From ef60e3273fdcd5658d799c36d85b8c0889e7fe99 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Sun, 5 Jul 2015 13:46:44 -0500
Subject: [PATCH] Compatibility with POCL, work around broken POCL calling
conventions
---
.gitlab-ci.yml | 12 +++
loopy/compiled.py | 148 +++++++++++++++++++++++++-----
loopy/target/pyopencl/__init__.py | 9 +-
test/test_dg.py | 27 +++---
test/test_linalg.py | 12 +++
test/test_loopy.py | 4 +
6 files changed, 175 insertions(+), 37 deletions(-)
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index fc5a1ca68..e71ea2c6d 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -23,6 +23,18 @@ Python 3.4 AMD CPU:
- amd-cl-cpu
except:
- tags
+Python 2.7 POCL:
+ script:
+ - export PY_EXE=python2.7
+ - export PYOPENCL_TEST=portable
+ - export EXTRA_INSTALL="numpy mako"
+ - curl -L -O -k https://gitlab.tiker.net/inducer/ci-support/raw/master/build-and-test-py-project.sh
+ - ". ./build-and-test-py-project.sh"
+ tags:
+ - python2.7
+ - pocl
+ except:
+ - tags
# PyPy AMD CPU:
# script:
# - export PY_EXE=pypy
diff --git a/loopy/compiled.py b/loopy/compiled.py
index c5928d6e7..da659eaba 100644
--- a/loopy/compiled.py
+++ b/loopy/compiled.py
@@ -27,6 +27,7 @@ THE SOFTWARE.
"""
+import sys
import numpy as np
from pytools import Record, memoize_method
from loopy.diagnostic import ParameterFinderWarning
@@ -296,46 +297,141 @@ def generate_integer_arg_finding_from_strides(gen, kernel, impl_arg_info, option
# {{{ value arg setup
-def generate_value_arg_setup(gen, kernel, impl_arg_info, options):
+def generate_value_arg_setup(gen, kernel, cl_kernel, impl_arg_info, options):
import loopy as lp
from loopy.kernel.array import ArrayBase
+ # {{{ arg counting bug handling
+
+ # For example:
+ # https://github.com/pocl/pocl/issues/197
+ # (but Apple CPU has a similar bug)
+
+ work_around_arg_count_bug = False
+ warn_about_arg_count_bug = False
+
+ from pyopencl.characterize import has_struct_arg_count_bug
+
+ devices = cl_kernel.context.devices
+
+ count_bug_per_dev = [
+ has_struct_arg_count_bug(dev)
+ for dev in devices]
+
+ if any(count_bug_per_dev):
+ if all(count_bug_per_dev):
+ work_around_arg_count_bug = True
+ else:
+ warn_about_arg_count_bug = True
+
+ # }}}
+
+ cl_arg_idx = 0
+ arg_idx_to_cl_arg_idx = {}
+
+ fp_arg_count = 0
+
for arg_idx, arg in enumerate(impl_arg_info):
+ arg_idx_to_cl_arg_idx[arg_idx] = cl_arg_idx
+
if arg.arg_class is not lp.ValueArg:
assert issubclass(arg.arg_class, ArrayBase)
+
+ # assume each of those generates exactly one...
+ cl_arg_idx += 1
+
continue
gen("# {{{ process %s" % arg.name)
gen("")
if not options.skip_arg_checks:
- gen("if %s is None:" % arg.name)
- with Indentation(gen):
- gen("raise RuntimeError(\"input argument '%s' must "
- "be supplied\")" % arg.name)
- gen("")
-
- if arg.dtype.kind == "i":
- gen("# cast to int to avoid numpy scalar trouble with Boost.Python")
- gen("%s = int(%s)" % (arg.name, arg.name))
+ gen("""
+ if {name} is None:
+ raise RuntimeError("input argument '{name}' must "
+ "be supplied")
+ """.format(name=arg.name))
+
+ if sys.version_info < (2, 7) and arg.dtype.kind == "i":
+ gen("# cast to long to avoid trouble with struct packing")
+ gen("%s = long(%s)" % (arg.name, arg.name))
gen("")
if arg.dtype.char == "V":
- gen("cl_kernel.set_arg(%d, %s)" % (arg_idx, arg.name))
+ gen("cl_kernel.set_arg(%d, %s)" % (cl_arg_idx, arg.name))
+ cl_arg_idx += 1
+
+ elif arg.dtype.kind == "c":
+ if warn_about_arg_count_bug:
+ from warnings import warn
+ warn("{knl_name}: arguments include complex numbers, and "
+ "some (but not all) of the target devices mishandle "
+ "struct kernel arguments (hence the workaround is "
+ "disabled".format(
+ knl_name=kernel.name))
+
+ if arg.dtype == np.complex64:
+ arg_char = "f"
+ elif arg.dtype == np.complex128:
+ arg_char = "d"
+ else:
+ raise TypeError("unexpected complex type: %s" % arg.dtype)
+
+ if (work_around_arg_count_bug
+ and arg.dtype == np.complex128
+ and fp_arg_count + 2 <= 8):
+ gen(
+ "buf = _lpy_pack('{arg_char}', {arg_var}.real)"
+ .format(arg_char=arg_char, arg_var=arg.name))
+ gen(
+ "cl_kernel.set_arg({cl_arg_idx}, buf)"
+ .format(cl_arg_idx=cl_arg_idx))
+ cl_arg_idx += 1
+
+ gen(
+ "buf = _lpy_pack('{arg_char}', {arg_var}.imag)"
+ .format(arg_char=arg_char, arg_var=arg.name))
+ gen(
+ "cl_kernel.set_arg({cl_arg_idx}, buf)"
+ .format(cl_arg_idx=cl_arg_idx))
+ cl_arg_idx += 1
+ else:
+ gen(
+ "buf = _lpy_pack('{arg_char}{arg_char}', "
+ "{arg_var}.real, {arg_var}.imag)"
+ .format(arg_char=arg_char, arg_var=arg.name))
+ gen(
+ "cl_kernel.set_arg({cl_arg_idx}, buf)"
+ .format(cl_arg_idx=cl_arg_idx))
+ cl_arg_idx += 1
+
+ fp_arg_count += 2
+
else:
- gen("cl_kernel.set_arg(%d, _lpy_pack(\"%s\", %s))"
- % (arg_idx, arg.dtype.char, arg.name))
+ if arg.dtype.kind == "f":
+ fp_arg_count += 1
+
+ gen("cl_kernel.set_arg(%d, _lpy_pack('%s', %s))"
+ % (cl_arg_idx, arg.dtype.char, arg.name))
+
+ cl_arg_idx += 1
+
gen("")
gen("# }}}")
gen("")
+ assert cl_arg_idx == cl_kernel.num_args
+
+ return arg_idx_to_cl_arg_idx
+
# }}}
# {{{ array arg setup
-def generate_array_arg_setup(gen, kernel, impl_arg_info, options):
+def generate_array_arg_setup(gen, kernel, impl_arg_info, options,
+ arg_idx_to_cl_arg_idx):
import loopy as lp
from loopy.kernel.array import ArrayBase
@@ -356,12 +452,12 @@ def generate_array_arg_setup(gen, kernel, impl_arg_info, options):
is_written = arg.base_name in kernel.get_written_variables()
kernel_arg = kernel.impl_arg_to_arg.get(arg.name)
- gen("# {{{ process %s" % arg.name)
- gen("")
-
if not issubclass(arg.arg_class, ArrayBase):
continue
+ gen("# {{{ process %s" % arg.name)
+ gen("")
+
if not options.no_numpy:
gen("if isinstance(%s, _lpy_np.ndarray):" % arg.name)
with Indentation(gen):
@@ -552,10 +648,12 @@ def generate_array_arg_setup(gen, kernel, impl_arg_info, options):
gen("del _lpy_made_by_loopy")
gen("")
+ cl_arg_idx = arg_idx_to_cl_arg_idx[arg_idx]
+
if arg.arg_class in [lp.GlobalArg, lp.ConstantArg]:
- gen("cl_kernel.set_arg(%d, %s.base_data)" % (arg_idx, arg.name))
+ gen("cl_kernel.set_arg(%d, %s.base_data)" % (cl_arg_idx, arg.name))
else:
- gen("cl_kernel.set_arg(%d, %s)" % (arg_idx, arg.name))
+ gen("cl_kernel.set_arg(%d, %s)" % (cl_arg_idx, arg.name))
gen("")
gen("# }}}")
@@ -567,7 +665,7 @@ def generate_array_arg_setup(gen, kernel, impl_arg_info, options):
# }}}
-def generate_invoker(kernel, impl_arg_info, options):
+def generate_invoker(kernel, cl_kernel, impl_arg_info, options):
system_args = [
"cl_kernel", "queue", "allocator=None", "wait_for=None",
# ignored if options.no_numpy
@@ -584,7 +682,7 @@ def generate_invoker(kernel, impl_arg_info, options):
gen.add_to_preamble("import pyopencl.array as _lpy_cl_array")
gen.add_to_preamble("import pyopencl.tools as _lpy_cl_tools")
gen.add_to_preamble("import numpy as _lpy_np")
- gen.add_to_preamble("from pyopencl._pvt_struct import pack as _lpy_pack")
+ gen.add_to_preamble("from struct import pack as _lpy_pack")
gen.add_to_preamble("")
gen("if allocator is None:")
@@ -596,8 +694,10 @@ def generate_invoker(kernel, impl_arg_info, options):
generate_integer_arg_finding_from_offsets(gen, kernel, impl_arg_info, options)
generate_integer_arg_finding_from_strides(gen, kernel, impl_arg_info, options)
- generate_value_arg_setup(gen, kernel, impl_arg_info, options)
- generate_array_arg_setup(gen, kernel, impl_arg_info, options)
+ arg_idx_to_cl_arg_idx = \
+ generate_value_arg_setup(gen, kernel, cl_kernel, impl_arg_info, options)
+ generate_array_arg_setup(gen, kernel, impl_arg_info, options,
+ arg_idx_to_cl_arg_idx)
# {{{ generate invocation
@@ -763,7 +863,7 @@ class CompiledKernel:
cl_kernel=cl_kernel,
