Merge pull request #2948 from devitocodes/smarter-tuner-3

compiler: Enable blocking-before-CIRE

Author: FabioLuporini

devito/arch/compiler.py

@@ -769,7 +769,7 @@ def __init_finalize__(self, **kwargs):
         if not configuration['safe-math']:
             self.cflags.append('--use_fast_math')
 
-        if configuration['profiling'] == 'advanced2':
+        if configuration['profiling'] in ('advanced2', 'ncu'):
             # Optionally print out per-kernel shared memory and register usage
             self.cflags.append('--ptxas-options=-v')
 

devito/core/cpu.py

@@ -5,8 +5,8 @@
 from devito.operator.operator import rcompile
 from devito.passes import stream_dimensions
 from devito.passes.clusters import (
-    Lift, blocking, buffering, cire, cse, factorize, fission, fuse, optimize_hyperplanes,
-    optimize_pows
+    Lift, apply_par_tiles, blocking, buffering, cire, cse, factorize, fission, fuse,
+    optimize_hyperplanes, optimize_pows
 )
 from devito.passes.equations import collect_derivatives
 from devito.passes.iet import (
@@ -67,6 +67,7 @@ def _normalize_kwargs(cls, **kwargs):
                                 reduce=oo.pop('par-tile-reduce', None))
 
         # CIRE
+        o['cire-block-temps'] = oo.pop('cire-block-temps', cls.CIRE_BLOCK_TEMPS)
         o['min-storage'] = oo.pop('min-storage', False)
         o['cire-rotate'] = oo.pop('cire-rotate', False)
         o['cire-maxpar'] = oo.pop('cire-maxpar', False)
@@ -198,6 +199,9 @@ def _specialize_clusters(cls, clusters, **kwargs):
         if options['blocklazy']:
             clusters = blocking(clusters, sregistry, options)
 
+        # Unfold the `par-tile`s, if any
+        clusters = apply_par_tiles(clusters, **kwargs)
+
         return clusters
 
     @classmethod

devito/core/gpu.py

@@ -7,8 +7,8 @@
 from devito.operator.operator import rcompile
 from devito.passes import is_on_device, stream_dimensions
 from devito.passes.clusters import (
-    Lift, blocking, buffering, cire, cse, factorize, fission, fuse, memcpy_prefetch,
-    optimize_pows, tasking
+    Lift, apply_par_tiles, blocking, buffering, cire, cse, factorize, fission, fuse,
+    memcpy_prefetch, optimize_pows, tasking
 )
 from devito.passes.equations import collect_derivatives
 from devito.passes.iet import (
@@ -38,7 +38,9 @@
 
 class DeviceOperatorMixin:
 
+    # Overrides the default values in the main Operator class
     BLOCK_LEVELS = 0
+    CIRE_BLOCK_TEMPS = False
     MPI_MODES = (True, 'basic',)
 
     GPU_FIT = 'all-fallback'
@@ -76,9 +78,10 @@ def _normalize_kwargs(cls, **kwargs):
         o['skewing'] = oo.pop('skewing', False)
 
         # CIRE
+        o['cire-block-temps'] = oo.pop('cire-block-temps', cls.CIRE_BLOCK_TEMPS)
         o['min-storage'] = False
         o['cire-rotate'] = False
-        o['cire-maxpar'] = oo.pop('cire-maxpar', True)
+        o['cire-maxpar'] = oo.pop('cire-maxpar', 'basic')
         o['cire-ftemps'] = oo.pop('cire-ftemps', False)
         o['cire-mingain'] = oo.pop('cire-mingain', cls.CIRE_MINGAIN)
         o['cire-minmem'] = oo.pop('cire-minmem', cls.CIRE_MINMEM)
@@ -239,6 +242,9 @@ def _specialize_clusters(cls, clusters, **kwargs):
         if options['blocklazy']:
             clusters = blocking(clusters, sregistry, options)
 
+        # Unfold the `par-tile`s, if any
+        clusters = apply_par_tiles(clusters, **kwargs)
+
         return clusters
 
     @classmethod

devito/core/operator.py

@@ -58,6 +58,13 @@ class BasicOperator(Operator):
     situations where the performance impact might be detrimental.
     """
 
+    CIRE_BLOCK_TEMPS = True
+    """
+    If an aliasing expression is computed within a blocked loop nest, all CIRE-
+    generated temporaries will inherit the block shape. If set to False, the
+    temporaries shape will systematically be defined by the root Dimensions.
+    """
+
     CIRE_MINGAIN = 10
     """
     Minimum operation count reduction for a redundant expression to be optimized
@@ -240,6 +247,9 @@ def _check_kwargs(cls, **kwargs):
         if oo['mpi'] and oo['mpi'] not in cls.MPI_MODES:
             raise InvalidOperator(f"Unsupported MPI mode `{oo['mpi']}`")
 
