Merge pull request #2719 from devitocodes/pytest-mark-name

compiler: improve rcompile

Author: FabioLuporini

conftest.py

@@ -4,6 +4,7 @@
 
 import pytest
 from sympy import Add
+from sympy.printing import sstr
 
 from devito import Eq, configuration, Revolver  # noqa
 from devito.checkpointing import NoopRevolver
@@ -287,6 +288,20 @@ def pytest_runtest_makereport(item, call):
             result.outcome = 'passed'
 
 
+def pytest_make_parametrize_id(config, val, argname):
+    """
+    Prevents pytest from making obscure parameter names (param0, param1, ...)
+    and default to sympy.sstr(val) instead for better log readability.
+    """
+    # First see if it has a name
+    if hasattr(val, '__name__'):
+        return val.__name__
+    try:
+        return sstr(val)
+    except Exception:
+        return None  # Fall back to default behavior
+
+
 # A list of optimization options/pipelines to be used in testing
 # regarding spatial and/or temporal blocking.
 opts_tiling = ['advanced',

devito/arch/compiler.py

@@ -180,6 +180,7 @@ def __init__(self):
     """
 
     fields = {'cc', 'ld'}
+    linker_opt = '-Wl,'
     _default_cpp = False
     _cxxstd = 'c++14'
     _cstd = 'c99'
@@ -425,7 +426,7 @@ def add_library_dirs(self, dirs, rpath=False):
         if rpath:
             # Add rpath flag to embed library dir
             for d in as_list(dirs):
-                self.ldflags.append(f'-Wl,-rpath,{d}')
+                self.ldflags.append(f'{self.linker_opt}-rpath,{d}')
 
     def add_libraries(self, libs):
         self.libraries = filter_ordered(self.libraries + as_list(libs))
@@ -662,6 +663,7 @@ def __lookup_cmds__(self):
 class CudaCompiler(Compiler):
 
     _default_cpp = True
+    linker_opt = "--linker-options="
 
     def __init_finalize__(self, **kwargs):
 

devito/core/__init__.py

@@ -6,7 +6,8 @@
     Cpu64CustomOperator, Cpu64CustomCXXOperator,
     Cpu64CXXNoopCOperator, Cpu64CXXNoopOmpOperator,
     Cpu64AdvCXXOperator, Cpu64AdvCXXOmpOperator,
-    Cpu64FsgCXXOperator, Cpu64FsgCXXOmpOperator
+    Cpu64FsgCXXOperator, Cpu64FsgCXXOmpOperator,
+    Cpu64CustomCXXOmpOperator, Cpu64CustomCOperator
 )
 from devito.core.intel import (
     Intel64AdvCOperator, Intel64AdvOmpOperator,
@@ -33,11 +34,11 @@
 from devito.operator.registry import operator_registry
 
 # Register CPU Operators
-operator_registry.add(Cpu64CustomOperator, Cpu64, 'custom', 'C')
+operator_registry.add(Cpu64CustomCOperator, Cpu64, 'custom', 'C')
 operator_registry.add(Cpu64CustomOperator, Cpu64, 'custom', 'openmp')
 operator_registry.add(Cpu64CustomOperator, Cpu64, 'custom', 'Copenmp')
 operator_registry.add(Cpu64CustomCXXOperator, Cpu64, 'custom', 'CXX')
-operator_registry.add(Cpu64CustomCXXOperator, Cpu64, 'custom', 'CXXopenmp')
+operator_registry.add(Cpu64CustomCXXOmpOperator, Cpu64, 'custom', 'CXXopenmp')
 
 operator_registry.add(Cpu64NoopCOperator, Cpu64, 'noop', 'C')
 operator_registry.add(Cpu64NoopOmpOperator, Cpu64, 'noop', 'openmp')

devito/core/cpu.py

@@ -321,11 +321,24 @@ def _make_iet_passes_mapper(cls, **kwargs):
     assert not (set(_known_passes) & set(_known_passes_disabled))
 
 
-class Cpu64CustomCXXOperator(Cpu64CustomOperator):
+class Cpu64CustomCXXOmpOperator(Cpu64CustomOperator):
 
     _Target = CXXOmpTarget
     LINEARIZE = True
 
+
+class Cpu64CustomCOperator(Cpu64CustomOperator):
+
+    _Target = CTarget
+    LINEARIZE = False
+
+
+class Cpu64CustomCXXOperator(Cpu64CustomOperator):
+
+    _Target = CXXTarget
+    LINEARIZE = True
+
+
 # Language level
 