impl_arg_info=impl_arg_info,
invoker=generate_invoker(
- kernel, impl_arg_info, self.kernel.options))
+ kernel, cl_kernel, impl_arg_info, self.kernel.options))
# {{{ debugging aids
diff --git a/loopy/target/pyopencl/__init__.py b/loopy/target/pyopencl/__init__.py
index 174506cd6..ee9366800 100644
--- a/loopy/target/pyopencl/__init__.py
+++ b/loopy/target/pyopencl/__init__.py
@@ -272,8 +272,13 @@ class PyOpenCLTarget(OpenCLTarget):
return vec.types[base, count]
def alignment_requirement(self, type_decl):
- import pyopencl._pvt_struct as _struct
- return _struct.calcsize(type_decl.struct_format())
+ import struct
+
+ fmt = (type_decl.struct_format()
+ .replace("F", "ff")
+ .replace("D", "dd"))
+
+ return struct.calcsize(fmt)
# }}}
diff --git a/test/test_dg.py b/test/test_dg.py
index 581562da8..0eb5be224 100644
--- a/test/test_dg.py
+++ b/test/test_dg.py
@@ -43,10 +43,10 @@ def test_dg_volume(ctx_factory):
order = "F"
- N = 3
- Np = (N+1)*(N+2)*(N+3)//6
+ N = 3 # noqa
+ Np = (N+1)*(N+2)*(N+3)//6 # noqa
- K = 10000
+ K = 10000 # noqa
knl = lp.make_kernel([
"{[n,m,k]: 0<= n,m < Np and 0<= k < K}",
@@ -146,16 +146,21 @@ def test_dg_volume(ctx_factory):
parameters_dict = dict(K=K)
- for variant in [
+ variants = [
variant_basic,
variant_more_per_work_group,
- variant_image_d,
variant_prefetch_d,
variant_prefetch_fields,
variant_k_ilp,
variant_simple_padding,
variant_fancy_padding
- ]:
+ ]
+
+ if (ctx.devices[0].image_support
+ and ctx.devices[0].platform.name != "Portable Computing Language"):
+ variants.append(variant_image_d)
+
+ for variant in variants:
lp.auto_test_vs_ref(
seq_knl, ctx, variant(knl), parameters=parameters_dict,
#codegen_kwargs=dict(with_annotation=True)
@@ -169,12 +174,12 @@ def no_test_dg_surface(ctx_factory):
order = "F"
- N = 3
- Np = (N+1)*(N+2)*(N+3)//6
- Nfp = (N+1)*(N+2)//2
- Nfaces = 4
+ N = 3 # noqa
+ Np = (N+1)*(N+2)*(N+3)//6 # noqa
+ Nfp = (N+1)*(N+2)//2 # noqa
+ Nfaces = 4 # noqa
- K = 10000
+ K = 10000 # noqa
knl = lp.make_kernel(
[
diff --git a/test/test_linalg.py b/test/test_linalg.py
index 9dbedf320..c019eb67f 100644
--- a/test/test_linalg.py
+++ b/test/test_linalg.py
@@ -422,6 +422,10 @@ def test_magma_fermi_matrix_mul(ctx_factory):
n = get_suitable_size(ctx)
+ if (not ctx.devices[0].image_support
+ or ctx.devices[0].platform.name == "Portable Computing Language"):
+ pytest.skip("crashes on pocl")
+
image_format = cl.ImageFormat(cl.channel_order.R, cl.channel_type.FLOAT)
if image_format not in cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
@@ -469,6 +473,10 @@ def test_image_matrix_mul(ctx_factory):
n = get_suitable_size(ctx)
+ if (not ctx.devices[0].image_support
+ or ctx.devices[0].platform.name == "Portable Computing Language"):
+ pytest.skip("crashes on pocl")
+
image_format = cl.ImageFormat(cl.channel_order.R, cl.channel_type.FLOAT)
if image_format not in cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
@@ -505,6 +513,10 @@ def test_image_matrix_mul_ilp(ctx_factory):
ctx = ctx_factory()
order = "C"
+ if (not ctx.devices[0].image_support
+ or ctx.devices[0].platform.name == "Portable Computing Language"):
+ pytest.skip("crashes on pocl")
+
image_format = cl.ImageFormat(cl.channel_order.R, cl.channel_type.FLOAT)
if image_format not in cl.get_supported_image_formats(
ctx, cl.mem_flags.READ_ONLY, cl.mem_object_type.IMAGE2D):
diff --git a/test/test_loopy.py b/test/test_loopy.py
index 22c0ce47c..1fa35101d 100644
--- a/test/test_loopy.py
+++ b/test/test_loopy.py
@@ -379,6 +379,10 @@ def test_stencil_with_overfetch(ctx_factory):
],
assumptions="n>=1")
+ if ctx.devices[0].platform.name == "Portable Computing Language":
+ # https://github.com/pocl/pocl/issues/205
+ pytest.skip("takes very long to compile on pocl")
+
knl = lp.add_and_infer_dtypes(knl, dict(a=np.float32))
ref_knl = knl
--
GitLab