18b5be92d Fix issue with CUDA and ArrayHandleMultiplexer
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !2168
`ArrayHandle::PrepareForOutput` often has to reallocate the array to the
specified size. Previously, this allocation was not happening with the
`Token` that is passed to `PrepareForOutput`. If the `ArrayHandle` is
already attached or enqueued for that `Token`, then the allocation would
deadlock.
You can now pass a `Token` object to `Allocate`, which is what
`PrepareForOutput` does.
When you try to call the `Reduce` operation in the CUDA device adapter
with a sufficently complex interator type, you get a compile error
that says `error: cannot pass an argument with a user-provided
copy-constructor to a device-side kernel launch`.
This appears to be a bug in either nvcc or Thrust. I believe it is
related to the following reported issues:
* https://github.com/thrust/thrust/issues/928
* https://github.com/thrust/thrust/issues/1044
Work around this problem by making a special condition for calling
`Reduce` with an `ArrayHandleMultiplexer` that calls the generic
algorithm in `DeviceAdapterAlgorithmGeneral` instead of the algorithm in
Thrust.
Often when a user gives memory to an `ArrayHandle`, she wants data to be
written into the memory given to be used elsewhere. Previously, the
`Buffer` objects would delete the given buffer as soon as a write buffer
was created elsewhere. That was a problem if a user wants VTK-m to write
results right into a given buffer.
Instead, when a user provides memory, "pin" that memory so that the
`ArrayHandle` never deletes it.
The buffer class encapsulates the movement of raw C arrays between
host and devices.
The `Buffer` class itself is not associated with any device. Instead,
`Buffer` is used in conjunction with a new templated class named
`DeviceAdapterMemoryManager` that can allocate data on a given
device and transfer data as necessary. `DeviceAdapterMemoryManager`
will eventually replace the more complicated device adapter classes
that manage data on a device.
The code in `DeviceAdapterMemoryManager` is actually enclosed in
virtual methods. This allows us to limit the number of classes that
need to be compiled for a device. Rather, the implementation of
`DeviceAdapterMemoryManager` is compiled once with whatever compiler
is necessary, and then the `RuntimeDeviceInformation` is used to
get the correct object instance.
Previously, the VariantArrayHandle::AsMultiplexer operation silently
failed and returned an invalid ArrayHandleMultiplexer. This is now
changed to throw an exception if the desired ArrayHandleMultiplexer
cannot hold the VariantArrayHandle's array.
Because ArrayHandle currently only supports one device at a time, it is
possible that a `PrepareForInput` might actually need to wait for write
access so that it could move the data between devices. However, we don't
want the `Token` to be attached for writing because that could block
other read operations.
To get around this, add a hack to WaitToWrite to allow it to attach for
reading instead of writing.
An issue that was identified for the thread safety of `ArrayHandle` is
that if several threads are waiting to use an `ArrayHandle`, there might
be an expectation of the order in which the operations happen. For
example, if one thread is modifying the contents of an `ArrayHandle` and
another is reading those results, we would need the first one to start
before the second one.
To solve this, a queue is added to `ArrayHandle` such that when waiting
to read or write an `ArrayHandle` the `Token` has to be at the top of
the queue in addition to other requirements being met.
Additionally, an `Enqueue` method is added to add a `Token` to the queue
without blocking. This allows a control thread to queue the access and
then spawn a thread where the actual work will be done. As long as
everything is enqueued on the main thread, the operations will happen in
the expected order.
Initially, the probe filter would simply not set a value if a sample was
outside the input `DataSet`. This is not great as the memory could be
left uninitalized and lead to unpredictable results. The testing
compared these invalid results to 0, which seemed to work but is
probably unstable.
This was partially fixed by a previous change that consolidated to
mapping of cell data with a general routine that permuted data. However,
the fix did not extend to point data in the input, and it was not
possible to specify a particular invalid value.
This change specifically updates the probe filter so that invalid values
are set to a user-specified value.
If you gave ReduceByKey a fancy output array that decorated another
array, you could get a runtime error for using an invalid array (if the
device adapter used the generic algorithm). The problem was that
ReduceByKey creates a temporary array, and that array was given the same
storage as the output array. That might not be valid for fancy arrays,
so instead use the default storage for the temporary array.
If you gave ScanInclusiveByKey a fancy output array that decorated
another array, you would get a runtime error for using an invalid array.
The problem was that ScanInclusiveByKey creates a temporary output array
and then copies the result to the actual output array. The problem was
that the temporary output array was given the same storage as the output
array, which won't work if the output array is fancy. Instead, make the
storage for the temporary array default.
