From 602db4a829abfb7a9f813dae55ceb52b5611b04c Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Sun, 5 Jul 2015 00:29:34 -0500
Subject: [PATCH] Code generation for kernel enqueue and work around pocl
complex arg split issue
---
pyopencl/__init__.py | 342 ++++++++++++++++++++++++++++++++++-------
pyopencl/cffi_cl.py | 30 ++--
setup.py | 2 +-
test/test_algorithm.py | 3 -
test/test_array.py | 8 -
test/test_clmath.py | 13 +-
6 files changed, 305 insertions(+), 93 deletions(-)
diff --git a/pyopencl/__init__.py b/pyopencl/__init__.py
index aa8abdd7..0e7412b1 100644
--- a/pyopencl/__init__.py
+++ b/pyopencl/__init__.py
@@ -42,6 +42,8 @@ except ImportError:
"its source directory. This likely won't work.")
raise
+_CPY2 = _cl._CPY2
+_CPY26 = _cl._CPY2 and sys.version_info < (2, 7)
import numpy as np
@@ -167,6 +169,8 @@ CONSTANT_CLASSES = [
and name[0].islower() and name not in ["zip", "map", "range"]]
+# {{{ diagnostics
+
class CompilerWarning(UserWarning):
pass
@@ -185,6 +189,25 @@ def compiler_output(text):
class _ErrorRecord(_Record):
pass
+# }}}
+
+
+# {{{ arg packing helpers
+
+_size_t_char = ({
+ 8: 'Q',
+ 4: 'L',
+ 2: 'H',
+ 1: 'B',
+})[_cl._ffi.sizeof('size_t')]
+_type_char_map = {
+ 'n': _size_t_char.lower(),
+ 'N': _size_t_char
+}
+del _size_t_char
+
+# }}}
+
# {{{ find pyopencl shipped source code
@@ -609,72 +632,260 @@ def _add_functionality():
kernel_old_init(self, prg, name)
self._source = getattr(prg, "_source", None)
- def kernel_call(self, queue, global_size, local_size, *args, **kwargs):
- global_offset = kwargs.pop("global_offset", None)
- g_times_l = kwargs.pop("g_times_l", False)
- wait_for = kwargs.pop("wait_for", None)
+ self._generate_naive_call()
- if kwargs:
- raise TypeError(
- "Kernel.__call__ recived unexpected keyword arguments: %s"
- % ", ".join(list(kwargs.keys())))
+ # {{{ code generation for __call__, set_args
- self.set_args(*args)
+ def kernel__set_set_args_body(self, body, num_passed_args):
+ from pytools.py_codegen import (
+ PythonFunctionGenerator,
+ PythonCodeGenerator,
+ Indentation)
- return enqueue_nd_range_kernel(queue, self, global_size, local_size,
- global_offset, wait_for, g_times_l=g_times_l)
+ arg_names = ["arg%d" % i for i in xrange(num_passed_args)]
- def kernel_set_scalar_arg_dtypes(self, arg_dtypes):
- assert len(arg_dtypes) == self.num_args, (
- "length of argument type array (%d) and "
- "CL-generated number of arguments (%d) do not agree"
- % (len(arg_dtypes), self.num_args))
+ # {{{ wrap in error handler
- arg_type_chars = []
+ err_gen = PythonCodeGenerator()
- for arg_dtype in arg_dtypes:
- if arg_dtype is None:
- arg_type_chars.append(None)
+ err_gen("try:")
+ with Indentation(err_gen):
+ err_gen.extend(body)
+ err_gen("except TypeError as e:")
+ with Indentation(err_gen):
+ err_gen("""
+ if current_arg is not None:
+ args = [{args}]
+ advice = ""
+ from pyopencl.array import Array
+ if isinstance(args[current_arg], Array):
+ advice = " (perhaps you meant to pass 'array.data' " \
+ "instead of the array itself?)"
