Defines TaggableCLArray

arraycontext/context.py

@@ -328,9 +328,10 @@ def einsum(self, spec, *args, arg_names=None, tagged=()):
             arg_names = tuple("arg%d" % i for i in range(len(args)))
 
         prg = self._get_einsum_prg(spec, arg_names, tagged)
-        return self.call_loopy(
+        out_ary = self.call_loopy(
             prg, **{arg_names[i]: arg for i, arg in enumerate(args)}
         )["out"]
+        return self.tag(tagged, out_ary)
 
     @abstractmethod
     def clone(self):

arraycontext/impl/pyopencl/__init__.py

@@ -1,6 +1,7 @@
 """
 .. currentmodule:: arraycontext
 .. autoclass:: PyOpenCLArrayContext
+.. automodule:: arraycontext.impl.pyopencl.taggable_cl_array
 """
 
 __copyright__ = """
@@ -35,6 +36,7 @@
 from pytools.tag import Tag
 
 from arraycontext.context import ArrayContext, _ScalarLike
+from arraycontext.container.traversal import rec_map_array_container
 
 
 if TYPE_CHECKING:
@@ -65,6 +67,8 @@ class PyOpenCLArrayContext(ArrayContext):
         of arrays are created (e.g. as results of computation), the associated cost
         may become significant. Using e.g. :class:`pyopencl.tools.MemoryPool`
         as the allocator can help avoid this cost.
+
+    .. automethod:: transform_loopy_program
     """
 
     def __init__(self,
@@ -109,7 +113,7 @@ def __init__(self,
                     DeprecationWarning, stacklevel=2)
 
         import pyopencl as cl
-        import pyopencl.array as cla
+        import pyopencl.array as cl_array
 
         super().__init__()
         self.context = queue.context
@@ -138,7 +142,9 @@ def __init__(self,
         self._loopy_transform_cache: \
                 Dict["lp.TranslationUnit", "lp.TranslationUnit"] = {}
 
-        self.array_types = (cla.Array,)
+        # TODO: Ideally this should only be `(TaggableCLArray,)`, but
+        # that would break the logic in the downstream users.
+        self.array_types = (cl_array.Array,)
 
     def _get_fake_numpy_namespace(self):
         from arraycontext.impl.pyopencl.fake_numpy import PyOpenCLFakeNumpyNamespace
@@ -147,18 +153,27 @@ def _get_fake_numpy_namespace(self):
     # {{{ ArrayContext interface
 
     def empty(self, shape, dtype):
-        import pyopencl.array as cl_array
-        return cl_array.empty(self.queue, shape=shape, dtype=dtype,
-                allocator=self.allocator)
+        from arraycontext.impl.pyopencl.taggable_cl_array import TaggableCLArray
+
+        return TaggableCLArray(self.queue, shape=shape, dtype=dtype,
+                               allocator=self.allocator)
 
     def zeros(self, shape, dtype):
         import pyopencl.array as cl_array
-        return cl_array.zeros(self.queue, shape=shape, dtype=dtype,
-                allocator=self.allocator)
+        from arraycontext.impl.pyopencl.taggable_cl_array import to_tagged_cl_array
+        return to_tagged_cl_array(cl_array.zeros(self.queue, shape=shape,
+                                                 dtype=dtype,
+                                                 allocator=self.allocator),
+                                  axes=None, tags=frozenset())
 
     def from_numpy(self, array: Union[np.ndarray, _ScalarLike]):
         import pyopencl.array as cl_array
-        return cl_array.to_device(self.queue, array, allocator=self.allocator)
+        from arraycontext.impl.pyopencl.taggable_cl_array import to_tagged_cl_array
+        return to_tagged_cl_array(cl_array
+                                  .to_device(self.queue,
+                                             array,
+                                             allocator=self.allocator),
+                                  axes=None, tags=frozenset())
 
     def to_numpy(self, array):
         if np.isscalar(array):
@@ -186,14 +201,33 @@ def call_loopy(self, t_unit, **kwargs):
             if len(wait_event_queue) > self._wait_event_queue_length:
                 wait_event_queue.pop(0).wait()
 