+        if oo['cire-maxpar'] not in (False, 'basic', 'compact'):
+            raise InvalidOperator("Illegal `cire-maxpar` value")
+
         if oo['cse-algo'] not in ('basic', 'smartsort', 'advanced'):
             raise InvalidOperator("Illegal `cse-algo` value")
 

devito/finite_differences/differentiable.py

@@ -985,8 +985,9 @@ def value(self, idx):
 class IndexDerivative(IndexSum):
 
     __rargs__ = ('expr', 'mapper')
+    __rkwargs__ = IndexSum.__rkwargs__ + ('deriv_order',)
 
-    def __new__(cls, expr, mapper, **kwargs):
+    def __new__(cls, expr, mapper, deriv_order=None, **kwargs):
         dimensions = as_tuple(set(mapper.values()))
 
         # Detect the Weights among the arguments
@@ -1008,6 +1009,8 @@ def __new__(cls, expr, mapper, **kwargs):
         obj._weights = weights
         obj._mapper = frozendict(mapper)
 
+        obj._deriv_order = deriv_order
+
         return obj
 
     def _hashable_content(self):
@@ -1040,6 +1043,10 @@ def weights(self):
     def mapper(self):
         return self._mapper
 
+    @property
+    def deriv_order(self):
+        return self._deriv_order
+
     @property
     def depth(self):
         iderivs = self.expr.find(IndexDerivative)
@@ -1216,7 +1223,8 @@ def _diff2sympy(obj):
 
         # Handle special objects
         if isinstance(obj, DiffDerivative):
-            return IndexDerivative(*args, obj.mapper), True
+            return IndexDerivative(*args, obj.mapper,
+                                   deriv_order=obj.deriv_order), True
 
         # Handle generic objects such as arithmetic operations
         try:

devito/finite_differences/finite_difference.py

@@ -210,7 +210,9 @@ def make_derivative(expr, dim, fd_order, deriv_order, side, matvec, x0, coeffici
         with suppress(AttributeError):
             expr = expr._evaluate(expand=False)
 
-        deriv = DiffDerivative(expr*weights, {dim: indices.free_dim})
+        deriv = DiffDerivative(
+            expr*weights, {dim: indices.free_dim}, deriv_order=deriv_order
+        )
     else:
         terms = []
         for i, c in zip(indices, weights, strict=True):

devito/ir/clusters/cluster.py

@@ -13,7 +13,7 @@
 )
 from devito.mpi.halo_scheme import HaloScheme, HaloTouch
 from devito.mpi.reduction_scheme import DistReduce
-from devito.symbolics import estimate_cost
+from devito.symbolics import estimate_cost, uxreplace
 from devito.tools import (
     CacheInstances, as_tuple, cached_hash, filter_ordered, flatten, infer_dtype
 )
@@ -112,7 +112,7 @@ def dimensions(self):
     @cached_property
     def exprs_dimensions(self):
         """
-        The Dimensions that appear explicitly in the Cluster expressions.
+        The Dimensions that appear explicitly in the expressions.
         """
         dims_explicit = {i for i in self.free_symbols if i.is_Dimension}
         dims_implicit = {d for e in self.exprs for d in e.implicit_dims}
@@ -121,7 +121,7 @@ def exprs_dimensions(self):
     @cached_property
     def guards_dimensions(self):
         """
-        The Dimensions that appear explicitly in the Cluster guards.
+        The Dimensions that appear explicitly in the guards.
         """
         syms_guards = {d for e in self.guards.values() for d in e.free_symbols}
         dims_guards = {i for i in syms_guards if i.is_Dimension}
@@ -142,7 +142,7 @@ def used_dimensions(self):
     @cached_property
     def dist_dimensions(self):
         """
-        The Cluster's distributed Dimensions.
+        The distributed Dimensions.
         """
         ret = set()
         for f in self.functions:
@@ -168,7 +168,7 @@ def grid(self):
         elif len(grids) == 1:
             return grids.pop()
         else:
-            raise ValueError("Cluster has no unique Grid")
+            raise ValueError("Multiple Grids detected")
 
     @cached_property
     def is_scalar(self):
@@ -296,31 +296,27 @@ def is_glb_load_to_mem_shared(self):
     @cached_property
     def is_async(self):
         """
-        True if an asynchronous Cluster, False otherwise.
+        True if asynchronous, False otherwise.
         """
         return any(isinstance(s, (WithLock, PrefetchUpdate))
                    for s in flatten(self.syncs.values()))
 
     @cached_property
     def is_wait(self):
         """
-        True if a Cluster waiting on a lock (that is a special synchronization
-        operation), False otherwise.
+        True if waiting on a lock (that is a special synchronization operation),
+        False otherwise.
         """
         return any(isinstance(s, WaitLock)
                    for s in flatten(self.syncs.values()))
 
     @cached_property
     def dtype(self):
         """
-        The arithmetic data type of the Cluster.
+        The arithmetic data type of the enclosed expressions.
 
