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.
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.
1f1688483 Initial infrastructure to allow WorkletMapField to have 3D scheduling
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Kenneth Moreland <kmorel@sandia.gov>
Merge-request: !1938
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.
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.
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.
Sometimes the CUDA runtime would not allocate sufficient stack
space for the particle advection code to run. This issue was exposed by
!1737 -- for some reason, once those changes to unrelated filters/worklets
are added to VTK, CUDA allocates less stack and the following tests would
fail:
UnitTestLagrangianFilterCUDA
UnitTestLagrangianStructuresFilterCUDA
UnitTestStreamlineFilterCUDA
UnitTestStreamSurfaceFilterCUDA
These were fixed by increasing the stack size in the particle advection
worklet Run(...) methods.
An RAII helper has been added that will restore the previous stack size
in case an exception is thrown, and the KDTree code has been updated
to use this helper when it adjusts the CUDA stack allocation.
- 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)`.
how did any of this work?
match other CellSet file layouts.
???
compile in CUDA.
unit tests.
also only serial.
make error message accurate
Well, this compiles and works now.
Did it ever?
use CellShapeTagGeneric
UnitTest matches previous changes.
whoops
Fix linking problems.
Need the same interface
as other ThreadIndices.
add filter test
okay, let's try duplicating CellSetStructure.
okay
inching...
change to wedge in CellSetListTag
Means changing these to support it.
switch back to wedge from generic
compiles and runs
remove ExtrudedType
need vtkm_worklet
vtkm_worklet needs to be included
fix segment count for wedge specialization
need to actually save the index
for the other constructor.
specialize on Explicit
clean up warning
angled brackets not quotes.
formatting
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.
The 8x8x8 is a better launch strategy for most VTK-m kernels.
The current problem is that a couple of VTK-m kernels use a
high number of registers and this number of threads combines to
require too many registers.
What we should do in the longer run is have more controls over
kernel launches on a per kernel basis. This will require VTK-m
to extract the number of registers being used by each kernel
The consistent API for control to execution memory transfers is
the ArrayHandle class. Previously the tests would verify memory
transfer by calling the ArrayManagerExecution class directly. This
is problematic as the class isn't used by ArrayHandle<T, StorageBasic>.
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.
VTK-m now offers a more GPU aware set of defaults for kernel scheduling.
When VTK-m first launches a kernel we do system introspection and determine
what GPU's are on the machine and than match this information to a preset
table of values. The implementation is designed in a way that allows for
VTK-m to offer both specific presets for a given GPU ( V100 ) or for
an entire generation of cards ( Pascal ).
Currently VTK-m offers preset tables for the following GPU's:
- Tesla V100
- Tesla P100
If the hardware doesn't match a specific GPU card we than try to find the
nearest know hardware generation and use those defaults. Currently we offer
defaults for
- Older than Pascal Hardware
- Pascal Hardware
- Volta+ Hardware
Some users have workloads that don't align with the defaults provided by
VTK-m. When that is the cause, it is possible to override the defaults
by binding a custom function to `vtkm::cont::cuda::InitScheduleParameters`.
As shown below:
```cpp
ScheduleParameters CustomScheduleValues(char const* name,
int major,
int minor,
int multiProcessorCount,
int maxThreadsPerMultiProcessor,
int maxThreadsPerBlock)
{
ScheduleParameters params {
64 * multiProcessorCount, //1d blocks
64, //1d threads per block
64 * multiProcessorCount, //2d blocks
{ 8, 8, 1 }, //2d threads per block
64 * multiProcessorCount, //3d blocks
{ 4, 4, 4 } }; //3d threads per block
return params;
}
vtkm::cont::cuda::InitScheduleParameters(&CustomScheduleValues);
```
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.
When reducing an input type that differs from the output type
you need to write a custom binary operator that also implements
how to do the unary transformation.
f1056affa Move select functions to host only to remove host/device suppressions
4f2156dfa Thrust detail::aligned_reinterpret_cast doesn't warn now
f4840618c Make sure ThrustPatches is included before thrust.
b2bbd66e6 Merge branch 'upstream-taotuple' into update_taoo
4ec6fc812 taotuple 2019-04-03 (8e70fa8a)
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Allison Vacanti <allison.vacanti@kitware.com>
Merge-request: !1607
When benchmarking the VTK-m algorithms on Summit I discovered
that our scheduling choices aren't optimal for the hardware.
This is a short term fix where we select good numbers for
Summit, and in the future make the defaults controllable
by the calling programming and/or environment variables.