The `ArrayHandleStreaming` class stems from an old research project
experimenting with bringing data from an `ArrayHandle` in parts and
overlapping device transfer and execution. It works, but only in very
limited contexts. Thus, it is not actually used today. Plus, the feature
requires global indexing to be permutated throughout the worklet
dispatching classes of VTK-m for no further reason.
Because it is not really used, there are other more promising approaches
on the horizon, and it makes further scheduling improvements difficult,
we are removing this functionality.
42bc9a393 Fix gaps in type support
dc112b516 Enable changing policy used for library compiles
76f870150 Type check input and output array arguments differently
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !1997
Previously, the PolicyDefault used to compile all the filters was hard-
coded. The problem was that if any external project that depends on VTK-
m needs a different policy, it had to recompile everything in its own
translation units with a custom policy.
This change allows an external project provide a simple header file that
changes the type lists used in the default policy. That allows VTK-m to
compile the filters exactly as specified by the external project.
Previously, the ArrayPortalCheck wrapper did not allow access to the
superclass' Get for 3D indices. This solves that problem and also fixes
it for Set (assuming there is ever an instance of that).
Previously, when a ReadPortal or a WritePortal was returned from an
ArrayHandle, it had wrapped in it a Token that was attached to the
ArrayHandle. This Token would prevent other reads and writes from the
ArrayHandle.
This added safety in the form of making sure that the ArrayPortal was
always valid. Unfortunately, it also made deadlocks very easy. They
happened when an ArrayPortal did not leave scope immediately after use
(which is not all that uncommon).
Now, the ArrayPortal no longer locks up the ArrayHandle. Instead, when
an access happens on the ArrayPortal, it checks to make sure that
nothing has happened to the data being accessed. If it has, a fatal
error is reported to the log.
When `ArrayPortalToken::Detach` was called, the contained `Token` was
detached, but `ArrayPortal` being wrapped might also be holding its own
portal that it is decorating. To ensure that any dependent portals are
also detached, `ArrayPortalToken::Detach` resets itself.
This fixes an issue where getting a `ReadPortal` or a `WritePortal` from
an `ArrayHandleVirtual` could cause a deadlock from a held token even if
the returned portal was detached or destroyed.
The problem was that `ArrayHandleVirtual` was keeping a reference to the
`ArrayPortal` from the concrete array. This was because the returned
`ArrayPortalRef`, which was designed to work on both control and
execution environments, had no good way to destroy the portal. This
meant that the `ArrayHandleVirtual` was caching a copy of the concrete
array's portal. This was not a great idea before because the array could
get invalidated. It is worse now because it keeps the concrete array
locked.
Fixed the problem by subclassing `vtkm::ArrayPortalRef` to make a
control-specific version that will delete the concrete portal on its own
destruction.
The ArrayPortalWrapper is used for both execution and control portals.
When it was wrapped around a control portal that does not work on CUDA
devices, we were getting ugly warnings even though the intention was
only to use it in the control environment.
To get a portal to access ArrayHandle values in the control
environment, you now use the ReadPortal and WritePortal methods.
The portals returned are wrapped in an ArrayPortalToken object
so that the data between the portal and the ArrayHandle are
guaranteed to be consistent.
It is questionable whether there is a point to having a token object
when transfering a virtual object to a device (since there is a handle
object that is managing it anyway). Back out of passing the token all
the way down unless there is an actual need for that.
Marked the old versions of PrepareFor* that do not use tokens as
deprecated and moved all of the code to use the new versions that
require a token. This makes the scope of the execution object more
explicit so that it will be kept while in use and can potentially be
reclaimed afterward.
When a single `ArrayHandle` is given to multiple arguments of a worklet
dispatch, the `PrepareFor*` methods will be called multiple times with
the same token. If one of them is a `PrepareForInPlace` or
`PrepareForOutput`, then the two requests will deadlock. To prevent
this, allow the `PrepareFor*` to happen if the same token was used
previously.
The old version of ExecutionObject (that only takes a device) is still
supported, but you will get a deprecated warning if that is what is
defined.
Supporing this also included sending vtkm::cont::Token through the
vtkm::cont::arg::Transport mechanism, which was a change that propogated
through a lot of code.
Duplicated the new versions of PrepareFor* methods from the basic
ArrayHandle that take a token in addition to the other arguments. The
ArrayHandle attaches itself to the token and will not allow operaitons
that make the returned portal invalid until the token goes out of scope.
Later the old versions will be deprecated.
