From 607f74af71e4f4c7d3eb75c01a60d3f05fe00ab4 Mon Sep 17 00:00:00 2001
From: Matt Wala <wala1@illinois.edu>
Date: Wed, 6 Dec 2017 15:25:12 -0600
Subject: [PATCH 1/2] Dep resolution: Speed up resolution in the common case
 where the dependency is not a wildcard.

Related: #110

Also fixes treatment for self-dependencies in wildcards (closes #111).
---
 loopy/kernel/creation.py | 15 +++++++++++++--
 test/test_loopy.py       | 21 +++++++++++++++++++++
 2 files changed, 34 insertions(+), 2 deletions(-)

diff --git a/loopy/kernel/creation.py b/loopy/kernel/creation.py
index fb935476d..c1236ba90 100644
--- a/loopy/kernel/creation.py
+++ b/loopy/kernel/creation.py
@@ -1656,6 +1656,13 @@ def apply_default_order_to_args(kernel, default_order):
 
 # {{{ resolve instruction dependencies
 
+WILDCARD_SYMBOLS = "*?["
+
+
+def _is_wildcard(s):
+    return any(c in s for c in WILDCARD_SYMBOLS)
+
+
 def _resolve_dependencies(knl, insn, deps):
     from loopy import find_instructions
     from loopy.match import MatchExpressionBase
@@ -1670,12 +1677,16 @@ def _resolve_dependencies(knl, insn, deps):
                 if new_dep.id != insn.id:
                     new_deps.append(new_dep.id)
                     found_any = True
-        else:
+        elif _is_wildcard(dep):
             from fnmatch import fnmatchcase
             for other_insn in knl.instructions:
-                if fnmatchcase(other_insn.id, dep):
+                if other_insn.id != insn.id and fnmatchcase(other_insn.id, dep):
                     new_deps.append(other_insn.id)
                     found_any = True
+        else:
+            if dep in knl.id_to_insn:
+                new_deps.append(dep)
+                found_any = True
 
         if not found_any and knl.options.check_dep_resolution:
             raise LoopyError("instruction '%s' declared a depency on '%s', "
diff --git a/test/test_loopy.py b/test/test_loopy.py
index d6f2d6df6..7b875ef08 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -2568,6 +2568,27 @@ def test_execution_backend_can_cache_dtypes(ctx_factory):
     knl(queue)
 
 
+def test_wildcard_dep_matching():
+    knl = lp.make_kernel(
+            "{[i]: 0 <= i < 10}",
+            """
+            <>a = 0 {id=insn1}
+            <>b = 0 {id=insn2,dep=insn?}
+            <>c = 0 {id=insn3,dep=insn*}
+            <>d = 0 {id=insn4,dep=insn[12]}
+            <>e = 0 {id=insn5,dep=insn[!1]}
+            """,
+            "...")
+
+    all_insns = set("insn%d" % i for i in range(1,6))
+
+    assert knl.id_to_insn["insn1"].depends_on == set()
+    assert knl.id_to_insn["insn2"].depends_on == all_insns - set(["insn2"])
+    assert knl.id_to_insn["insn3"].depends_on == all_insns - set(["insn3"])
+    assert knl.id_to_insn["insn4"].depends_on == set(["insn1", "insn2"])
+    assert knl.id_to_insn["insn5"].depends_on == all_insns - set(["insn1", "insn5"])
+
+
 def test_preamble_with_separate_temporaries(ctx_factory):
     from loopy.kernel.data import temp_var_scope as scopes
     # create a function mangler
-- 
GitLab


From 4fac320bd1a8bf97a6ea3b38d79723e5a7a778f0 Mon Sep 17 00:00:00 2001
From: Matt Wala <wala1@illinois.edu>
Date: Wed, 6 Dec 2017 16:38:32 -0500
Subject: [PATCH 2/2] Flake8 fix

---
 test/test_loopy.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/test/test_loopy.py b/test/test_loopy.py
index 7b875ef08..79d849727 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -2580,7 +2580,7 @@ def test_wildcard_dep_matching():
             """,
             "...")
 
-    all_insns = set("insn%d" % i for i in range(1,6))
+    all_insns = set("insn%d" % i for i in range(1, 6))
 
     assert knl.id_to_insn["insn1"].depends_on == set()
     assert knl.id_to_insn["insn2"].depends_on == all_insns - set(["insn2"])
-- 
GitLab