diff --git a/loopy/codegen/bounds.py b/loopy/codegen/bounds.py index 442985e6a624a1980bbc6ea85d8a43218dad8cbd..61f4b3a9b8c38dfc25ebc81243812aa963423f8a 100644 --- a/loopy/codegen/bounds.py +++ b/loopy/codegen/bounds.py @@ -63,23 +63,21 @@ def get_usable_inames_for_conditional(kernel, sched_index): result = find_active_inames_at(kernel, sched_index) crosses_barrier = has_barrier_within(kernel, sched_index) - # Find our containing subkernel, grab inames for all insns from there. - + # Find our containing subkernel. Grab inames for all insns from there. within_subkernel = False - for prev_sched_index, sched_item in enumerate(kernel.schedule): - if prev_sched_index == sched_index: - if not within_subkernel: - # Outside all subkernels - use only inames available to host. - return frozenset(result) - + for sched_item_index, sched_item in enumerate(kernel.schedule[:sched_index+1]): from loopy.schedule import CallKernel, ReturnFromKernel if isinstance(sched_item, CallKernel): within_subkernel = True - subkernel_index = prev_sched_index + subkernel_index = sched_item_index elif isinstance(sched_item, ReturnFromKernel): within_subkernel = False + if not within_subkernel: + # Outside all subkernels - use only inames available to host. + return frozenset(result) + insn_ids_for_subkernel = get_insn_ids_for_block_at( kernel.schedule, subkernel_index)