Added new versions of PrepareFor* methods that take a token in addition
to the other arguments. The ArrayHandle attaches itself to the token and
will not allow operations that make the returned portal invalid until
the token goes out of scope.
Later the old versions will be deprecated.
The convenience functions `ArrayPortalToIteratorBegin()` and
`ArrayPortalToIteratorEnd()` wouldn't detect specializations of
`ArrayPortalToIterators<PortalType>` since the specializations aren't
visible when the `Begin`/`End` functions are declared.
Since the CUDA iterators rely on a specialization, the convenience
functions would not compile on CUDA.
Now, instead of specializing `ArrayPortalToIterators` to provide custom
iterators for a particular portal, the portal may advertise custom
iterators by defining `IteratorType`, `GetIteratorBegin()`, and
`GetIteratorEnd()`. `ArrayPortalToIterators` will detect such portals
and automatically switch to using the specialized portals.
This eliminates the need for the specializations to be visible to the
convenience functions and allows them to be usable on CUDA.
This behaves just like `ScanExclusive`, but rather than returning the
total sum, it is appended to the end of the output array.
This is in preparation for the CellSetExplicit refactoring described in
issue #408.
- Use AtomicInterface to implement device-specific atomic operations.
- Remove DeviceAdapterAtomicArrayImplementations.
- Extend supported atomic types to include unsigned 32/64-bit ints.
- Add a static_assert to check that AtomicArray type is supported.
- Add documentation for AtomicArrayExecutionObject, including a CAS
example.
- Add a `T Get(idx)` method to AtomicArrayExecutionObject that does
an atomic load, and update existing CAS usage to use this instead
of `Add(idx, 0)`.
The `From` and `To` nomenclature for topology mapping has been confusing for
both users and developers, especially at lower levels where the intention of
mapping attributes from one element to another is easily conflated with the
concept of mapping indices (which maps in the exact opposite direction).
These identifiers have been renamed to `VisitTopology` and `IncidentTopology`
to clarify the direction of the mapping. The order in which these template
parameters are specified for `WorkletMapTopology` have also been reversed,
since eventually there may be more than one `IncidentTopology`, and having
`IncidentTopology` at the end will allow us to replace it with a variadic
template parameter pack in the future.
Other implementation details supporting these worklets, include `Fetch` tags,
`Connectivity` classes, and methods on the various `CellSet` classes (such as
`PrepareForInput` have also reversed their template arguments. These will need
to be cautiously updated.
The convenience implementations of `WorkletMapTopology` have been renamed for
clarity as follows:
```
WorkletMapPointToCell --> WorkletVisitCellsWithPoints
WorkletMapCellToPoint --> WorkletVisitPointsWithCells
```
The `ControlSignature` tags have been renamed as follows:
```
FieldInTo --> FieldInVisit
FieldInFrom --> FieldInMap
FromCount --> IncidentElementCount
FromIndices --> IncidentElementIndices
```
Previously, IsWriteableArrayHandle just checked to see if an
ArrayHandle's portal has a ValueType of void* because we had coded the
special read-only array handles to have fake portals for writing.
However, we recently removed that because it was more trouble than it
was worth. Now IsWritableArrayHandle checks for the existance of the Set
method. If it does not exist, then the portal is considered read-only.
Also corrected spelling (writeable -> writable).
As the RuntimeDeviceTracker is a per thread construct we now make
it explicit that you can only get a reference to the per-thread
version and can't copy it.
When TransferInfo is given memory from VirtualObjectTransferShareWithControl
it doesn't have a bound function ptr for the destruction. In those cases
we need to make sure the HostCopyOfDevice is properly deleted, otherwise
we will cause a memory leak.
ff687016e For VTK-m libs all includes of DeviceAdapterTagCuda happen from cuda files
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Kenneth Moreland <kmorel@sandia.gov>
Merge-request: !1648
It is very easy to cause ODR violations with DeviceAdapterTagCuda.
If you include that header from a C++ file and a CUDA file inside
the same program we an ODR violation. The reasons is that the C++
versions will say the tag is invalid, and the CUDA will say the
tag is valid.
The solution to this is that any compilation unit that includes
DeviceAdapterTagCuda from a version of VTK-m that has CUDA enabled
must be invoked by the cuda compiler.
Fixes#277
DeviceAdapterError existed to make sure that the default device adapter
template was being handled properly. Since the default device adapter doesn't
exist, and nothing is templated over it we can now remove DeviceAdapterError.
