With the major revision 2.0 of VTK-m, many items previously marked as
deprecated were removed. If updating to a new version of VTK-m, it is
recommended to first update to VTK-m 1.9, which will include the deprecated
features but provide warnings (with the right compiler) that will point to
the replacement code. Once the deprecations have been fixed, updating to
2.0 should be smoother.
When managing portals in the execution environment, `ArrayHandle` uses the
`Token` object to ensure that the memory associated with a portal exists
for the length of time that it is needed. This is done by creating the
portal with a `Token` object, and the associated portal objects are
guaranteed to be valid while that `Token` object exists. This is supported
by essentially locking the array from further changes.
`Token` objects are typically used when creating a control-side portal with
the `ReadPortal` or `WritePortal`. This is not to say that a `Token` would
not be useful; a control-side portal going out of scope is definitely a
problem. But the creation and distruction of portals in the control
environment is generally too much work for the possible benefits.
However, under certain circumstances it could be useful to use a `Token` to
get a control-side portal. For example, if the `PrepareForExecution` method
of an `ExecutionObjectBase` needs to fill a small `ArrayHandle` on the
control side to pass to the execution side, it would be better to use the
provided `Token` object when doing so. This change allows you to optionally
provide that `Token` when creating these control-side portals.
Previously, the number of buffers held by an `ArrayHandle` had to be
determined statically at compile time by the storage. Most of the time
this is fine. However, there are some exceptions where the number of
buffers need to be selected at runtime. For example, the
`ArrayHandleRecombineVec` does not specify the number of components it
uses, and it needed a hack where it stored buffers in the metadata of
another buffer, which is bad.
This change allows the number of buffers to vary at runtime (at least at
construction). The buffers were already managed in a `std::vector`. It
now no longer forces the vector to be a specific size.
`GetNumberOfBuffers` was removed from the `Storage`. Instead, if the
number of buffers was not specified at construction, an allocation of
size 0 is done to create default buffers.
The biggest change is to the interface of the storage object methods,
which now take `std::vector` instead of pointers to `Buffer` objects.
This adds a little hastle in having to copy subsets of this `vector`
when a storage object has multiple sub-arrays. But it does simplify some
of the templating.
Whan an `UnknownArrayHandler` is constructed from an `ArrayHandle`, it uses
the `VecTraits` of the component type to construct its internal functions.
This meant that you could not put an `ArrayHandle` with a component type
that did not have `VecTraits` into an `UnknownArrayHandle`.
`UnknownArrayHandle` now no longer needs the components of its arrays to
have `VecTraits`. If the component type of the array does not have
`VecTraits`, it treats the components as if they are a scalar type.
Previously, the `Allocate` method of `ArrayHandle` was _not_ declared as
`const`. Likewise, the methods that depended on `Allocate`, namely
`ReleaseResources` and `PrepareForOutput` were also not declared `const`.
The main consequence of this was that if an `ArrayHandle` were passed as a
constant reference argument to a method (e.g. `const ArrayHandle<T>& arg`),
then the array could not be reallocated.
This seems right at first blush. However, we have changed these methods to
be `const` so that you can in fact reallocate the `ArrayHandle`. This is
because the `ArrayHandle` is in principle a pointer to an array pointer.
Such a structure in C will allow you to change the pointer to the array,
and so in this context it makes sense for `ArrayHandle` to support that as
well.
Although this distinction will certainly be confusing to users, we think
this change is correct for a variety of reasons.
1. This change makes the behavior of `ArrayHandle` consistent with the
behavior of `UnknownArrayHandle`. The latter needed this behavior to
allow `ArrayHandle`s to be passed as output arguments to methods that
get automatically converted to `UnknownArrayHandle`.
2. Before this change, a `const ArrayHandle&` was still multible is many
way. In particular, it was possible to change the data in the array
even if the array could not be resized. You could still call things
like `WritePortal` and `PrepareForInOut`. The fact that you could
change it for some things and not others was confusing. The fact that
you could call `PrepareForInOut` but not `PrepareForOutput` was doubly
confusing.
