def _lower_iet(cls, stree, profiler, **kwargs):
"""
Iteration/Expression tree lowering:
* Turn a ScheduleTree into an Iteration/Expression tree;
* Introduce distributed-memory, shared-memory, and SIMD parallelism;
* Introduce optimizations for data locality;
* Finalize (e.g., symbol definitions, array casts)
"""
name = kwargs.get("name", "Kernel")
sregistry = kwargs['sregistry']
# Build an IET from a ScheduleTree
iet = iet_build(stree)
# Analyze the IET Sections for C-level profiling
profiler.analyze(iet)
# Wrap the IET with an EntryFunction (a special Callable representing
# the entry point of the generated library)
parameters = derive_parameters(iet, True)
iet = EntryFunction(name, iet, 'int', parameters, ())
# Lower IET to a target-specific IET
graph = Graph(iet)
graph = cls._specialize_iet(graph, **kwargs)
# Instrument the IET for C-level profiling
# Note: this is postponed until after _specialize_iet because during
# specialization further Sections may be introduced
instrument(graph, profiler=profiler, sregistry=sregistry)
return graph.root, graph
def test_strides_forwarding1():
grid = Grid(shape=(4, 4))
a = Array(name='a', dimensions=grid.dimensions, shape=grid.shape)
bar = Callable('bar',
DummyExpr(a[0, 0], 0),
'void',
parameters=[a.indexed])
call = Call(bar.name, [a.indexed])
foo = Callable('foo', call, 'void', parameters=[a])
# Emulate what the compiler would do
graph = Graph(foo)
graph.efuncs['bar'] = bar
linearize(graph, mode=True, sregistry=SymbolRegistry())
# Despite `a` is passed via `a.indexed`, and since it's an Array (which
# have symbolic shape), we expect the stride exprs to be placed in `bar`,
# and in `bar` only, as `foo` doesn't really use `a`, it just propagates it
# down to `bar`
foo = graph.root
bar = graph.efuncs['bar']
assert len(foo.body.body) == 1
assert foo.body.body[0].is_Call
assert len(bar.body.body) == 5
assert bar.body.body[0].write.name == 'y_fsz0'
assert bar.body.body[2].write.name == 'y_stride0'
def _lower_iet(cls, stree, profiler, **kwargs):
"""
Iteration/Expression tree lowering:
* Turn a ScheduleTree into an Iteration/Expression tree;
* Perform analysis to detect optimization opportunities;
* Introduce distributed-memory, shared-memory, and SIMD parallelism;
* Introduce optimizations for data locality;
* Finalize (e.g., symbol definitions, array casts)
"""
name = kwargs.get("name", "Kernel")
iet = iet_build(stree)
# Instrument the IET for C-level profiling
iet = profiler.instrument(iet)
# Wrap the IET with a Callable
parameters = derive_parameters(iet, True)
iet = Callable(name, iet, 'int', parameters, ())
# Lower IET to a target-specific IET
graph = Graph(iet)
graph = cls._specialize_iet(graph, **kwargs)
return graph.root, graph
def test_strides_forwarding0():
grid = Grid(shape=(4, 4))
f = Function(name='f', grid=grid)
bar = Callable('bar',
DummyExpr(f[0, 0], 0),
'void',
parameters=[f.indexed])
call = Call(bar.name, [f.indexed])
foo = Callable('foo', call, 'void', parameters=[f])
# Emulate what the compiler would do
graph = Graph(foo)
graph.efuncs['bar'] = bar
linearize(graph, mode=True, sregistry=SymbolRegistry())
# Since `f` is passed via `f.indexed`, we expect the stride exprs to be
# lifted in `foo` and then passed down to `bar` as arguments
foo = graph.root
bar = graph.efuncs['bar']
assert foo.body.body[0].write.name == 'y_fsz0'
assert foo.body.body[2].write.name == 'y_stride0'
assert len(foo.body.body[4].arguments) == 2
assert len(bar.parameters) == 2
assert bar.parameters[1].name == 'y_stride0'
assert len(bar.body.body) == 1