 

devito/operator/operator.py

@@ -18,8 +18,8 @@
                            switch_log_level)
 from devito.ir.equations import LoweredEq, lower_exprs, concretize_subdims
 from devito.ir.clusters import ClusterGroup, clusterize
-from devito.ir.iet import (Callable, CInterface, EntryFunction, FindSymbols,
-                           MetaCall, derive_parameters, iet_build)
+from devito.ir.iet import (Callable, CInterface, EntryFunction, DeviceFunction,
+                           FindSymbols, MetaCall, derive_parameters, iet_build)
 from devito.ir.support import AccessMode, SymbolRegistry
 from devito.ir.stree import stree_build
 from devito.operator.profiling import create_profile
@@ -1270,9 +1270,12 @@ def rcompile(expressions, kwargs, options, target=None):
 
     # Recursive compilation is expensive, so we cache the result because sometimes
     # it is called multiple times for the same input
-    compiled = RCompiles(expressions, cls).compile(**kwargs)
+    irs, byproduct0 = RCompiles(expressions, cls).compile(**kwargs)
 
-    return compiled
+    key = lambda i: isinstance(i, (EntryFunction, DeviceFunction))
+    byproduct = byproduct0.filter(key)
+
+    return irs, byproduct
 
 
 # *** Misc helpers
@@ -1518,7 +1521,7 @@ def nbytes_snapshots(self):
         # Layers are sometimes aliases, so include aliases here
         for i in self._op_symbols:
             try:
-                if i._child is None:
+                if i._child is None and i.alias is not True:
                     # Use only the "innermost" layer to avoid counting snapshots
                     # twice. This layer will have no child.
                     v = self._apply_override(i)
@@ -1574,6 +1577,11 @@ def parse_kwargs(**kwargs):
     else:
         raise InvalidOperator(f"Illegal `opt={str(opt)}`")
 
+    # `openmp` in mode e.g `opt=('openmp', 'simd', {})`
+    if mode and 'openmp' in mode:
+        options['openmp'] = True
+        mode = tuple(i for i in as_tuple(mode) if i != 'openmp')
+
     # `opt`, deprecated kwargs
     kwopenmp = kwargs.get('openmp', options.get('openmp'))
     if kwopenmp is None:

devito/passes/iet/definitions.py

@@ -86,7 +86,8 @@ class DataManager:
     The language used to express data allocations, deletions, and host-device transfers.
     """
 
-    def __init__(self, rcompile=None, sregistry=None, platform=None, **kwargs):
+    def __init__(self, rcompile=None, sregistry=None, platform=None,
+                 options=None, **kwargs):
         self.rcompile = rcompile
         self.sregistry = sregistry
         self.platform = platform

devito/passes/iet/engine.py

@@ -1,13 +1,13 @@
-from collections import OrderedDict, defaultdict
+from collections import defaultdict
 from functools import partial, singledispatch, wraps
 
 import numpy as np
 from sympy import Mul
 
 from devito.ir.iet import (
     Call, ExprStmt, Expression, Iteration, SyncSpot, AsyncCallable, FindNodes,
-    FindSymbols, MapNodes, MetaCall, Transformer, EntryFunction,
-    ThreadCallable, Uxreplace, derive_parameters
+    FindSymbols, MapNodes, MetaCall, Transformer, EntryFunction, ThreadCallable,
+    Uxreplace, derive_parameters
 )
 from devito.ir.support import SymbolRegistry
 from devito.mpi.distributed import MPINeighborhood
@@ -28,7 +28,31 @@
 __all__ = ['Graph', 'iet_pass', 'iet_visit']
 
 
-class Graph:
+class Byproduct:
+
+    """
+    A Byproduct is a mutable collection of metadata produced by one or more
+    compiler passes.
+
+    This metadata may be used internally or by the caller itself, typically
+    for code generation purposes.
+    """
+
+    def __init__(self, efuncs=None, includes=None, headers=None, namespaces=None,
+                 globs=None):
+        self.efuncs = efuncs or {}
+        self.includes = includes or []
+        self.headers = headers or []
+        self.namespaces = namespaces or []
+        self.globals = globs or []
+
+    @property
+    def funcs(self):
+        return tuple(MetaCall(v, True) for v in self.efuncs.values()
+                     if not isinstance(v, EntryFunction))
+
+
+class Graph(Byproduct):
 
