From 0a9f6ed2f245aa93dae08042d3fa5353728c7185 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Sat, 5 Nov 2011 14:04:30 -0500
Subject: [PATCH] Add support for constant array arguments.
---
loopy/__init__.py | 2 +-
loopy/codegen/__init__.py | 15 ++++++++++-----
loopy/cse.py | 4 ++--
loopy/kernel.py | 3 ++-
4 files changed, 15 insertions(+), 9 deletions(-)
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 843356bf3..379f1c3b8 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -20,7 +20,7 @@ class LoopyAdvisory(UserWarning):
# {{{ imported user interface
-from loopy.kernel import ScalarArg, ArrayArg, ImageArg
+from loopy.kernel import ScalarArg, ArrayArg, ConstantArrayArg, ImageArg
from loopy.kernel import AutoFitLocalIndexTag, get_dot_dependency_graph
from loopy.cse import realize_cse
diff --git a/loopy/codegen/__init__.py b/loopy/codegen/__init__.py
index dd4b4dec5..fb3e00bd3 100644
--- a/loopy/codegen/__init__.py
+++ b/loopy/codegen/__init__.py
@@ -183,7 +183,7 @@ def generate_code(kernel):
POD, Value, ArrayOf, Module, Block,
Line, Const, LiteralLines, Initializer)
- from cgen.opencl import (CLKernel, CLGlobal, CLRequiredWorkGroupSize,
+ from cgen.opencl import (CLKernel, CLGlobal, CLConstant, CLRequiredWorkGroupSize,
CLLocal, CLImage, CLConstant)
from loopy.symbolic import LoopyCCodeMapper
@@ -207,11 +207,11 @@ def generate_code(kernel):
has_double = False
has_image = False
- from loopy.kernel import ArrayArg, ImageArg
+ from loopy.kernel import ArrayArg, ConstantArrayArg, ImageArg, ScalarArg
args = []
for arg in kernel.args:
- if isinstance(arg, ArrayArg):
+ if isinstance(arg, (ConstantArrayArg, ArrayArg)):
arg_decl = restrict_ptr_if_not_nvidia(
POD(arg.dtype, arg.name))
if arg_decl.name not in kernel.get_written_variables():
@@ -219,7 +219,10 @@ def generate_code(kernel):
arg_decl = CLConstant(Const(arg_decl))
else:
arg_decl = Const(arg_decl)
- arg_decl = CLGlobal(arg_decl)
+ if isinstance(arg, ConstantArrayArg):
+ arg_decl = CLConstant(arg_decl)
+ else:
+ arg_decl = CLGlobal(arg_decl)
elif isinstance(arg, ImageArg):
if arg.name in kernel.get_written_variables():
mode = "w"
@@ -229,8 +232,10 @@ def generate_code(kernel):
arg_decl = CLImage(arg.dimensions, mode, arg.name)
has_image = True
- else:
+ elif isinstance(arg, ScalarArg):
arg_decl = Const(POD(arg.dtype, arg.name))
+ else:
+ raise ValueError("argument type not understood: '%s'" % type(arg))
if arg.dtype in [np.float64, np.complex128]:
has_double = True
diff --git a/loopy/cse.py b/loopy/cse.py
index ab576a701..2fe4340ab 100644
--- a/loopy/cse.py
+++ b/loopy/cse.py
@@ -342,8 +342,8 @@ def realize_cse(kernel, cse_tag, dtype, independent_inames=[],
if not set(independent_inames) <= kernel.all_inames():
raise ValueError("In CSE realization for '%s': "
- "cannot make inames '%s' independent--"
- "they don't already exist" % (
+ "cannot make iname(s) '%s' independent--"
+ "it/they don't already exist" % (
cse_tag,
",".join(
set(independent_inames)-kernel.all_inames())))
diff --git a/loopy/kernel.py b/loopy/kernel.py
index 6af2b8726..408fe45fa 100644
--- a/loopy/kernel.py
+++ b/loopy/kernel.py
@@ -156,7 +156,8 @@ class ArrayArg:
def __repr__(self):
return "<ArrayArg '%s' of type %s>" % (self.name, self.dtype)
-
+class ConstantArrayArg(ArrayArg):
+ pass
class ImageArg:
def __init__(self, name, dtype, dimensions=None, shape=None):
--
GitLab