diff --git a/loopy/codegen/control.py b/loopy/codegen/control.py
index 7b73093873e7a5364945f8888723be639f37dcf4..fbe4e4c0390a70080e80889e85984a56add4edf3 100644
--- a/loopy/codegen/control.py
+++ b/loopy/codegen/control.py
@@ -75,7 +75,7 @@ def generate_code_for_sched_index(codegen_state, sched_index):
 
         from loopy.codegen.result import generate_host_or_device_program
         codegen_result = generate_host_or_device_program(
-                new_codegen_state, sched_index + 1)
+                new_codegen_state, sched_index)
 
         glob_grid, loc_grid = kernel.get_grid_sizes_for_insn_ids_as_exprs(
                 get_insn_ids_for_block_at(kernel.schedule, sched_index))
diff --git a/loopy/codegen/loop.py b/loopy/codegen/loop.py
index a9eb44f84d44015fa11ce83afe410e833bb8e214..648c3fe6f5b748dcc47de5ac972bb82ce605a9a9 100644
--- a/loopy/codegen/loop.py
+++ b/loopy/codegen/loop.py
@@ -233,17 +233,23 @@ def set_up_hw_parallel_loops(codegen_state, schedule_index, next_func,
     from loopy.kernel.data import (
             UniqueTag, HardwareParallelTag, LocalIndexTag, GroupIndexTag)
 
+    from loopy.schedule import get_insn_ids_for_block_at
+    insn_ids_for_block = get_insn_ids_for_block_at(kernel.schedule, schedule_index)
+
     if hw_inames_left is None:
+        all_inames_by_insns = set()
+        for insn_id in insn_ids_for_block:
+            all_inames_by_insns |= kernel.insn_inames(insn_id)
+
         hw_inames_left = [iname
-                for iname in kernel.all_inames()
+                for iname in all_inames_by_insns
                 if isinstance(kernel.iname_to_tag.get(iname), HardwareParallelTag)]
 
     if not hw_inames_left:
         return next_func(codegen_state)
 
-    from loopy.schedule import get_insn_ids_for_block_at
     global_size, local_size = kernel.get_grid_sizes_for_insn_ids(
-            get_insn_ids_for_block_at(kernel.schedule, schedule_index))
+            insn_ids_for_block)
 
     hw_inames_left = hw_inames_left[:]
     iname = hw_inames_left.pop()
diff --git a/loopy/codegen/result.py b/loopy/codegen/result.py
index 3306d30e4486418b7b4f78e4f1d95a4fd39b45bc..a46500ea515974b00b5bfce4faadfadd010106a4 100644
--- a/loopy/codegen/result.py
+++ b/loopy/codegen/result.py
@@ -246,14 +246,17 @@ def generate_host_or_device_program(codegen_state, schedule_index):
     from functools import partial
 
     from loopy.codegen.control import build_loop_nest
-    next_func = partial(build_loop_nest, schedule_index=schedule_index)
-
     if codegen_state.is_generating_device_code:
+        from loopy.schedule import CallKernel
+        assert isinstance(codegen_state.kernel.schedule[schedule_index], CallKernel)
+
         from loopy.codegen.loop import set_up_hw_parallel_loops
         codegen_result = set_up_hw_parallel_loops(
-                codegen_state, schedule_index, next_func=next_func)
+                codegen_state, schedule_index,
+                next_func=partial(build_loop_nest,
+                    schedule_index=schedule_index + 1))
     else:
-        codegen_result = next_func(codegen_state)
+        codegen_result = build_loop_nest(codegen_state, schedule_index)
 
     codegen_result = merge_codegen_results(
             codegen_state,