The namespaces need to be different for each test, or else only the first
implementation of the function will be used for all tests that call that
function.
Also updated the test to verify that we can count starting from a non zero
number.
NVCC is unable to handle finding the worklets when they are in an anonymous
namespace. It only looks at the the anonymous namespaces included by the
files that device code uses, and misses our anon namespace. Moving to a named
namespace solves these issues.
They were declared as in both control and execution, but this would
cause problems when the FunctionInterface contained objects that could
only be copied in the control environment. Using these methods probably
only makes sense in the control environment anyway. (They are a bit
heavyweight to use in an inner loop of the execution environment.)
The functors in the ForEach, StaticTransform, and DynamicTransform
methods sometimes can use the index of the parameter that they are
operating on. This can be a helpful diagnostic in compile and run-time
errors. It is also helpful when linking parameters from one
FunctionInterface with those of another.
This new features are now replacing implementations using the Zip
functionality that was removed earlier. The implementation is actually
simplified a bit.
The Zip function does not work when compiling with CUDA because it
forces the parameters to be in both the control and execution (host and
device) environments.
Porting the dax device adapter over to vtkm. Unlike the dax version, doesn't
use the thrust::device_vector, but instead uses thrust::system calls so that
we can support multiple thrust based backends.
Also this has Texture Memory support for input array handles. Some more work
will need to be done to ArrayHandle so that everything works when using an
ArrayHandle inplace with texture memory bindings.
When fixing a problem where the disabled test had left some unused
classes, which some compilers picked up on (SHA
eae8921dc714c9bdb12058db365f890425291ea2), I used a preprocessor wrapper
to enable/disable the code (mostly to preserve history). However, I
forgot to leave the code disabled. Disable that here.
A couple of tests were failing with the Intel compiler due to
imprecision in comparing floating point values.
Also snuck in some minor documentation fixes in a comment for
FunctionInterface.
There is a test that tries to determine that the Invoke methods in
FunctionInterface do not add an unreasonable overhead. However, this
test is unreliable. Also, the most critical performance hit would be in
invoking a worklet operation, but that is now done elsewhere anyway.
The unit test for StorageBasic tested the StealArray feature and then
used the delete[] operator on the stolen array to deallocate it. For
many standard libraries the default implementation for delete[] is
the same as (or at least compatible with) std::allocator, but for
the PGI compiler they were not compatible and this resulted in a
run-time error. This change fixes the problem with the test by using
the same allocator as the StorageBasic test.
ICC can be pretty thorough about finding unused elements. In this case
it was picking up an unused method in instances of a templated class
in an anonymous namespace. It was a method that should be there due to
the nature of the class, but it happened to not be used (which was OK,
too). To get around the problem, I just added some use of that method
in another method.
It's easy to put accidently put something that is not a valid tag in a
ControlSignature or ExecutionSignature. Previously, when you did that
you got a weird error at the end of a very long template instantiation
chain that made it difficult to find the offending worklet.
This adds some type checks when the dispatcher is instantated to check
the signatures. It doesn't point directly to the signature or its
parameter, but it is much closer.
MSVC is picky about type conversions. To get it to shut up, explicitly
cast the worklet return value to the fetch value in the
WorkletInvokeFunctor. The good is that it will help with needing
explicit conversions on these return values. But that is also bad in
that it might make some unexpected conversions possible.
Instead of just checking that a dispatcher's Invoke input is an
ArrayHandle, also check that the ValueType of the ArrayHandle is
compatible with the types of the worklet operator. This is done by
adding a template argument to the ControlSignature tags that is a type
list tag that gets passed to the type check.
One fix is a simple (pointless) compiler warning about precision. The
other fix is an error in one of the test codes that did not clear out
the message string in an error message buffer like it was supposed to.
Also reduced the maximum list size to 15 (which is the current longest
single list we have). Trying to reduce the size of the generated code a
bit, which is getting a little long.
This is a simple version of a dispatcher, but an important one.