-        If the Cluster performs floating point arithmetic, then the expressions
-        performing integer arithmetic are ignored, assuming that they are only
-        carrying out array index calculations.
-
-        If two expressions perform calculations with different precision,
-        the data type with highest precision is returned.
+        If two expressions perform calculations with different precision, the data
+        type with highest precision is returned.
         """
         dtypes = set()
         for i in self.exprs:
@@ -336,8 +332,8 @@ def dtype(self):
     @cached_property
     def dspace(self):
         """
-        Derive the DataSpace of the Cluster from its expressions,
-        IterationSpace, and Guards.
+        The DataSpace deriving from the enclosed expressions, IterationSpace,
+        and Guards.
         """
         accesses = detect_accesses(self.exprs)
 
@@ -421,8 +417,8 @@ def ops(self):
     @cached_property
     def traffic(self):
         """
-        The Cluster compulsory traffic (number of reads/writes), as a mapper
-        from Functions to IntervalGroups.
+        The compulsory traffic (number of reads/writes), as a mapper from
+        Functions to IntervalGroups.
 
         Notes
         -----
@@ -509,30 +505,6 @@ def __getattr__(self, name):
             raise AttributeError(name) from None
         return getattr(block, name)
 
-    @property
-    def exprs(self):
-        return self._block.exprs
-
-    @property
-    def ispace(self):
-        return self._block.ispace
-
-    @property
-    def guards(self):
-        return self._block.guards
-
-    @property
-    def properties(self):
-        return self._block.properties
-
-    @property
-    def syncs(self):
-        return self._block.syncs
-
-    @property
-    def halo_scheme(self):
-        return self._block.halo_scheme
-
     @classmethod
     def from_clusters(cls, *clusters):
         """
@@ -612,6 +584,33 @@ def rebuild(self, *args, **kwargs):
                               syncs=syncs,
                               halo_scheme=halo_scheme)
 
+    def subs(self, mapper, compact=()):
+        """
+        Build a new Cluster applying substitutions rules to `self`.
+        """
+        if not mapper:
+            return self
+
+        if self.halo_scheme:
+            raise NotImplementedError
+
+        key0 = lambda i: i.is_Block
+        subs0 = {d: self.ispace[d].promote(key0).dim for d in compact}
+
+        subs = {**mapper, **subs0}
+        exprs = [uxreplace(e, subs) for e in self.exprs]
+
+        ispace = self.ispace.switch(mapper)
+        key = lambda i: key0(i) and i in flatten(d._defines for d in subs0)
+        ispace = ispace.promote(key, mode='total')
+
+        guards = self.guards.subs(mapper).promote(subs0)
+        properties = self.properties.subs(mapper).promote(subs0)
+        syncs = self.syncs.subs(mapper)
+
+        return self.__class__(exprs=exprs, ispace=ispace, guards=guards,
+                              properties=properties, syncs=syncs)
+
 
 class ClusterGroup(tuple):
 
@@ -691,7 +690,15 @@ def dspace(self):
         """Return the DataSpace of this ClusterGroup."""
         return DataSpace.union(*[i.dspace.reset() for i in self])
 
-    @property
+    @cached_property
+    def is_dense(self):
+        return all(i.is_dense for i in self)
+
+    @cached_property
+    def is_wild(self):
+        return all(i.is_wild for i in self)
+
+    @cached_property
     def is_halo_touch(self):
         return all(i.is_halo_touch for i in self)
 

devito/ir/clusters/visitors.py

@@ -37,15 +37,7 @@ def _make_key(self, cluster, level):
         assert self._q_ispace_in_key
         ispace = cluster.ispace[:level]
 
-        if self._q_guards_in_key:
-            try:
-                guards = tuple(cluster.guards.get(i.dim) for i in ispace)
-            except AttributeError:
-                # `cluster` is actually a ClusterGroup
-                assert len(cluster.guards) == 1
-                guards = tuple(cluster.guards[0].get(i.dim) for i in ispace)
-        else:
-            guards = None
+        guards = self._make_key_guards(cluster, ispace)
 
         if self._q_properties_in_key:
             properties = cluster.properties.drop(cluster.ispace[level:].itdims)
@@ -68,6 +60,17 @@ def _make_key(self, cluster, level):
 
         return (prefix,) + subkey
 
+    def _make_key_guards(self, cluster, ispace):
+        if not self._q_guards_in_key:
+            return None
+
+        try:
+            return tuple(cluster.guards.get(i.dim) for i in ispace)
+        except AttributeError:
+            # `cluster` is actually a ClusterGroup
+            assert len(cluster.guards) == 1
+            return tuple(cluster.guards[0].get(i.dim) for i in ispace)
+
     def _make_key_hook(self, cluster, level):
         return ()