diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index 343c850149df72f90c6860aa0b5db16fa1528718..ae18a9294cfe203dde5800b4e8eb7db1900847d3 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -1897,8 +1897,12 @@ class FunctionScoper(IdentityMapper):
from loopy.symbolic import Reduction
+ # Adding _reduce at the end of the reduction in order to avoid
+ # confusion between reduce(max, ...) and max(a, b) in the
+ # `scoped_functions` dictionary.
+
return Reduction(
- ScopedFunction(expr.function.name),
+ ScopedFunction(expr.function.name+"_reduce"),
tuple(new_inames),
self.rec(expr.expr),
allow_simultaneous=expr.allow_simultaneous)
@@ -1921,7 +1925,10 @@ class ScopedFunctionCollector(CombineMapper):
from loopy.kernel.function_interface import CallableOnScalar
from loopy.symbolic import Reduction
- callable_reduction = CallableReduction(expr.function.name)
+ # Refer to map_reduction subroutine of FunctionScoper.
+ assert expr.function.name[-7:] == "_reduce"
+
+ callable_reduction = CallableReduction(expr.function.name[:-7])
# sanity checks
@@ -1986,7 +1993,7 @@ def scope_functions(kernel):
else:
raise NotImplementedError("scope_functions not implemented for %s" %
type(insn))
-
+
# Need to combine the scoped functions into a dict
scoped_function_dict = dict(scoped_functions)
return kernel.copy(instructions=new_insns, scoped_functions=scoped_function_dict)
diff --git a/loopy/kernel/function_interface.py b/loopy/kernel/function_interface.py
index 5066cff5c753089d78f3eb688f124a49ad2160c2..2fbb931cb5b37f0f1f0da3925b9c5bb95ab235f5 100644
--- a/loopy/kernel/function_interface.py
+++ b/loopy/kernel/function_interface.py
@@ -566,11 +566,14 @@ def next_indexed_name(name):
class FunctionScopeChanger(IdentityMapper):
- #TODO: Make it sophisticated as in I don't like the if-else systems. Needs
+ # TODO: Make it sophisticated as in I don't like the if-else systems. Needs
# something else.
+ # Explain what this is doing.
+ # The name should be more like "NameChanger" more like "GameChanger" LOl.
+ # Wow my jokes are baaad. Anyways back to work!!
+
def __init__(self, new_names):
self.new_names = new_names
- self.new_names_set = frozenset(new_names.values())
def map_call(self, expr):
if expr in self.new_names:
@@ -594,6 +597,18 @@ class FunctionScopeChanger(IdentityMapper):
else:
return IdentityMapper.map_call_with_kwargs(self, expr)
+ def map_reduction(self, expr):
+ from loopy.symbolic import Reduction
+
+ if self.new_names:
+ return Reduction(
+ ScopedFunction(self.new_names[expr]),
+ tuple(expr.inames),
+ self.rec(expr.expr),
+ allow_simultaneous=expr.allow_simultaneous)
+ else:
+ return IdentityMapper.map_reduction(self, expr)
+
def register_pymbolic_calls_to_knl_callables(kernel,
pymbolic_calls_to_knl_callables):
@@ -654,5 +669,4 @@ def register_pymbolic_calls_to_knl_callables(kernel,
# }}}
-
# vim: foldmethod=marker
diff --git a/loopy/library/reduction.py b/loopy/library/reduction.py
index f4444c8864a44840f624c63ab282769e0f321a9a..f1c5607fe27c750a8ec0ab9b3b22a2eb7e664034 100644
--- a/loopy/library/reduction.py
+++ b/loopy/library/reduction.py
@@ -428,8 +428,8 @@ def get_argext_preamble(kernel, func_id, arg_dtypes):
_REDUCTION_OPS = {
"sum": SumReductionOperation,
"product": ProductReductionOperation,
- "maximum": MaxReductionOperation,
- "minimum": MinReductionOperation,
+ "max": MaxReductionOperation,
+ "min": MinReductionOperation,
"argmax": ArgMaxReductionOperation,
"argmin": ArgMinReductionOperation,
"segmented(sum)": SegmentedSumReductionOperation,
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 3f3c1c472a8936a70dcfdda651478f47d53bcb4b..8950f15900d43ab28c258b650330b7a6b1cc613f 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -2357,7 +2357,22 @@ class ReadyForCodegen(CombineMapper):
map_function_symbol = map_constant
-def try_making_callable_ready_for_codegen(kernel):
+def specializing_incomplete_callables(kernel):
+ """
+ Transformation necessary to type-specialize the callables which are missed
+ in type inference. For example consider:
+ ```
+ knl = lp.make_kernel(
+ "{[i]: 0<=i<16}",
+ "a[i] = sin[b[i]]",
+ [lp.GlobalArg('a', dtype=np.float64),
+ lp.GlobalArg('b', dtype=np.float64)])
+ ```
+ In this case, none of the instructions undergo type inference as the type
+ inference is already resolved. But this would be a problem during
+ code-generation as `sin` is not type specialized.
+
+ """
from loopy.type_inference import TypeInferenceMapper
from loopy.symbolic import SubstitutionRuleExpander
from loopy.kernel.function_interface import (
@@ -2462,7 +2477,6 @@ def preprocess_kernel(kernel, device=None):
# - realize_reduction must happen after default dependencies are added
# because it manipulates the depends_on field, which could prevent
# defaults from being applied.
-
kernel = realize_reduction(kernel, unknown_types_ok=False)
# inferring the shape and dim_tags of the arguments involved in a function
@@ -2470,7 +2484,7 @@ def preprocess_kernel(kernel, device=None):
kernel = infer_arg_descr(kernel)
# try specializing callables one last time.
- kernel = try_making_callable_ready_for_codegen(kernel)
+ kernel = specializing_incomplete_callables(kernel)
# Ordering restriction:
# add_axes_to_temporaries_for_ilp because reduction accumulators
diff --git a/loopy/symbolic.py b/loopy/symbolic.py
index 32670c1cc26bef1dacc18c0a46f70c92a00d9405..831bab5c21ccfccb4d65465eb6dd6335c7b938b5 100644
--- a/loopy/symbolic.py
+++ b/loopy/symbolic.py
@@ -616,7 +616,7 @@ class Reduction(p.Expression):
self.allow_simultaneous = allow_simultaneous
def __getinitargs__(self):
- return (self.funciton, self.inames, self.expr, self.allow_simultaneous)
+ return (self.function, self.inames, self.expr, self.allow_simultaneous)
def get_hash(self):
return hash((self.__class__, self.function, self.inames, self.expr))