3. Passing a value by constant reference should be the same, from the
calling code's perspective, as passing by value. Although the function
can change an argument passed by value, that change is not propogated
back to the calling code. However, in the case of `ArrayHandle`,
calling by value would allow the array to be reallocated from the
calling side whereas a constant reference would prevent that. This
change makes the two behaviors consistent.
4. The supposed assurance that the `ArrayHandle` would not be reallocated
was easy to break even accidentally. If the `ArrayHandle` was assigned
to another `ArrayHandle` (for example as a class' member or wrapped
inside of an `UnknownArrayHandle`), then the array was free to be
reallocated.
This method is a remenant of when `ArrayHandle` could only store data on
one device at a time. It is now capable of storing data on any number of
devices (as well as the host), so asking for "the" device no longer
makes sense. Thus, this method is deprecated in favor of
`ArrayHandle::IsOnDevice`.
This deprecation leads to fixing some older functionality that still
assumed data could only be on one device at a time.
Fixes#592.
The newer version of `ArrayHandle` no longer supports different types of
portals for different devices. Thus, the `ReadPortalType` and
`WritePortalType` are sufficient for all types of portals across all
devices.
This significantly simplifies supporting execution objects on devices,
and thus this change also includes many changes to various execution
objects to remove their dependence on the device adapter tag.
cecd81d5d Add types appropriate for Ascent
865855ea0 Add changelog for making ArrayHandleSOA a default array
50ff9c22a Add support of `ArrayHandleSOA` as a default storage type
bc09a9cd1 Add precompiled versions of `ArrayRangeCompute` for `ArrayHandleSOA`
77f9ae653 Support `ArrayHandleSOA` only for `Vec` value types
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !2349
What was previously declared as `ArrayHandleNewStyle` is now just the
implementation of `ArrayHandle`. The old implementation of `ArrayHandle`
has been moved to `ArrayHandleDeprecated`, and `ArrayHandle`s still
using this implementation must declare `VTKM_ARRAY_HANDLE_DEPRECATED` to
use it.
Previously, `ArrayHandleSOA` worked with any value type that supported
`VecTraits`. That means that `ArrayHandleSOA` worked with scalar types
like `Float32`. However, for scalar types, the behavior is essentially
the same as `ArrayHandleBasic`, but with lots of extra templating and
code generation.
Although there is nothing _wrong_ with allowing `ArrayHandleSOA` holding
a scalar, there is no real reason to either (other than likely template
convenience). Generally, there is nothing wrong with supporting it, but
if you want to support `ArrayHandleSOA` in places where types are not
known (e.g. `Field`), then templating tends to iterate over the cross of
all supported types with all supported storage. That means such code
will automatically generate a bunch of code for `ArrayHandleSOA` with
scalars even if there is no reason for those code paths.
So, we can just disable the use of `ArrayHandleSOA` with scalars to
allow us to use `ArrayHandleSOA` as a default storage without creating
all these useless code paths.
The new-style `ArrayHandle` uses `Buffer` objects to manage data. Thus,
when one is decorating the other, it expects to find the `Buffer`
objects, which the old-style `ArrayHandle`s do not have. To make the two
work together, fake buffers in the old-style arrays.
The buffers in old-style arrays are empty, but have metadata that points
back to the `ArrayHandle.
It should be noted that a break in backward compatibility is introduced.
The implementation class passed to `ArrayHandleDecorator` changed the
specification of `AllocateSourceArrays`. Thus, decorators that were
previously allocatable now no longer will be until that method is
updated to the new form.
Most `ArrayHandle` implementations must create an
`std::vector<vtkm::cont::internal::Buffer>` object. It is most helpful
to create it in the constructor when initializing the superclass (so the
superclass does not have to create it's own). Added a `CreateBuffers`
convenience function to make it easy to build these vectors.
