diff --git a/MEMO b/MEMO
index 87725fac86ce357147af22ed60355ab30797eb9a..66f5f2f4e9ca3ba8657585ba04e6387d80e5abbd 100644
--- a/MEMO
+++ b/MEMO
@@ -55,10 +55,11 @@ To-do
- reg rolling
- nbody GPU
+ -> pending better prefetch spec
- Expose iname-duplicate-and-rename as a primitive.
-- String instructions?
+- add_prefetch gets a flag to separate out each access
- Making parameters run-time varying, substituting values that
depend on other inames?
@@ -68,6 +69,8 @@ To-do
Future ideas
^^^^^^^^^^^^
+- String instructions?
+
- How is intra-instruction ordering of ILP loops going to be determined?
(taking into account that it could vary even per-instruction?)
diff --git a/doc/reference.rst b/doc/reference.rst
index a6c3bc988409f6f037b8a86ea7ec3d05410448b1..0e538771935d373519eb074bb484550241a3ec50 100644
--- a/doc/reference.rst
+++ b/doc/reference.rst
@@ -14,6 +14,8 @@ Expressions
-----------
* `if`
+* `reductions`
+ * duplication of reduction inames
* complex-valued arithmetic
Assignments and Substitution Rules
@@ -28,13 +30,21 @@ Tag Meaning
`None` | `"for"` Sequential loop
`"l.N"` Local (intra-group) axis N
`"l.auto"` Automatically chosen local (intra-group) axis
+`"g.N"` Group-number axis N
`"unr"` Plain unrolling
-`"ilp"` Unroll using instruction-level parallelism
-`"inn"` Realize parallel iname as innermost loop
+`"ilp"` | `"ilp.unr"` Unroll using instruction-level parallelism
+`"ilp.seq"` Realize parallel iname as innermost loop
===================== ====================================================
(Throughout this table, `N` must be replaced by an actual number.)
+ILP is really three things combined:
+
+* Restricts loops to be innermost (excludes them from scheduling)
+* Duplicates reduction storage for any reductions nested around ILP usage
+* Causes a loop (unrolled or not) to be opened/generated for each
+ involved instruction
+
.. _automatic-axes:
Automatic Axis Assignment
diff --git a/loopy/check.py b/loopy/check.py
index 16b4c72845d490ce33c852334bfb0f4ec76f523d..c0bfeca134cd70a77aa78cbe1d0ffd78f75c4d7f 100644
--- a/loopy/check.py
+++ b/loopy/check.py
@@ -85,7 +85,7 @@ def check_for_inactive_iname_access(kernel):
def check_for_write_races(kernel):
from loopy.symbolic import DependencyMapper
- from loopy.kernel import ParallelTag, GroupIndexTag, IlpTag
+ from loopy.kernel import ParallelTag, GroupIndexTag, IlpBaseTag
depmap = DependencyMapper()
for insn in kernel.instructions:
@@ -136,7 +136,7 @@ def check_for_write_races(kernel):
ilp_inames = set(
iname
for iname in kernel.insn_inames(insn)
- if isinstance(kernel.iname_to_tag.get(iname), IlpTag))
+ if isinstance(kernel.iname_to_tag.get(iname), IlpBaseTag))
inames_without_write_dep = ilp_inames - (
assignee_inames & ilp_inames)
diff --git a/loopy/cse.py b/loopy/cse.py
index 3246a34e6267361f1e3d259eb18bbb05521a3fe8..82b43d25d0311bb9f8dff46e28eb8d054d7d8570 100644
--- a/loopy/cse.py
+++ b/loopy/cse.py
@@ -278,7 +278,7 @@ def precompute(kernel, subst_name, dtype, sweep_axes=[],
:arg sweep_axes: A :class:`list` of inames and/or rule argument names to be swept.
:arg storage_dims: A :class:`list` of inames and/or rule argument names/indices to be used as storage axes.
- If `storage_dims` is not specified, it defaults to the arrangement
+ If `storage_axes` is not specified, it defaults to the arrangement
`<direct sweep axes><arguments>` with the direct sweep axes being the
slower-varying indices.
@@ -318,6 +318,8 @@ def precompute(kernel, subst_name, dtype, sweep_axes=[],
from loopy.symbolic import SubstitutionCallbackMapper
scm = SubstitutionCallbackMapper([(subst_name, subst_instance)], gather_substs)
+ # We need to work on the fully expanded form of an expression.