661fb64de AtomicInterfaceControl functions are marked with VTKM_SUPPRESS_EXEC_WARNINGS
0c70f9b9a Add BitFieldIn/Out/InOut worklet signature tags.
a66510e81 Add ArrayHandleBitField, a boolean-valued AH backed by a BitField.
56cc5c3d3 Add support for BitFields.
d01b97382 Allow VTKM_SUPPRESS_EXEC_WARNINGS to be used inside macros.
2f2ca9370 Add bit operations FindFirstSetBit and CountSetBits to Math.h.
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !1629
BitFields are:
- Stored in memory using a contiguous buffer of bits.
- Accessible via portals, a la ArrayHandle.
- Portals operate on individual bits or words.
- Operations may be atomic for safe use from concurrent kernels.
The new BitFieldToUnorderedSet device algorithm produces an ArrayHandle
containing the indices of all set bits, in no particular order.
The new AtomicInterface classes provide an abstraction into bitwise
atomic operations across control and execution environments and are used
to implement the BitPortals.
18ff6681f Less vtkm_cont cxx files bring in the cuda device
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Kenneth Moreland <kmorel@sandia.gov>
Merge-request: !1612
This adds an ExecutionSignature tag named Device that passes the
DeviceAdapterTag as an argument to the worklet's operator(). This allows
worklets to specialize their code based on the device.
Previously we just took the optionparser.h file and stuck it right in
our source code. That was problematic for a variety of reasons.
1. It incorrectly assigned our license to external code.
2. It made lots of unnecessary changes to the original source (like
reformatting).
3. It made it near impossible to track patches we make and updates to
the original software.
Instead, use the third-party system to track changes to optionparser.h
in a different repository and then pull that into ours.
The RuntimeDeviceTracker had grown organically to handle multiple
different roles inside VTK-m. Now that we have device tags
that can be passed around at runtime, large portions of
the RuntimeDeviceTracker API aren't needed.
Additionally the RuntimeDeviceTracker had a dependency on knowing
the names of each device, and this wasn't possible
as that information was part of its self. Now we have moved that
information into RuntimeDeviceInformation and have broken
the recursion.
e1f5c4dd9 Modify VariantAH::AsVirtual to cast to new ValueType if needed.
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !1585
E.g:
```
ArrayHandle<Float64> doubleArray;
VariantArrayHandle varHandle{doubleArray};
ArrayHandleVirtual<Float32> = varHandle.AsVirtual<Float32>();
```
If there is a loss in range and/or precision, a warning is logged. If
the ValueTypes are Vecs with mismatched widths, an ErrorBadType is thrown.
Internally, an ArrayHandleCast is used between the VariantArrayHandle's
stored array and the ArrayHandleVirtual.
Query and report number of components at runtime so that it can work with
arrays of types with runtime varying number of components
Eg. ArrayHandleGroupVecVariable.
The ArrayPortalValueReference is supposed to behave just like the value
it encapsulates and does so by automatically converting to the base type
when necessary. However, when it is possible to convert that to
something else, it is possible to get errors about ambiguous overloads.
To avoid these, add specialized versions of the operators to specify
which ones should be used.
Also consolidated the CUDA version of an ArrayPortalValueReference to the
standard one. The two implementations were equivalent and we would like
changes to apply to both.
`vtkm::cont::testing` now initializes with logging enabled and support
for device being passed on the command line, `vtkm::testing` only
enables logging.
When you call VariantArrayHandle::CastAndCall, it now tries both basic
storage and virtual storage. You can modify the types of storages tried
by giving a type list of storage tags as the first argument.
We want to make sure that VariantArrayHandleContainer has as little
overhead when launch worklets as possible. To do so we cache
type information to make deducing the `T` type of ArrayHandles
as fast as possible.
Previously you had to exactly match the case of a device adapter's name to
construct it, which was a source of lots of problems ( OpenMP versus OPENMP, CUDA or Cuda ).
Now `vtkm::cont::make_DeviceAdapterId` and `vtkm::cont::RuntimeDeviceTracker` support
case-insensitive device construction.
ArrayHandleVirtual can automatically be constructed from any ArrayHandle.
In the cases where the input ArrayHandle doesn't derived from ArrayHandleVirtual,
it will automatically construct StorageAny to hold the array.
Also
- Renamed vtkm::cont::make_DeviceAdapterIdFromName to just overload
make_DeviceAdapterId.
- Refactored CMake logic for unit tests
- Since we're now querying the device tracker for the names, they
cannot be all caps.
- Updated usages of InitLogging to use Initialize instead.
- Added changelog.