This portal only works on the control environment, which means it cannot
work with the new `ArrayHandle` type. Recent changes to
`ArrayHandleMultiplexer` also do not allow this, so just remove it
rather than try to fix it.
The `ReleaseResourcesExecution` method makes changes by booting data off
of the execution environment. But logically the array does not change.
It remains the same size with the same contents. Thus, it makes sense
for this to be a const method.
Also modified some deprecated methods a bit to remove some unnecessary
templates.
The new method `ArrayHandle::DeepCopy` had the pattern such that the
data in the `this` array was pushed to the provided destination array.
However, this is the opposite pattern used in the equivalent method in
VTK, which takes the data from the provided array and copies it to
`this` array.
So, swap the pattern to match that of VTK. Also, make the method name
more descriptive by renaming it to `DeepCopyFrom`. Hopefully, users will
read that to mean given the `ArrayHandle`, you copy data from the other
provided `ArrayHandle`.
A recent modification to `ArrayCopy` created a fast path for when
copying arrays of the same type. This fast path just deep copies the
buffers. The issue was that the buffer copy was creating new buffers
instead of updating the existing buffers. The passed in `ArrayHandle`
would get updated to the new buffers, but any other `ArrayHandle`s
sharing those buffers would still have the old versions. That would lead
to unexpected errors in code like this.
```cpp
vtkm::cont::ArrayHandle<T> outArray1;
vtkm::cont::ArrayHandle<T> outArray2 = outArray1;
vtkm::cont::ArrayCopy(inArray, outArray2);
```
If `inArray` was a different type than `outArray2`, then the data for
both `outArray1` and `outArray2` would be updated (which is the expected
behavior for something considered a "pointer"). However, if `inArray`
was the same type as `outArray2`, then the fast path would change
`outArray2` but leave `outArray1` the same.
This behavior has been corrected so that, in this case, the data of
`outArray1` always follows that of `outArray2`.
The old style `ArrayHandle` stored most of its state, including the
data, in the `vtkm::cont::internal::Storage` object (templated to the
type of array). The new style of `ArrayHandle` stores the data itself in
`Buffer` objects, and recent changes to `Buffer` allow metadata to be
stored there, too.
These changes make it pretty unnecessary to hold any state at all in the
`Storage` object. This is good since the sharing of state from one type
of `ArrayHandle` to another (such as by transforming the data), can be
done by just sharing the `Buffer` objects.
To reinforce this behavior, the `Storage` object has been changed to
make it completely stateless. All the methods of `Storage` must be
marked as `static`.
While in the transition between two types of `ArrayHandle`
implementations, we need to declare when an `ArrayHandle` is implemented
with the new style. To consolidate, create a
`VTKM_ARRAY_HANDLE_NEW_STYLE` to override the default `ArrayHandle`
implementation with the `ArrayHandleNewStyle` implementation.
`ArrayHandle::DeepCopy` creates a new `ArrayHandle` of the same type and
deep copies the data into it.
This functionality is similar to `ArrayCopy`. However, it can be used
without having to compile for the device on which the copy happens.
This has no real change in the operation, but it will simplify code as
we convert `ArrayHandle`s to the new type. We will be able to write
simple runtime code rather than complex metaprogramming to determine the
number of buffers to use.
We are moving to deprecate `ArrayHandleVirtual`, so we are removing the
feature where `VariantArrayHandle::CastAndCall` automatically casts to
an `ArrayHandleVirtual` if possible.
The big reason to make this change now (as opposed to later when
`ArrayHandleVirtual` is deprecated) is to improve compile times.
This prevents us from having to compile an extra code path using
`ArrayHandleVirtual`.
`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.
This makes it clear that it returns true for an invalid array handle.
The previous name implied that it was looking for an ArrayHandle in some
"valid" set, which is the opposite.
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.
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.
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.
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.
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.