Performance numbers can be found at:
https://gitlab.kitware.com/snippets/755
Previously, when Stop was called on a Cuda timer, it would record a stop
event but it would not synchronize it at that time. Instead, the
synchronize was only called when GetElapsedTime was called. The problem
is that the time of the event is only marked when synchronize is called.
Thus, if the event completed before GetElapsedTime was called, it would
record the time from when the event acutally happened to the time when
GetElapsedTime was called as part of the elapsed time, which is
incorrect.
Fix the problem by synchronizing when Stop is called. Although this
makes the Timer more invasive, generally using the Timer can cause
synchronization to happen. This behavior is consistent with the Timer
implementation for other devices.
It should be possible to query a vtkm::cont::Timer without modifying it.
As such, its query functions (such as Stopped and GetElapsedTime) should
be const.
The timer class now is asynchronous and device independent. it's using an
similiar API as vtkOpenGLRenderTimer with Start(), Stop(), Reset(), Ready(),
and GetElapsedTime() function. For convenience and backward compability, Each
Start() function call will call Reset() internally and each GetElapsedTime()
function call will call Stop() function if it hasn't been called yet for keeping
backward compatibility purpose.
Bascially it can be used in two modes:
* Create a Timer without any device info. vtkm::cont::Timer time;
* It would enable timers for all enabled devices on the machine. Users can get a
specific elapsed time by passing a device id into the GetElapsedtime function.
If no device is provided, it would pick the maximum of all timer results - the
logic behind this decision is that if cuda is disabled, openmp, serial and tbb
roughly give the same results; if cuda is enabled it's safe to return the
maximum elapsed time since users are more interested in the device execution
time rather than the kernal launch time. The Ready function can be handy here
to query the status of the timer.
* Create a Timer with a device id. vtkm::cont::Timer time((vtkm::cont::DeviceAdapterTagCuda()));
* It works as the old timer that times for a specific device id.
The kernel launch component of the runtime device adapter is fairly
pointless. If the hardware supports CUDA we should expect that
VTK-m has the correct kernel versions.
Plus in the original version if the CUDA device was being used
and the kernel launch returns cudaErrorDevicesUnavailable it
was never possible to restore CUDA support. Now what happens
is that the runtime tracker is marked as failed, but the
calling code can always go back and trying the device again.
This is only set while compiling device code, and is useful
for code that needs different implementations on devices (e.g.
they call CUDA device intrinsics, etc).
`vtkm::cont::testing` now initializes with logging enabled and support
for device being passed on the command line, `vtkm::testing` only
enables logging.
The purpose of the TestBuild infrastructure was to confirm that
VTK-m didn't have any lexical issues when it was a pure header
only project. As we now move to have more compiled components
the need for this form of testing is mitigated. Combined
with the issue of TestBuilds causing MSVC issues, we should
just remove this infrastructure.
This change allows you to set a subclass of
vtkm::cont::ExecutionObjectBase as a functor
used in ArrayHandleTransform. This latter class will then detect that
the functor is an ExecObject and will call PrepareForExecution with the
appropriate device to get the actual Functor object.
This change allows you to use virtual objects and other device dependent
objects as functors for ArrayHandleTransform without knowing a priori
what device the portal will be used on.
Rather than force all dispatchers to be templated on a device adapter,
instead use a TryExecute internally within the invoke to select a device
adapter.
Because this removes the need to declare a device when invoking a
worklet, this commit also removes the need to declare a device in
several other areas of the code.
e34301eca Allow disabling/enabling of CUDA managed memory via an env variable
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !1359
By setting the environment variable "VTKM_MANAGEDMEMO_DISABLED" to be 1,
users are able to disable CUDA managed memory even though the hardware is
capable of doing so.
Calls to 'cudaFree' block execution on all cuda devices. Reduce the number of
times this happens by having a deferred free mechanism that frees a pool
of pointers together when a threshold is reached.
Especially helpful during virtual object transfers that requires a few small
allocations and frees.
The original design of invoke and the transport infrastructure
relied on the implementation behavior of vtkm::cont types
such as ArrayHandle that used an internal shared_ptr to managed
state. This allowed passing by value instead of passing by
non-const ref when needing to transfer information to the device.
As VTK-m adds support for classes that use virtuals the ability
to pass by base pointer type allows for us to invoke worklets
using a base type without the risk of type slicing.
Additional by moving over to a non-const ref Invocation we
can update all transports that have 'output' to now be
by ref and therefore support types that can't be copied while
being 'more' correct.