Note that there is an issue brought up with UnitTestWorkletMapField in
that there needs to be better ways to specify worklet argument types.
These changes support the implementation of DispatcherBase. This class
provides the basic functionality for calling an Invoke method in the
control environment, transferring data to the execution environment,
scheduling threads in the execution environment, pulling data for each
calling of the worklet method, and actually calling the worklet.
The zip capability allows you to parameter-wise combine two
FunctionInterface objects. The result is another FunctionInterface with
each parameter a Pair containing the respective values of the two
inputs.
Being able to zip allows you to do transforms and invokes on data that
is divided among multiple function interface objects.
The Fetch class is responsible for moving data in and out of some
collection in the execution environment. The Fetch class is templated
with a pair of tags (the type of fetch and the aspect) that control the
mechanism used for the fetch.
The Transport class is responsible for moving data from the control
environment to the execution environment. (Actually, it might be more
accurate to say it gets the execution environment associated with a
given control object.) The Transport class is templated with a tag that
controls the mechanism used for the transport.
Lots of tests have to move values in and out of arrays and check them
against expected values. It is also often the case that these tests are
run on lots of different types. There is some repeated code for
generating known values for particular indices. This change unifies some
of that. This can probably also encourage making more generic tests.
The previous commits had TypeListTagAll containing a subset of Vec
types. This commit adds all possible vectors with 2 to 4 components
containing one of the basic C types.
Providing these types tends to "lock in" the precision of the algorithms
used in VTK-m. Since we are using templating anyway, our templates
should be generic enough to handle difference precision in the data.
Usually the appropriate type can be determined by the data provided. In
the case where there is no hint on the precision of data to use (for
example, in the code that provides coordinates for uniform data), there
is a vtkm::FloatDefault.
Before we assumed that we would only use the basic types specified by
the widths of vtkm::Scalar and vtkm::Id. We want to expand this to make
sure the code works on whatever data precision we need.
Since we want our code to generally handle data of different precision
(for example either float or double) expand the types in our list types
to include multiple precision.
Previously we just hand coded base lists up to 4 entries, which was fine
for what we were using it for. However, now that we want to support base
types of different sizes, we are going to need much longer lists.
There are multiple reasons for this name change:
* The name Tuple conflicts with the boost::Tuple class, which as a
different interface and feature set. This gets confusing, especially
since VTK-m uses boost quite a bit.
* The use of this class is usually (although not always) as a
mathematical vector.
* The vtkm::Scalar and vtkm::Vector* classes are going to go away soon
to better support multiple base data type widths. Having this
abbriviated name will hopefully make the code a bit nicer when these
types have to be explicitly specified.
Also modified the implementation a bit to consolidate some of the code.
In preparation for supporting base types with more widths, add typedefs
for the base types with explicit widths (number of bits).
Also added a IdComponent type that should be used for indices for
components into tuples and vectors. There now should be no reason to use
"int" inside of VTK-m code (especially for indexing). This change cleans
up many of the int types that were used throughout.
We have a test for FunctionInterface to make sure that calling a function
indirectly through that is about as fast as directly. On MSVC we sometimes
observe that this timing fails in debug mode. This is probably the compiler
adding some code to each function invocation. That won't happen in production
compiles, so we don't care about it too much. Make an exception in this case.
When compiling with 32-bit Ids for a 64 bit machine (which is not
uncommon), it is possible that the distance between two iterators
is larger than the maximum value that can be stored in vtkm::Id.
If two such iterators were passed to ArrayPortalFromIterators, that
would cause problems.
This change checks for that condition and throws an out of memory
exception if it occurs. That would be a pretty darn big array and
is more likely to be the cause of an error somewhere else in the
code, but either way the check and error is good. This change also
fixes a warning we have been getting with MSVC.
For MSVC we use the non-portable wrapper stdext::checked_array_iterators
because the compiler insists on it for safety. When we check to make sure
our templates are giving us raw pointers, we have to check for this wrapper
instead of the raw pointer itself.
