Merge pull request #2955 from devitocodes/mpi-redistribution-engine

MPI: revamp mpi indexing for full arbitrary indexing support

Author: mloubout

.github/workflows/examples-mpi.yaml

@@ -71,7 +71,7 @@ jobs:
         ipcluster start --profile=mpi --engines=mpi -n 4 --daemonize
         # A few seconds to ensure workers are ready
         sleep 10
-        pytest -v --nbval examples/mpi
+        pytest -vvv --nbval examples/mpi
         ipcluster stop --profile=mpi
 
     - name: Test seismic examples

devito/data/data.py

@@ -296,7 +296,14 @@ def index_glb_to_loc(self, *args, rel=True):
                         loc_min = loc_abs_min - base \
                             + np.mod(glb_idx.step - np.mod(base, glb_idx.step),
                                      glb_idx.step)
-                    else:  # glb start is given explicitly
+                    elif glb_idx_min >= loc_abs_min:
+                        # The slice starts within this subdomain, so its first
+                        # selected index is the start itself (the stride phase
+                        # begins here, not at loc_abs_min).
+                        loc_min = glb_idx_min - base
+                    else:  # the slice started in an earlier subdomain
+                        # Advance to the first in-stride index at or after this
+                        # subdomain's start.
                         loc_min = loc_abs_min - base \
                             + np.mod(glb_idx.step - np.mod(base - glb_idx.start,
                                                            glb_idx.step), glb_idx.step)
@@ -319,7 +326,14 @@ def index_glb_to_loc(self, *args, rel=True):
                         loc_max = top - base \
                             + np.mod(glb_idx.step - np.mod(top - glb_max,
                                                            glb_idx.step), glb_idx.step)
-                    else:
+                    elif glb_idx_max <= loc_abs_max:
+                        # The slice's high end (its start) is within this
+                        # subdomain, so the first selected index is the start
+                        # itself (the stride phase begins here, not at loc_abs_max).
+                        loc_max = glb_idx_max - base
+                    else:  # the slice's high end is in a later subdomain
+                        # Step back to the first in-stride index at or below this
+                        # subdomain's end.
                         loc_max = top - base \
                             + np.mod(glb_idx.step - np.mod(top - glb_idx.start,
                                                            glb_idx.step), glb_idx.step)
@@ -555,3 +569,50 @@ def reshape(self, *args):
         items = [i + nleft for i in items]
 
         return Decomposition(items, self.local)
+
+    def index_decomposition(self, idx):
+        """
+        The decomposition induced by NumPy-indexing this axis with a slice.
+
+        Each subdomain keeps the result positions, in index order, of the global
+        indices it owns. A strided or reversed slice therefore just relabels
+        which rank owns which result index -- indexing never moves data across
+        ranks. Unlike `reshape` (which adjusts subdomain *boundaries* and
+        is used for halos), this follows exact NumPy slicing semantics.
+
+        Examples
+        --------
+        >>> d = Decomposition([[0, 1, 2, 3], [4, 5, 6, 7]], 0)
+        >>> d.index_decomposition(slice(None, None, -1))
+        Decomposition(<<[4,7]>>, [0,3])
+        >>> d.index_decomposition(slice(None, None, -3))
+        Decomposition(<<[2,2]>>, [0,1])
+        """
+        if self.glb_max is None:
+            return Decomposition([np.array([], dtype=np.int64)]*len(self),
+                                 self.local)
+        size = self.glb_max - self.glb_min + 1
+
+        # The global indices selected by `idx`, in result order. For a slice
+        # (the only caller) this is built directly, without materialising the
+        # whole axis.
+        if isinstance(idx, slice):
+            start, stop, step = idx.indices(size)
+            selected = self.glb_min + np.arange(start, stop, step)
+        else:
+            selected = np.arange(self.glb_min, self.glb_min + size)[idx]
+
+        # Bucket the selected indices by owning subdomain. A subdomain is a
+        # contiguous range, so membership is a bounds check in O(len(selected))
+        # -- avoiding an O(axis) `isin` on every slice. A non-contiguous
+        # subdomain (not produced by gridding/slicing) falls back to `isin`.
+        items = []
+        for sd in self:
+            if sd.size == 0:
+                items.append(np.array([], dtype=np.int64))
+            elif sd[-1] - sd[0] + 1 == sd.size:
+                owned = (selected >= sd[0]) & (selected <= sd[-1])
+                items.append(np.nonzero(owned)[0])
+            else:
+                items.append(np.nonzero(np.isin(selected, sd))[0])
+        return Decomposition(items, self.local)