+ # To that end, instantiate a substitutor.
from loopy.symbolic import ParametrizedSubstitutor
rules_except_mine = kernel.substitutions.copy()
del rules_except_mine[subst_name]
@@ -328,7 +330,8 @@ def precompute(kernel, subst_name, dtype, sweep_axes=[],
# arguments. Therefore, fully expand each instruction and look at
# the invocations in subst_name occurring there.
- scm(subst_expander(insn.expression))
+ expanded_expr = subst_expander(insn.expression)
+ scm(expanded_expr)
if not invocation_descriptors:
raise RuntimeError("no invocations of '%s' found" % subst_name)
diff --git a/loopy/kernel.py b/loopy/kernel.py
index 4ca9b5540b05c0abb0808d7933b365d3cb023b9b..8b5010c44964b09d31228a327f3ad61d2d2b4349 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -65,9 +65,16 @@ class AutoFitLocalIndexTag(AutoLocalIndexTagBase):
def __str__(self):
return "l.auto"
-class IlpTag(ParallelTag):
+class IlpBaseTag(ParallelTag):
+ pass
+
+class UnrolledIlpTag(IlpBaseTag):
+ def __str__(self):
+ return "ilp.unr"
+
+class LoopedIlpTag(IlpBaseTag):
def __str__(self):
- return "ilp"
+ return "ilp.seq"
class UnrollTag(IndexTag):
def __str__(self):
@@ -87,8 +94,10 @@ def parse_tag(tag):
return None
elif tag in ["unr"]:
return UnrollTag()
- elif tag == "ilp":
- return IlpTag()
+ elif tag in ["ilp", "ilp.unr"]:
+ return UnrolledIlpTag()
+ elif tag == "ilp.seq":
+ return LoopedIlpTag()
elif tag.startswith("g."):
return GroupIndexTag(int(tag[2:]))
elif tag.startswith("l."):
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 43e93ddd0e2e77e8c4cd57b3efa8f4b71c74cd64..40c702fb9dbe90262729dc1696383d33e2b9e172 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -132,17 +132,17 @@ def realize_reduction(kernel, insn_id_filter=None):
new_insns = []
new_temporary_variables = kernel.temporary_variables.copy()
- from loopy.kernel import IlpTag
+ from loopy.kernel import IlpBaseTag
def map_reduction(expr, rec):
# Only expand one level of reduction at a time, going from outermost to
# innermost. Otherwise we get the (iname + insn) dependencies wrong.
- # {{{ see if this reduction is nested inside some ILP loops
+ # {{{ see if this reduction is nested around some ILP loops
ilp_inames = [iname
for iname in temp_kernel.insn_inames(insn)
- if isinstance(temp_kernel.iname_to_tag.get(iname), IlpTag)]
+ if isinstance(temp_kernel.iname_to_tag.get(iname), IlpBaseTag)]
from loopy.isl_helpers import static_max_of_pw_aff
diff --git a/test/test_fem_assembly.py b/test/test_fem_assembly.py
index 1c42082f2b9a77db3fa38f8f9f34822b84b0c9da..c90b556bd6a8e8265451bc4ed7ca41543654091c 100644
--- a/test/test_fem_assembly.py
+++ b/test/test_fem_assembly.py
@@ -83,7 +83,7 @@ def test_laplacian_stiffness(ctx_factory):
knl = lp.split_dimension(knl, "K", Ncloc,
outer_iname="Ko", inner_iname="Kloc")
knl = lp.precompute(knl, "dPsi.one", np.float32, ["dx_axis"], default_tag=None)
- knl = lp.tag_dimensions(knl, {"j": "ilp"})
+ knl = lp.tag_dimensions(knl, {"j": "ilp.seq"})
return knl, ["Ko", "Kloc"]
@@ -124,7 +124,7 @@ def test_laplacian_stiffness(ctx_factory):
# Plug in variant name here
# |
# v
- for variant in [variant_simple_gpu_prefetch]:
+ for variant in [variant_fig33]:
var_knl, loop_prio = variant(knl)
kernel_gen = lp.generate_loop_schedules(var_knl,
loop_priority=loop_prio)