VTK-m has been updated to replace old per device worklet testing executables with a device
dependent shared library so that it's able to accept a device adapter
at runtime.
Meanwhile, it updates the testing infrastructure APIs. vtkm::cont::testing::Run
function would call ForceDevice when needed and if users need the device
adapter info at runtime, RunOnDevice function would pass the adapter into the functor.
Optional Parser is bumped from 1.3 to 1.7.
Also add a throwFailedRuntimeDeviceTransfer that throws a nicely
detailed message on why a something couldn't be transfered to
the requested device adapter.
The OpenMP Device Reduction algorithm previously used a std::vector<T>
to store the reduction results of each thread. This caused problems
when T=bool as the types became a proxy type which isn't usable
with vtkm BinaryOperators.
Additionally by fixing this issue in the FunctorsOpenMP we
can remove a workaround in FunctorsGeneral that caused
compile failures when using complex BinaryOperators
such as MinAndMax.
Previously ArrayHandleBasicImpl had no support for OpenMP since
we forgot to update the implementation. This version will
work when adding new devices without any changes.
Benchmarking in VTK showed significant overhead in the computation
of the reverse connectivity calculation in
ConnectivityExplicitInternals::ComputeCellToPointConnectivity.
This patch adds a ReverseConnectivityBuilder that reduces the amount of
time and memory needed to build the table by using an atomic histogram
approach that avoids a costly radix SortByKey.
Key operations in the new helper class are templated to allow this
approach to be reused by VTK-specific cell array converters.
There is no real reason why you cannot construct an
ArrayPortalFromIterators on a device, so go ahead and let that happen.
(This removes some CUDA warnings about calling __host__ from
__device__.)
Error: Throwing an exception in CUDA code.
Fix: Change method throwing exception to VTKM_CONT.
New warning: host/device warning in taotuple.
Fix: Markup additional taotuple methods with suppressions.
This also updates our taotuple checkout to match upstream master.
Previously, to query whether an ArrayHandle was writable with
IsWriteableArrayHandle, you had to specify a device adapter. The idea
was that it would look at the portal used for that device adapter.
Instead, check the control pointer, which should give the same
indication without having to have a separate check for every type of
device.
Found via `codespell` and `grep`
more typos
includes source typo change and a typo that needs further review
follow-up typos
Follow-up typos
Revert a commit
183bcf109 Add initial version of an OpenMP backend.
7b5ad3e80 Expand device scheduler test to check for overlap.
e621b6ba3 Generalize the TBB radix sort implementation.
d60278434 Specialize swap for ArrayPortalValueReference types.
761f8986f Cache inputs to SSI::Unique benchmark.
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !1099
c3c8d0fd Refactor ArrayHandleCompositeVector to simplify usage and impl.
ec88b7dd Markup array portals for use in the exec env.
3159b376 Make Swizzle and ExtractComponent array parameters runtime vars.
19344707 Add vtkm_taotuple to build system.
d23c4452 Merge branch 'upstream-taotuple' into tmp
b3b14de7 taotuple 2018-05-15 (e3de8c97)
b00f6c1c Add taocpp/tuple as a thirdparty package.
Acked-by: Kenneth Moreland <kmorel@sandia.gov>
Merge-request: !1199
The previous implementation of DeviceAdapterRuntimeDetector caused
multiple differing definitions of the same class to exist and
was causing the runtime device tracker to report CUDA as disabled
when it actually was enabled.
The ODR was caused by having a default implementation for
DeviceAdapterRuntimeDetector and a specific specialization for
CUDA. If a library had both CUDA and C++ sources it would pick up
both implementations and would have undefined behavior. In general
it would think the CUDA backend was disabled.
To avoid this kind of situation in the future I have reworked VTK-m
so that each device adapter must implement DeviceAdapterRuntimeDetector
for that device.
Previously memory that was allocated outside of VTK-m was impossible to transfer to
VTK-m as we didn't know how to free it. By extending the ArrayHandle constructors
to support a Storage object that is being moved, we can clearly express that
the ArrayHandle now owns memory it didn't allocate.
Here is an example of how this is done:
```cpp
T* buffer = new T[100];
auto user_free_function = [](void* ptr) { delete[] static_cast<T*>(ptr); };
vtkm::cont::internal::Storage<T, vtkm::cont::StorageTagBasic>
storage(buffer, 100, user_free_function);
vtkm::cont::ArrayHandle<T> arrayHandle(std::move(storage));
```
By hard coding the PrepareForDevice to know about all the different VTK-m
devices, we can have a single base class do the execution allocation, and not
have that logic repeated in each child class.