-        return result
+        from arraycontext.impl.pyopencl.taggable_cl_array import to_tagged_cl_array
+        # FIXME: Inherit loopy tags for these arrays
+        return {name: to_tagged_cl_array(ary, axes=None, tags=frozenset())
+                for name, ary in result.items()}
 
     def freeze(self, array):
         array.finish()
         return array.with_queue(None)
 
     def thaw(self, array):
-        return array.with_queue(self.queue)
+        from arraycontext.impl.pyopencl.taggable_cl_array import (TaggableCLArray,
+                                                                  to_tagged_cl_array)
+        import pyopencl.array as cl_array
+
+        if isinstance(array, TaggableCLArray):
+            return array.with_queue(self.queue)
+        elif isinstance(array, cl_array.Array):
+            from warnings import warn
+            warn("Invoking PyOpenCLArrayContext.thaw with pyopencl.Array"
+                 " will be unsupported in 2023. Use `to_tagged_cl_array`"
+                 " to convert instances of pyopencl.Array to TaggableCLArray.",
+                 DeprecationWarning, stacklevel=2)
+            return (to_tagged_cl_array(array, axes=None, tags=frozenset())
+                    .with_queue(self.queue))
+        else:
+            raise ValueError("array should be a cl.array.Array,"
+                             f" got '{type(array)}'")
 
     # }}}
 
@@ -268,12 +302,37 @@ def transform_loopy_program(self, t_unit):
         return t_unit
 
     def tag(self, tags: Union[Sequence[Tag], Tag], array):
-        # Sorry, not capable.
-        return array
+        import pyopencl.array as cl_array
+        from arraycontext.impl.pyopencl.taggable_cl_array import (TaggableCLArray,
+                                                                  to_tagged_cl_array)
+
+        def _rec_tagged(ary):
+            if isinstance(ary, TaggableCLArray):
+                return ary.tagged(tags)
+            elif isinstance(ary, cl_array.Array):
+                return to_tagged_cl_array(ary, axes=None,  tags=tags)
+            else:
+                raise ValueError("array should be a cl.array.Array,"
+                                 f" got '{type(ary)}'")
+
+        return rec_map_array_container(_rec_tagged, array)
 
     def tag_axis(self, iaxis, tags: Union[Sequence[Tag], Tag], array):
-        # Sorry, not capable.
-        return array
+        import pyopencl.array as cl_array
+        from arraycontext.impl.pyopencl.taggable_cl_array import (TaggableCLArray,
+                                                                  to_tagged_cl_array)
+
+        def _rec_tagged(ary):
+            if isinstance(ary, TaggableCLArray):
+                return ary.with_tagged_axis(iaxis, tags)
+            elif isinstance(ary, cl_array.Array):
+                return (to_tagged_cl_array(ary, axes=None,  tags=tags)
+                        .with_tagged_axis(iaxis, tags))
+            else:
+                raise ValueError("array should be a cl.array.Array,"
+                                 f" got '{type(ary)}'")
+
+        return rec_map_array_container(_rec_tagged, array)
 