This moves the ability to get an iterator from an array portal out of
the portal itself. The next step is to move the GetIteratorBegin/End out
of ArrayPortal. This should make the implemenation a bit cleaner.
MSVC likes to warn about using raw pointers as iterators in generic
algorithms because they have been known to lead to problems. When
compiling with that compiler, wrap raw pointers in
stdext::checked_array_pointer to suppress the error and also add a bit
more checking.
I wanted to test ArrayHandleCounting with some non-standard data type.
I was using a class that looked like a number that counts by two, but
the operator behavior was not a proper group and that was causing issues.
Replaced that with a class that inefficiently represents an unsigned
integer as a string with that many characters. The inefficiency does not
matter because it is just a test.
It appears that when the Intel compiler is optimizing, constant floating
point values can be slightly different than the same value stored in memory
and never changed. This change uses the test_equal method to compare
these floating point values that might have a slight numeric error.
Although we cannot expect every developer to have pyexpander, for those
that do the build will automatically run it and check the expanded file
in the source code. If they match, a descriptive error is given.
We don't automatically update the file because subtle problems might
occur. It is better to alert a developer to fix the problem properly.
This commit removes the usage of the boost preprocessor library to
iteratively generate templates with a variable number of parameters. It
is replaced with a template that is expanded by running it through the
pyexpander macro processing tool (http://pyexpander.sourceforge.net).
One reason for this change is to make the code easier to read. In
particular, it is difficult to understand compiler errors when they
occur deep within an iterating macro. Another reason for this change is
that the Intel compiler currently has a bug that breaks with the boost
preprocessor library.
One issue with this approach is that the macro expansion is not part of
the build process. Although open, pyexpander is not a tool most
developers will have readily installed on their system. Thus, if you
want to make changes to any of the macro code, you have to make sure
pyexpander is installed, then make changes to the input files, then
manually run pyexpander from the command line.
After a talk with Robert Maynard, we decided to change the name
ArrayContainerControl to Storage. There are several reasons for this
change.
1. The name ArrayContainerControl is unwieldy. It is long, hard for
humans to parse, and makes for long lines and wraparound. It is also
hard to distinguish from other names like ArrayHandleFoo and
ArrayExecutionManager.
2. The word container is getting overloaded. For example, there is a
SimplePolymorphicContainer. Container is being used for an object that
literally acts like a container for data. This class really manages
data.
3. The data does not necessarily have to be on the control side.
Implicit containers store the data nowhere. Derivative containers might
have all the real data on the execution side. It is possible in the
future to have storage on the execution environment instead of the
control (think interfacing with a simulator on the GPU).
Storage is not a perfect word (what does implicit storage really mean?),
but its the best English word we came up with.
It was originally put there to support CopyInto in ArrayHandle, but that
has already been removed. It really only makes sense for trivial
examples and testing code, and it sometimes causes complications with
coding.
There is a special version of the testing methods for use in the control
environment that handles execeptions that can be thrown there. There are
tests to make sure you are using the correct version of the testing
framework, but it was broken until the last commit. Now that it's fixed,
here are two places where the wrong testing method was used.
We made this change a while ago to help with completion in IDEs.
(Completion was matching a bunch of wrapper macros that were almost
never used anywhere.) Most of the changes are in comments, but there are
a few bad macro definitions.
Whenever creating a functor to be launched in the execution environment
using the device adapter Schedule algorithm, you had to also create a
couple of methods to handle error message buffers. For convenience, lots
of code started to just inherit from WorkletBase. Although this worked,
it was a misnomer (and might cause problems in the future if worklets
later require different things from its base). To get around this
problem, add a FunctorBase class that is intended to be used as the
superclass to functors called with Schedule.
Previously, VTKM_DEVICE_ADAPTER_UNDEFINED and
VTKM_ARRAY_CONTAINER_CONTROL_UNDEFINED were set to 0. The problem is
that if someone set VTKM_DEVICE_ADAPTER or VTKM_ARRAY_CONTAINER_CONTROL
to something invalid, that would test positive when compared to 0. Thus,
you get an error about not defining the default when in fact the problem
is setting an invalid flag.
