diff --git a/MEMO b/MEMO
index d6300e1dc6cb51351d31cd5e394365dbb673e000..ebd410f4f9f333041fafb805a743cc18b6a87b56 100644
--- a/MEMO
+++ b/MEMO
@@ -42,11 +42,25 @@ To-do
- What if no universally valid precompute base index expression is found?
(test_intel_matrix_mul with n = 6*16, e.g.?)
+- "No schedule found" debug help:
+
+ - Find longest dead-end
+ - Automatically report on what hinders progress there
+
+- When duplicating, use iname aliases to relieve burden on isl
+
+- Differentiate ilp.unr from ilp.seq
+
+- Expose iname-duplicate-and-rename as a primitive.
+
- Fix all tests
Future ideas
^^^^^^^^^^^^
+- How is intra-instruction ordering of ILP loops going to be determined?
+ (taking into account that it could vary even per-instruction?)
+
- Barriers for data exchanged via global vars?
- Float4 joining on fetch/store?
diff --git a/doc/reference.rst b/doc/reference.rst
index c8844f4b82b5caf577bfa6b36680fd2cabb78b5e..71062604e2e2d6fe8462e327b07bbf3e0444a520 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -114,6 +114,11 @@ Precomputation and Prefetching
Uses :func:`extract_subst` and :func:`precompute`.
+Manipulating Reductions
+-----------------------
+
+.. autofunction:: realize_reduction
+
Finishing up
------------
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 912532ae104c3566c8b792cbb84a33db9dc3e4c7..94616f25ddab2f60d6a1b2170ec728268d71f4be 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -25,7 +25,7 @@ from loopy.kernel import ScalarArg, ArrayArg, ConstantArrayArg, ImageArg
from loopy.kernel import AutoFitLocalIndexTag, get_dot_dependency_graph, LoopKernel
from loopy.subst import extract_subst, apply_subst
from loopy.cse import precompute
-from loopy.preprocess import preprocess_kernel
+from loopy.preprocess import preprocess_kernel, realize_reduction
from loopy.schedule import generate_loop_schedules
from loopy.codegen import generate_code
from loopy.compiled import CompiledKernel, drive_timing_run, auto_test_vs_ref
diff --git a/loopy/codegen/loop.py b/loopy/codegen/loop.py
index fb94fbecaf819537d829249197bf1c9756914207..d1ffea7a395722c477b4fca0d6771adde4924443 100644
--- a/loopy/codegen/loop.py
+++ b/loopy/codegen/loop.py
@@ -149,7 +149,7 @@ def generate_unroll_loop(kernel, sched_index, codegen_state):
# }}}
-# {{{ parallel loop
+# {{{ hw-parallel loop
def set_up_hw_parallel_loops(kernel, sched_index, codegen_state, hw_inames_left=None):
from loopy.kernel import UniqueTag, HardwareParallelTag, LocalIndexTag, GroupIndexTag
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index a9e39c9fbd5581b5349833a40ef07888fe2a4345..43e93ddd0e2e77e8c4cd57b3efa8f4b71c74cd64 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -110,7 +110,25 @@ def duplicate_reduction_inames(kernel):
# {{{ rewrite reduction to imperative form
-def realize_reduction(kernel):
+def realize_reduction(kernel, insn_id_filter=None):
+ """Rewrites reductions into their imperative form. With *insn_id_filter* specified,
+ operate only on the instruction with an instruction id matching insn_id_filter.
+
+ If *insn_id_filter* is given, only the outermost level of reductions will be
+ expanded, inner reductions will be left alone (because they end up in a new
+ instruction with a different ID, which doesn't match the filter).
+
+ If *insn_id_filter* is not given, all reductions in all instructions will
+ be realized.
+
+ This routine also implicitly performs (global) reduction iname duplication,
+ if requested by '@' prefixes on any reduction iname.
+ """
+
+ # Reduction iname duplication needs to happen beforehand, and it is
+ # idempotent. So just call it now.
+ kernel = duplicate_reduction_inames(kernel)
+
new_insns = []
new_temporary_variables = kernel.temporary_variables.copy()
@@ -130,7 +148,9 @@ def realize_reduction(kernel):
ilp_iname_lengths = []
for iname in ilp_inames:
- # original kernel ok here--we're not messing with inames
+ # Using the original kernel is ok here. Nothing in realize_reductions
+ # messes with inames. This is useful because it takes advantage
+ # of bounds caching.
bounds = kernel.get_iname_bounds(iname)
from loopy.symbolic import pw_aff_to_expr
@@ -206,6 +226,10 @@ def realize_reduction(kernel):
insn = insn_queue.pop(0)
+ if insn_id_filter is not None and insn.id != insn_id_filter:
+ new_insns.append(insn)
+ continue
+
# Run reduction expansion.
new_expression = cb_mapper(insn.expression)
@@ -681,9 +705,8 @@ def preprocess_kernel(kernel):
from loopy.subst import apply_subst
kernel = apply_subst(kernel)
- kernel = mark_local_temporaries(kernel)
- kernel = duplicate_reduction_inames(kernel)
kernel = realize_reduction(kernel)
+ kernel = mark_local_temporaries(kernel)
kernel = assign_automatic_axes(kernel)
kernel = add_boostability_and_automatic_dependencies(kernel)
kernel = limit_boostability(kernel)
diff --git a/loopy/schedule.py b/loopy/schedule.py
index 64204f20da7f76800a17ce68edef2a04626d1384..9bdf9ca5c8fb6bf285a631509eea4fbd4850440d 100644
--- a/loopy/schedule.py
+++ b/loopy/schedule.py
@@ -267,6 +267,9 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
# {{{ decide about debug mode
debug_mode = False
+ #if len(schedule) == 15:
+ #debug_mode = True
+
if debug is not None:
if (debug.debug_length is not None
and len(schedule) >= debug.debug_length):
@@ -313,6 +316,8 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
# nested loop without harm.
if allow_boost:
+ # Note that the inames in 'insn.boostable_into' necessarily won't
+ # be contained in 'want'.
have = have - insn.boostable_into
if want != have:
@@ -330,10 +335,6 @@ def generate_loop_schedules_internal(kernel, loop_priority, schedule=[], allow_b
if (not schedule_now and have <= want):
reachable_insn_ids.add(insn_id)
- else:
- if debug_mode:
- print (" '%s' also not reachable because it won't work under '%s'"
- % (insn.id, ",".join(have-want)))
# }}}