     """
     DAG representation of a call graph.
@@ -49,16 +73,9 @@ class Graph:
     """
 
     def __init__(self, iet, options=None, sregistry=None, **kwargs):
-        self.efuncs = OrderedDict([(iet.name, iet)])
-
         self.sregistry = sregistry
 
-        self.includes = []
-        self.headers = []
-        self.namespaces = []
-        self.globals = []
-
-        # Stash immutable information useful for one or more compiler passes
+        super().__init__({iet.name: iet})
 
         # All written user-level objects
         writes = FindSymbols('writes').visit(iet)
@@ -79,10 +96,6 @@ def __init__(self, iet, options=None, sregistry=None, **kwargs):
     def root(self):
         return self.efuncs[list(self.efuncs).pop(0)]
 
-    @property
-    def funcs(self):
-        return tuple(MetaCall(v, True) for v in self.efuncs.values())[1:]
-
     @property
     def sync_mapper(self):
         """
@@ -146,7 +159,7 @@ def apply(self, func, **kwargs):
             new_efuncs = metadata.get('efuncs', [])
 
             efuncs[i] = efunc
-            efuncs.update(OrderedDict([(i.name, i) for i in new_efuncs]))
+            efuncs.update(dict([(i.name, i) for i in new_efuncs]))
 
             # Update the parameters / arguments lists since `func` may have
             # introduced or removed objects
@@ -176,10 +189,24 @@ def visit(self, func, **kwargs):
         dag = create_call_graph(self.root.name, self.efuncs)
         toposort = dag.topological_sort()
 
-        mapper = OrderedDict([(i, func(self.efuncs[i], **kwargs)) for i in toposort])
+        mapper = dict([(i, func(self.efuncs[i], **kwargs)) for i in toposort])
 
         return mapper
 
+    def filter(self, key):
+        """
+        Return a Byproduct containing only the Callables in the Graph
+        for which `key` evaluates to True. The resulting object cannot be
+        further modified by an IET pass.
+        """
+        return Byproduct(
+            efuncs={i: v for i, v in self.efuncs.items() if key(v)},
+            includes=as_tuple(self.includes),
+            headers=as_tuple(self.headers),
+            namespaces=as_tuple(self.namespaces),
+            globs=as_tuple(self.globals)
+        )
+
 
 def iet_pass(func):
     if isinstance(func, tuple):
@@ -732,7 +759,7 @@ def _filter(v, efunc=None):
 
         return processed
 
-    efuncs = OrderedDict(efuncs)
+    efuncs = dict(efuncs)
     efuncs[root.name] = root._rebuild(parameters=_filter(root.parameters, root))
 
     # Update all call sites to use the new signature

examples/cfd/01_convection.ipynb

@@ -382,7 +382,7 @@
     "# Now combine the BC expressions with the stencil to form operator\n",
     "expressions = [Eq(u.forward, stencil)]\n",
     "expressions += [bc_left, bc_right, bc_top, bc_bottom]\n",
-    "op = Operator(expressions=expressions, opt=None, openmp=False)  # <-- Turn off performance optimisations\n",
+    "op = Operator(expressions=expressions, opt=None, language='C')  # <-- Turn off performance optimisations\n",
     "op(time=nt, dt=dt)\n",
     "\n",
     "plot_field(u.data[0])\n",

examples/cfd/01_convection_revisited.ipynb

@@ -359,7 +359,7 @@
     "# Now combine the BC expressions with the stencil to form operator.\n",
     "expressions = [Eq(u.forward, stencil, subdomain=grid.interior)]\n",
     "expressions += [bc_left, bc_right, bc_top, bc_bottom]\n",
-    "op = Operator(expressions=expressions, opt=None, openmp=False)  # <-- Turn off performance optimisations\n",
+    "op = Operator(expressions=expressions, opt=None, language='C')  # <-- Turn off performance optimisations\n",
     "op(time=nt, dt=dt)\n",
     "\n",
     "plot_field(u.data[0])\n",

examples/userapi/05_conditional_dimension.ipynb

@@ -792,8 +792,8 @@
     "for n, i in enumerate(f.dimensions):\n",
     "    eqs.append(Eq(g[k, n], f.dimensions[n], implicit_dims=(f.dimensions + (ci,))))\n",
     "\n",
-    "# TODO: Must be openmp=False for now due to issue #2061\n",
-    "op0 = Operator(eqs, openmp=False)\n",
+    "# TODO: Must be language='C' for now due to issue #2061\n",
+    "op0 = Operator(eqs, language='C')\n",
     "op0.apply()\n",
     "print(op0.body.body[-1])\n",
     "\n",