This change makes the undefined constants -1 so that the comparison will
fail unless the macro is actually properly set.
Getting the type right for ArrayHandleCompositeVector can be a bit
tricky. It is expressed in a somewhat strange function signature format
and you have to extract the right component type for the return. This
adds an ArrayHandleCompositeVectorType that makes it easier (although no
less verbose).
Using alignment on basic types when vtkm only targetted Linux/BSD/OSX was
'okay' because of how the alignment operators worked, but potential was going
to cause issues in the long run if we failed to detect the correct size and the
compiler was than forced to not use intrinsics.
Now with adding windows support we have run into another problem. Basically
using an alignment operator on a typedef means that the type must never
be passed by value, but must always be passed by reference. The reason for
this is that passing by value doesn't respect alignment requirements, and
can cause very subtle errors or crashes.
A really good read for people more interested in these problems:
http://eigen.tuxfamily.org/dox/group__TopicPassingByValue.htmlhttp://eigen.tuxfamily.org/dox-devel/group__DenseMatrixManipulation__Alignement.html
Each type of point coordinates has its own class with the name
PointCoordinates*. Currently there is a PointCoordiantesArray that contains
an ArrayHandle holding the point coordinates and a PointCoordinatesUniform
that takes the standard extent, origin, and spacing for a uniform rectilinear
grid and defines point coordiantes for that. Creating new PointCoordinates
arrays is pretty easy, and we will almost definitely add more. For example,
we should have an elevation version that takes uniform coordinates for
a 2D grid and then an elevation in the third dimension. We can probably
also use a basic composite point coordinates that can build them from
other coordinates.
There is also a DynamicPointCoordinates class that polymorphically stores
an instance of a PointCoordinates class. It has a CastAndCall method that
behaves like DynamicArrayHandle; it can call a functor with an array handle
(possible implicit) that holds the point coordinates.
This derived array handle creates an array of vectors whose components come
from other arrays of vectors. In either case ArrayHandleCompositeVector
handles scalars as vectors of size 1.
This will allow a faster conversion than the dynamic transform and will
allow you to define compile-time types for transformation unlike dynamic
transform or invoke with transform.
This is used with the FunctionInterface::DynamicTransformCont method to
convert a call of arguments using dynamic array handles to a function
templated on concrete types.
The FunctionInterface class is a convienient way to wrap up a variable
number of arguments and pass them around templated interfaces without
requiring variadic template arguments. It also correctly hands return
arguments.
Use this mechanism in the dynamic array handle to skip over trying
invalid array handle types (and thereby incurring a compiler error even
though we never intended to use these classes).
The dynamic array handle holds a reference to an array handle of an
unknown type. It contains the ability to try to cast it to an instance
of array handle or to try lists of types and containers.
There is currently an issue that is causing the test code not to
compile. It is the case that some combinations of types and containers
are not compatible. For example, an implict container is bound to a
certain type, and the container is undefined if they do not agree. There
needs to be a mechanism to detect these invalid combinations and skip
over them in the MTP for each.
Provies a list of types in a template like boost::mpl::vector and a
method to call a functor on each type. However, rather than explicitly
list each type, uses tags to identify the list. This provides the
following main advantages:
1. Can use these type lists without creating horrendously long class
names based on them, making compiler errors easier to read. For example,
you would have a typename like MyClass<TypeListTagVectors> instead of
MyClass<TypeList<Id3,Vector2,Vector3,Vector4> > (or worse if variadic
templates are not supported). This is the main motivation for this
implementation.
2. Do not require variadic templates and usually few constructions. That
should speed compile times.
There is one main disadvantage to this approach: It is difficult to get
a printed list of items in a list during an error. If necessary, it
probably would not be too hard to make a template to convert a tag to a
boost mpl vector.