     def clone(self):
         return type(self)(self.queue, self.allocator,

arraycontext/impl/pyopencl/taggable_cl_array.py

@@ -0,0 +1,129 @@
+"""
+.. autoclass:: TaggableCLArray
+.. autoclass:: Axis
+
+.. autofunction:: to_tagged_cl_array
+"""
+
+import pyopencl.array as cla
+from typing import FrozenSet, Union, Sequence, Optional, Tuple
+from pytools.tag import Taggable, Tag
+from dataclasses import dataclass
+from pytools import memoize
+
+
+@dataclass(frozen=True, eq=True)
+class Axis(Taggable):
+    """
+    Records the tags corresponding to a dimensions of :class:`TaggableCLArray`.
+    """
+    tags: FrozenSet[Tag]
+
+    def copy(self, **kwargs):
+        from dataclasses import replace
+        return replace(self, **kwargs)
+
+
+@memoize
+def _construct_untagged_axes(ndim: int) -> Tuple[Axis, ...]:
+    return tuple(Axis(frozenset()) for _ in range(ndim))
+
+
+class TaggableCLArray(cla.Array, Taggable):
+    """
+    A :class:`pyopencl.array.Array` with additional metadata. This is used by
+    :class:`~arraycontext.PytatoPyOpenCLArrayContext` to preserve tags for data
+    while frozen, and also in a similar capacity by
+    :class:`~arraycontext.PyOpenCLArrayContext`.
+
+    .. attribute:: axes
+
+       A :class:`tuple` of instances of :class:`Axis`, with one :class:`Axis`
+       for each dimension of the array.
+
+    .. attribute:: tags
+
+        A :class:`frozenset` of :class:`pytools.tag.Tag`. Typically intended to
+        record application-specific metadata to drive the optimizations in
+        :meth:`arraycontext.PyOpenCLArrayContext.transform_loopy_program`.
+    """
+    def __init__(self, cq, shape, dtype, order="C", allocator=None,
+                 data=None, offset=0, strides=None, events=None, _flags=None,
+                 _fast=False, _size=None, _context=None, _queue=None,
+                 axes=None, tags=frozenset()):
+
+        super().__init__(cq=cq, shape=shape, dtype=dtype,
+                         order=order, allocator=allocator,
+                         data=data, offset=offset,
+                         strides=strides, events=events,
+                         _flags=_flags, _fast=_fast,
+                         _size=_size, _context=_context,
+                         _queue=_queue)
+
+        self.tags = tags
+        axes = axes if axes is not None else _construct_untagged_axes(len(self
+                                                                          .shape))
+        self.axes = axes
+
+    def copy(self, queue=cla._copy_queue, tags=None, axes=None, _new_class=None):
+        """
+        :arg _new_class: The class of the copy. :func:`to_tagged_cl_array` is
+            sets this to convert instances of :class:`pyopencl.array.Array` to
+            :class:`TaggableCLArray`. If not provided, defaults to
+            ``self.__class__``.
+        """
+        _new_class = self.__class__ if _new_class is None else _new_class
+
+        if queue is not cla._copy_queue:
+            # Copying command queue is an involved operation, use super-class'
+            # implementation.
+            base_instance = super().copy(queue=queue)
+        else:
+            base_instance = self
+
+        if tags is None and axes is None and _new_class is self.__class__:
+            # early exit
+            return base_instance
+
+        tags = getattr(base_instance, "tags", frozenset()) if tags is None else tags
+        axes = getattr(base_instance, "axes", None) if axes is None else axes
+
+        return _new_class(None,
+                          base_instance.shape,
+                          base_instance.dtype,
+                          allocator=base_instance.allocator,
+                          strides=base_instance.strides,
+                          data=base_instance.base_data,
+                          offset=base_instance.offset,
+                          events=base_instance.events, _fast=True,
+                          _context=base_instance.context,
+                          _queue=base_instance.queue,
+                          _size=base_instance.size,
+                          tags=tags,
+                          axes=axes,
+                          )
+
+    def with_tagged_axis(self, iaxis: int,
+                         tags: Union[Sequence[Tag], Tag]) -> "TaggableCLArray":
+        """
+        Returns a copy of *self* with *iaxis*-th axis tagged with *tags*.
+        """
+        new_axes = (self.axes[:iaxis]
+                    + (self.axes[iaxis].tagged(tags),)
+                    + self.axes[iaxis+1:])
+        return self.copy(axes=new_axes)
+
+
+def to_tagged_cl_array(ary: cla.Array,
+                       axes: Optional[Tuple[Axis, ...]],
+                       tags: FrozenSet[Tag]) -> TaggableCLArray:
+    """
+    Returns a :class:`TaggableCLArray` that is constructed from the data in
+    *ary* along with the metadata from *axes* and *tags*.
+
+    :arg axes: An instance of :class:`Axis` for each dimension of the
+        array. If passed *None*, then initialized to a :class:`pytato.Axis`
+        with no tags attached for each dimension.
+    """
+    return TaggableCLArray.copy(ary, axes=axes, tags=tags,
+                                _new_class=TaggableCLArray)