This proved problematic because
values read from the array in the worklet were typed as const T rather
than simply T. Any Vec or Matrix built from that type would then fail
because they are not meant to work with a const value (which means they
have to be set on construction and never changed.
Instead, declare ConstArrayPortalFromThrust<T> and internally set all
the Thrust pointers to have type const T. Also declare other thrust
pointers used as method parameters to have const T rather than T. This
should work as conversion from T to const T should be fine, but not the
other way around.
There is a strange nvcc warning in CUDA 7.5 that sometimes happens on MSVC
that causes it to emit a warning for an undefined method that is clearly
defined. The CUDA development team is aware of the problem and is going
to fix it, but these changes will work around the problem for now.
Thanks to Tom Fogal from NVIDIA for these fixes.
This now allows for even more efficient construction of uniform point
coordinates when running under the 3d scheduler, since we don't need to go
from 3d index to flat index to 3d index, instead we stay in 3d index
Array handles for cuda device pointers have been implemented. The data for
these handles exists solely on the exec side (info such as length can be
queried from the cont side).
ArrayHandles in DAX have a CopyInto function which allows the user to copy an array handle's data into a compatible STL type iterator. Originally this was fairly straight forward to implement since array handles in DAX are templated on the DeviceAdapterTag. In contrast, VTKm array handles use a polymorphic ArrayHandleExecutionManager under the hood allowing a single array handle to interface with multiple devices at runtime. To achieve this virtual functions are used. This makes implementing the CopyInto function difficult since it is templated on the IteratorType and virtual functions cannot be templated.
To work around this, I've implemented a concrete templated CopyInto function in the class derived from ArrayHandleExecutionManagerBase. In the ArrayHandle class, CopyInto dynamically casts the base class into the derived class, then calls the CopyInto function defined in the derived class.
The drawback to this approach is that, should the user define their own class that inherits from ArrayHandleExectionManagerBase, they are not forced to implement the CopyInto function, unlike the other virtual functions.
Previously it was really hard to verify if a device adapter was valid. Since
you would have to check for the existence of the tag. Now the tag always
exists, but instead you query the traits of the DeviceAdapter to see if
it is a valid adapter.
This makes compiling with multiple backends alot easier.
By introducing our own custom thrust execution policy we can make sure
to hit the fastest code paths in thrust for the sort operation. This makes
sure that for UInt32,Int32, and Float32 we use the radix sort from thrust
which offers a 2x to 3x speed improvement over the merge sort implementation.
Secondly by telling thrust that our BinaryOperators are commutative we
make sure that we get the fastest code paths when executing Inclusive
and Exclusive Scan
Benchmark 'Radix Sort on 1048576 random values vtkm::Int32' results:
median = 0.0117049s
median abs dev = 0.00324614s
mean = 0.0167615s
std dev = 0.00786269s
min = 0.00845875s
max = 0.0389063s
Benchmark 'Radix Sort on 1048576 random values vtkm::Float32' results:
median = 0.0234463s
median abs dev = 0.000317249s
mean = 0.021452s
std dev = 0.00470307s
min = 0.011255s
max = 0.0250643s
Benchmark 'Merge Sort on 1048576 random values vtkm::Int32' results:
median = 0.0310486s
median abs dev = 0.000182129s
mean = 0.0286914s
std dev = 0.00634102s
min = 0.0116225s
max = 0.0317379s
Benchmark 'Merge Sort on 1048576 random values vtkm::Float32' results:
median = 0.0310617s
median abs dev = 0.000193583s
mean = 0.0295779s
std dev = 0.00491531s
min = 0.0147257s
max = 0.032307s
Starting in thrust 1.8 the implementation of scan inclusive inside
thrust became highly optimized by using parallel task groups. This
new implementation has a bug that only exists when using custom
binary operators, large size arrays, release mode, and no
debugger or mem-checker attached.
While I have submitted the issue to thrust, we need to be able
to work around the existing issue. The solution I have chosen is
to mark all vtkm::exec::cuda::interal::WrappedBinaryOperators
as being commutative as far as thrust is concerened. To make
sure we don't get any unexpected behavior I have also had
to create WrappedBinaryPredicate so that we don't mark any
predicate as commutative.
The original implementing tried to run 2^31 kernels and detect a
launch failure to determine this use-case. The issue with this approach
is that on a cuda 3+ gpu, this would take multiple seconds and cause
the gpu to terminate the kernel when opengl was also loaded.
The initial implementation forgot about the fact that SM2.X architectures
can only handle 65k blocks. Now we gracefully handle when compiling for
SM2.X.