diff --git a/MEMO b/MEMO
index 38f88de62a33bc0ef8b83e1483780ada4bc377c6..170019de51d8514a051de9c13a5b55e00b50a709 100644
--- a/MEMO
+++ b/MEMO
@@ -53,12 +53,8 @@ To-do
- rename IndexTag -> InameTag
- Data implementation tags
- TODO initial bringup:
- - Adapt padding
- - Adapt automatic padding of temp variables
- - turn base_indices into offset
-
TODO further:
+ - turn base_indices into offset
- vectorization
- automatic copies
- write_image()
diff --git a/loopy/cse.py b/loopy/cse.py
index 8d8c7477baabe2b8f65c50d33330452c1ae0e592..6d9cc491308d220faee337fe4344799b51754a33 100644
--- a/loopy/cse.py
+++ b/loopy/cse.py
@@ -864,7 +864,7 @@ def precompute(kernel, subst_use, sweep_inames=[], within=None,
name=target_var_name,
dtype=dtype,
base_indices=(0,)*len(non1_storage_shape),
- shape=non1_storage_shape,
+ shape=tuple(non1_storage_shape),
is_local=None)
new_temporary_variables[target_var_name] = temp_var
diff --git a/loopy/kernel/array.py b/loopy/kernel/array.py
index beb2b6d9f8e072489c9d2290d0af981defe32a13..831996ba210c7b6ae8104afbb9dd31029e5898aa 100644
--- a/loopy/kernel/array.py
+++ b/loopy/kernel/array.py
@@ -285,6 +285,7 @@ def _parse_shape_or_strides(x):
x = _pymbolic_parse_if_necessary(x)
if isinstance(x, lp.auto):
return x
+ assert not isinstance(x, list)
if not isinstance(x, tuple):
assert x is not lp.auto
x = (x,)
diff --git a/loopy/kernel/data.py b/loopy/kernel/data.py
index 6d8e27c4f1528c74bc85e6f1b86154346b477a3f..60cba66f0eb808e281970f250096b5c94572a376 100644
--- a/loopy/kernel/data.py
+++ b/loopy/kernel/data.py
@@ -262,8 +262,9 @@ class TemporaryVariable(ArrayBase):
# FIXME take into account storage_shape, or something like it
storage_shape = self.shape
- for l in storage_shape:
- temp_var_decl = ArrayOf(temp_var_decl, l)
+ if storage_shape:
+ temp_var_decl = ArrayOf(temp_var_decl,
+ " * ".join(str(s) for s in storage_shape))
if self.is_local:
temp_var_decl = CLLocal(temp_var_decl)
diff --git a/loopy/padding.py b/loopy/padding.py
index 65d90a8f48c86579e0d444306e8de3e6d10dfa85..ee43f2590c915c582ab11d5891e2a23d0fdbf266 100644
--- a/loopy/padding.py
+++ b/loopy/padding.py
@@ -108,24 +108,34 @@ def split_arg_axis(kernel, args_and_axes, count):
# }}}
- # {{{ adjust strides
+ # {{{ adjust dim tags
- new_strides = list(arg.strides)
- old_stride = new_strides[axis]
+ if arg.dim_tags is None:
+ raise RuntimeError("dim_tags of '%s' are not known" % arg.name)
+ new_dim_tags = list(arg.dim_tags)
+
+ old_dim_tag = arg.dim_tags[axis]
+
+ from loopy.kernel.array import FixedStrideArrayDimTag
+ if not isinstance(old_dim_tag, FixedStrideArrayDimTag):
+ raise RuntimeError("axis %d of '%s' is not tagged fixed-stride"
+ % (axis, arg.name))
+
+ old_stride = old_dim_tag.stride
outer_stride = count*old_stride
if order == "F":
- new_strides.insert(axis+1, outer_stride)
+ new_dim_tags.insert(axis+1, FixedStrideArrayDimTag(outer_stride))
elif order == "C":
- new_strides.insert(axis, outer_stride)
+ new_dim_tags.insert(axis, FixedStrideArrayDimTag(outer_stride))
else:
raise RuntimeError("order '%s' not understood" % order)
- new_strides = tuple(new_strides)
+ new_dim_tags = tuple(new_dim_tags)
# }}}
- new_args[arg_idx] = arg.copy(shape=new_shape, strides=new_strides)