devito/data/decomposition.py

@@ -0,0 +1,22 @@
+"""
+Indexing engine for MPI-distributed arrays.
+
+Indexing an MPI-distributed array is treated as a redistribution between layouts.
+The engine is built as four pure layers plus a value object:
+
+* `selection` -- what an index expression means (serial NumPy semantics).
+* `layout`    -- where a global coordinate physically lives.
+* `plan`      -- the rank-to-rank routing, built without communication.
+* `transport` -- the sparse point-to-point exchange (NBX).
+* `exchange`  -- `Exchange(data, idx).get()/.put(value)`.
+
+Only `Exchange` is needed by `Data`; the lower layers are independently
+testable (in serial, or with toy buffers).
+"""
+
+from devito.data.distributed.exchange import Exchange, cached_exchange  # noqa
+from devito.data.distributed.layout import Layout  # noqa
+from devito.data.distributed.plan import ExchangePlan  # noqa
+from devito.data.distributed.redistribution import redistribute_set  # noqa
+from devito.data.distributed.selection import Selection  # noqa
+from devito.data.distributed.transport import sparse_exchange  # noqa

devito/data/distributed/__init__.py

@@ -0,0 +1,113 @@
+"""
+Exchange: the value object tying a Selection and a Layout into a reusable plan.
+
+`Exchange(data, idx)` builds the routing for one index expression without any
+communication. `get` pulls `data[idx]` and `put` assigns `data[idx] =
+value`. Because the plan is communication-free to build and independent of the
+data *values*, it can be cached and replayed across calls with the same index
+(e.g. sparse injection every timestep), so the steady state is pure
+point-to-point with no re-planning. `cached_exchange` provides that cache;
+`Data` uses it so `data[idx] = value` is automatically plan-cached.
+"""
+
+from functools import lru_cache
+
+import numpy as np
+
+from devito.data.distributed.layout import Layout
+from devito.data.distributed.plan import ExchangePlan
+from devito.data.distributed.selection import Selection
+from devito.tools import as_tuple
+
+__all__ = ['Exchange', 'cached_exchange']
+
+
+class Exchange:
+
+    """
+    A reusable redistribution plan for `data[idx]` on distributed `data`.
+
+    Parameters
+    ----------
+    data : Data
+        The MPI-distributed array being indexed.
+    idx : index expression
+        Any NumPy index (scalars, slices, integer arrays, boolean masks).
+    """
+
+    def __init__(self, data, idx):
+        # Keep the data reference (not a snapshot): the underlying buffer is
+        # stable across `f.data` views, so a cached plan always reads/writes the
+        # current values.
+        self._data = data
+        global_shape = tuple(
+            dec.size if dec is not None else size
+            for dec, size in zip(data._decomposition, data.shape, strict=True)
+        )
+        self._layout = Layout(data._distributor, data._decomposition, global_shape)
+        self._selection = Selection.from_index(idx, global_shape)
+        self._plan = ExchangePlan.build(self._selection, self._layout)
+
+    def get(self):
+        """Return `data[idx]` as a NumPy array."""
+        return self._plan.get(np.asarray(self._data))
+
+    def put(self, value):
+        """Assign `data[idx] = value`."""
+        self._plan.put(np.asarray(self._data), value)
+
+
+class _ExchangeKey:
+
+    """
+    Hashable, content-addressed key wrapping `(data, idx)` for plan caching.
+
+    Generates a signature from the `data` identity and the `idx` content, so that
+    the same plan is reused across calls with the same index expression, even if
+    the `data` object is a different view of the same underlying buffer.
+    The signature does not include the `data` values,
+    but only the `data` metadata (shape, decomposition, distributor)
+    and the `idx` content so that the same plan is reused across calls.
+    """
+
+    __slots__ = ('data', 'idx', '_sig')
+
+    def __init__(self, data, idx):
+        self.data = data
+        self.idx = idx
+        self._sig = _signature(self.data, self.idx)
+
+    def __hash__(self):
+        return hash(self._sig)
+
+    def __eq__(self, other):
+        return self._sig == other._sig
+
+
+@lru_cache(maxsize=64)
+def _build(key):
+    return Exchange(key.data, key.idx)
+
+
+def cached_exchange(data, idx):
+    """Return a (cached) `Exchange` for `data[idx]`."""
+    return _build(_ExchangeKey(data, idx))
+
+
+def _signature(data, idx):
+    # The decomposition/distributor are keyed by identity: while cached, the key
+    # keeps them alive so their ids cannot be recycled, and distinct live objects
+    # always have distinct ids.
+    sig = [id(data._distributor), id(data._decomposition), tuple(data.shape)]
+    for component in as_tuple(idx):
+        if isinstance(component, np.ndarray):
+            sig.append(('arr', component.shape, component.dtype.str,
+                        component.tobytes()))
+        elif isinstance(component, slice):
+            sig.append(('slc', component.start, component.stop, component.step))
+        elif isinstance(component, (list, tuple)):
+            arr = np.asarray(component)
+            sig.append(('seq', arr.shape, arr.dtype.str, arr.tobytes()))
+        else:
+            sig.append(('obj', component))
+    return tuple(sig)

