diff --git a/loopy/preprocess.py b/loopy/preprocess.py index c0eb91ea60317ef8cad1c594571d46bba2d1a671..23c4b7fbd9e55006dd17ed9b127e598a14ee17a2 100644 --- a/loopy/preprocess.py +++ b/loopy/preprocess.py @@ -289,7 +289,7 @@ def _classify_reduction_inames(kernel, inames): nonlocal_par = [] from loopy.kernel.data import ( - LocalIndexTagBase, UnrolledIlpTag, UnrollTag, VectorizeTag, + LocalIndexTagBase, UnrolledIlpTag, UnrollTag, ConcurrentTag, filter_iname_tags_by_type) for iname in inames: @@ -303,7 +303,7 @@ def _classify_reduction_inames(kernel, inames): elif filter_iname_tags_by_type(iname_tags, LocalIndexTagBase): local_par.append(iname) - elif filter_iname_tags_by_type(iname_tags, (ConcurrentTag, VectorizeTag)): + elif filter_iname_tags_by_type(iname_tags, ConcurrentTag): nonlocal_par.append(iname) else: diff --git a/loopy/schedule/__init__.py b/loopy/schedule/__init__.py index fb0d0e2c17005ecf051d7034fd7903ed5262bdfc..e6f8d1ff394d35d2413c09ecb20ac9f86b4b646a 100644 --- a/loopy/schedule/__init__.py +++ b/loopy/schedule/__init__.py @@ -212,12 +212,12 @@ def find_loop_nest_with_map(kernel): """ result = {} - from loopy.kernel.data import ConcurrentTag, IlpBaseTag, VectorizeTag + from loopy.kernel.data import ConcurrentTag, IlpBaseTag all_nonpar_inames = set( iname for iname in kernel.all_inames() if not kernel.iname_tags_of_type(iname, - (ConcurrentTag, IlpBaseTag, VectorizeTag))) + (ConcurrentTag, IlpBaseTag))) iname_to_insns = kernel.iname_to_insns() @@ -276,7 +276,7 @@ def find_loop_insn_dep_map(kernel, loop_nest_with_map, loop_nest_around_map): result = {} - from loopy.kernel.data import ConcurrentTag, IlpBaseTag, VectorizeTag + from loopy.kernel.data import ConcurrentTag, IlpBaseTag for insn in kernel.instructions: for iname in kernel.insn_inames(insn): if kernel.iname_tags_of_type(iname, ConcurrentTag): @@ -310,7 +310,7 @@ def find_loop_insn_dep_map(kernel, loop_nest_with_map, loop_nest_around_map): continue if kernel.iname_tags_of_type(dep_insn_iname, - (ConcurrentTag, IlpBaseTag, VectorizeTag)): + (ConcurrentTag, IlpBaseTag)): # Parallel tags don't really nest, so we'll disregard # them here. continue @@ -1878,6 +1878,7 @@ def generate_loop_schedules_inner(kernel, debug_args={}): iname for iname, tags in six.iteritems(kernel.iname_to_tags) if filter_iname_tags_by_type(tags, ConcurrentTag)) + # (ConcurrentTag includes VectorizeTag) loop_nest_with_map = find_loop_nest_with_map(kernel) loop_nest_around_map = find_loop_nest_around_map(kernel)