+ new_args[arg_idx] = arg.copy(shape=new_shape, dim_tags=new_dim_tags)
# }}}
@@ -187,9 +197,22 @@ def split_arg_axis(kernel, args_and_axes, count):
def find_padding_multiple(kernel, variable, axis, align_bytes, allowed_waste=0.1):
arg = kernel.arg_dict[variable]
- stride = arg.strides[axis]
+ if arg.dim_tags is None:
+ raise RuntimeError("cannot find padding multiple--dim_tags of '%s' "
+ "are not known" % variable)
+
+ dim_tag = arg.dim_tags[axis]
+
+ from loopy.kernel.array import FixedStrideArrayDimTag
+ if not isinstance(dim_tag, FixedStrideArrayDimTag):
+ raise RuntimeError("cannot find padding multiple--"
+ "axis %d of '%s' is not tagged fixed-stride"
+ % (axis, variable))
+
+ stride = dim_tag.stride
+
if not isinstance(stride, int):
- raise RuntimeError("cannot find padding multi--stride is not a "
+ raise RuntimeError("cannot find padding multiple--stride is not a "
"known integer")
from pytools import div_ceil
@@ -212,21 +235,31 @@ def add_padding(kernel, variable, axis, align_bytes):
new_args = kernel.args[:]
arg = new_args[arg_idx]
- new_strides = list(arg.strides)
- stride = new_strides[axis]
+ if arg.dim_tags is None:
+ raise RuntimeError("cannot add padding--dim_tags of '%s' "
+ "are not known" % variable)
+
+ new_dim_tags = list(arg.dim_tags)
+ dim_tag = new_dim_tags[axis]
+
+ from loopy.kernel.array import FixedStrideArrayDimTag
+ if not isinstance(dim_tag, FixedStrideArrayDimTag):
+ raise RuntimeError("cannot find padding multiple--"
+ "axis %d of '%s' is not tagged fixed-stride"
+ % (axis, variable))
+
+ stride = dim_tag.stride
if not isinstance(stride, int):
raise RuntimeError("cannot find split granularity--stride is not a "
"known integer")
+
from pytools import div_ceil
- new_strides[axis] = div_ceil(stride, align_bytes) * align_bytes
+ new_dim_tags[axis] = FixedStrideArrayDimTag(
+ div_ceil(stride, align_bytes) * align_bytes)
- new_args[arg_idx] = arg.copy(strides=tuple(new_strides))
+ new_args[arg_idx] = arg.copy(dim_tags=tuple(new_dim_tags))
return kernel.copy(args=new_args)
-
-
-
-
# vim: foldmethod=marker
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 09adb1278fb45bfafd91a2603c5253a163f421ff..dc32e01454e361b96d9bb5c16d1e994d4bf7bdde 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -710,14 +710,24 @@ def get_auto_axis_iname_ranking_by_stride(kernel, insn):
ary_name = aae.aggregate.name
arg = kernel.arg_dict.get(ary_name)
- ary_strides = arg.strides
- if ary_strides is None and len(index_expr) == 1:
- ary_strides = (1,)
+ if arg.dim_tags is None:
+ from warnings import warn
+ warn("Strides for '%s' are not known. Local axis assignment "
+ "is likely suboptimal." % arg.name)
+ ary_strides = [1] * len(index_expr)
+ else:
+ ary_strides = []
+ from loopy.kernel.array import FixedStrideArrayDimTag
+ for dim_tag in arg.dim_tags:
+ if isinstance(dim_tag, FixedStrideArrayDimTag):
+ ary_strides.append(dim_tag.stride)
# {{{ construct iname_to_stride_expr
iname_to_stride_expr = {}
for iexpr_i, stride in zip(index_expr, ary_strides):
+ if stride is None:
+ continue
coeffs = CoefficientCollector()(iexpr_i)
for var_name, coeff in coeffs.iteritems():
if var_name in auto_axis_inames: # excludes '1', i.e. the constant