devito/data/distributed/exchange.py

@@ -0,0 +1,108 @@
+"""
+Layout layer: where a global coordinate physically lives.
+
+A `Layout` wraps a distributor and a per-axis decomposition and answers,
+for any axis, "which rank owns global index g, and at what local offset". It is
+the single bridge between the layout-independent `Selection` and the
+physical MPI placement, and it is what makes replicated and distributed axes
+look uniform to the planner. All maps are computed locally from replicated
+metadata; no communication happens here.
+"""
+
+from functools import cached_property
+
+import numpy as np
+
+from devito.tools import prod
+
+__all__ = ['Layout']
+
+
+class Layout:
+
+    """
+    Physical placement of a distributed array's axes.
+
+    Parameters
+    ----------
+    distributor : Distributor
+        Provides the communicator, topology and rank/coord maps.
+    decomposition : tuple
+        One entry per array axis: a Decomposition for a distributed axis, or
+        `None` for a replicated axis.
+    global_shape : tuple of int
+        The global array shape (full size on every axis).
+    """
+
+    def __init__(self, distributor, decomposition, global_shape):
+        self.distributor = distributor
+        self.decomposition = tuple(decomposition)
+        self.global_shape = tuple(global_shape)
+
+    @cached_property
+    def distributed_axes(self):
+        """The array axes that are MPI-distributed, in increasing order."""
+        return tuple(a for a, d in enumerate(self.decomposition) if d is not None)
+
+    @cached_property
+    def replicated_axes(self):
+        """The array axes that are replicated on every rank."""
+        return tuple(a for a, d in enumerate(self.decomposition) if d is None)
+
+    @cached_property
+    def replicated_size(self):
+        """Product of the full local sizes of the replicated axes."""
+        return prod(self.global_shape[a] for a in self.replicated_axes)
+
+    def axis_maps(self, axis):
+        """
+        Lookup tables for one distributed axis.
+
+        Returns
+        -------
+        owner : ndarray
+            `owner[g]` is the topology sub-rank owning global index `g` along
+            this axis (`-1` if out of bounds).
+        local : ndarray
+            `local[g]` is the offset of `g` within its owner's subdomain.
+        sizes : ndarray
+            `sizes[i]` is the number of indices owned by sub-rank `i`.
+        """
+        return self._axis_maps[axis]
+
+    @cached_property
+    def _axis_maps(self):
+        maps = {}
+        for axis in self.distributed_axes:
+            dec = self.decomposition[axis]
+            gmin = dec.glb_min or 0
+            n = self.global_shape[axis]
+            owner = np.full(n, -1, dtype=np.int64)
+            local = np.full(n, -1, dtype=np.int64)
+            sizes = np.zeros(len(dec), dtype=np.int64)
+            for i, sub in enumerate(dec):
+                pos = np.asarray(sub, dtype=np.int64) - gmin
+                in_range = (pos >= 0) & (pos < n)
+                owner[pos[in_range]] = i
+                local[pos[in_range]] = np.arange(pos.size)[in_range]
+                sizes[i] = pos.size
+            maps[axis] = (owner, local, sizes)
+        return maps
+
+    @cached_property
+    def topology_shape(self):
+        """Number of sub-ranks along each distributed axis."""
+        return tuple(len(self.decomposition[a]) for a in self.distributed_axes)
+
+    @cached_property
+    def coord_to_rank(self):
+        """Map a topology coordinate tuple (over distributed axes) to a flat rank.
+
+        Grid distributors expose the inverse Cartesian map via `all_coords`;
+        single-axis distributors (e.g. sparse) lay ranks out linearly, so the
+        sole sub-rank index is the flat rank.
+        """
+        if hasattr(self.distributor, 'all_coords'):
+            return {tuple(int(c) for c in coord): r
+                    for r, coord in enumerate(self.distributor.all_coords)}
+        return {(r,): r for r in range(self.distributor.nprocs)}