+
+ raise _cl.LogicError(
+ "when processing argument #%d (1-based): %s%s"
+ % (current_arg+1, str(e), advice))
+ else:
+ raise
+ """
+ .format(args=", ".join(arg_names)))
+ err_gen("")
+
+ # }}}
+
+ def add_preamble(gen):
+ gen.add_to_preamble(
+ "import numpy as np")
+ gen.add_to_preamble(
+ "import pyopencl.cffi_cl as _cl")
+ gen.add_to_preamble(
+ "from pyopencl.cffi_cl import _lib, "
+ "_ffi, _handle_error, _CLKernelArg")
+ gen.add_to_preamble("from pyopencl import status_code")
+ gen.add_to_preamble("from struct import pack")
+ gen.add_to_preamble("")
+
+ # {{{ generate _enqueue
+
+ gen = PythonFunctionGenerator("enqueue_knl_%s" % self.function_name,
+ ["self", "queue", "global_size", "local_size"]
+ + arg_names
+ + ["global_offset=None", "g_times_l=None", "wait_for=None"])
+
+ add_preamble(gen)
+ gen.extend(err_gen)
+
+ gen("""
+ return _cl.enqueue_nd_range_kernel(queue, self, global_size, local_size,
+ global_offset, wait_for, g_times_l=g_times_l)
+ """)
+
+ self._enqueue = gen.get_function()
+
+ # }}}
+
+ # {{{ generate set_args
+
+ gen = PythonFunctionGenerator("_set_args", ["self"] + arg_names)
+
+ add_preamble(gen)
+ gen.extend(err_gen)
+
+ self._set_args = gen.get_function()
+
+ # }}}
+
+ def kernel__generate_buffer_arg_setter(self, gen, arg_idx, buf_var):
+ from pytools.py_codegen import Indentation
+
+ if _CPY2:
+ # https://github.com/numpy/numpy/issues/5381
+ gen("if isinstance({buf_var}, np.generic):".format(buf_var=buf_var))
+ with Indentation(gen):
+ gen("{buf_var} = np.getbuffer({buf_var})".format(buf_var=buf_var))
+
+ gen("""
+ c_buf, sz, _ = _cl._c_buffer_from_obj({buf_var})
+ status = _lib.kernel__set_arg_buf(self.ptr, {arg_idx}, c_buf, sz)
+ if status != _ffi.NULL:
+ _handle_error(status)
+ """
+ .format(arg_idx=arg_idx, buf_var=buf_var))
+
+ def kernel__generate_generic_arg_handler(self, gen, arg_idx, arg_var):
+ from pytools.py_codegen import Indentation
+
+ gen("""
+ if {arg_var} is None:
+ status = _lib.kernel__set_arg_null(self.ptr, {arg_idx})
+ if status != _ffi.NULL:
+ _handle_error(status)
+ elif isinstance({arg_var}, _CLKernelArg):
+ self.set_arg({arg_idx}, {arg_var})
+ """
+ .format(arg_idx=arg_idx, arg_var=arg_var))
+
+ gen("else:")
+ with Indentation(gen):
+ self._generate_buffer_arg_setter(gen, arg_idx, arg_var)
+
+ def kernel__generate_naive_call(self):
+ num_args = self.num_args
+
+ from pytools.py_codegen import PythonCodeGenerator
+ gen = PythonCodeGenerator()
+
+ for i in range(num_args):
+ gen("# process argument {arg_idx}".format(arg_idx=i))
+ gen("")
+ gen("current_arg = {arg_idx}".format(arg_idx=i))
+ self._generate_generic_arg_handler(gen, i, "arg%d" % i)
+ gen("")
+
+ self._set_set_args_body(gen, num_args)
+
+ def kernel_set_scalar_arg_dtypes(self, scalar_arg_dtypes):
+ # {{{ 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
+
+ count_bug_per_dev = [
+ has_struct_arg_count_bug(dev)
+ for dev in self.context.devices]
+
+ if any(count_bug_per_dev):
+ if all(count_bug_per_dev):
+ work_around_arg_count_bug = True
else:
- arg_type_chars.append(np.dtype(arg_dtype).char)
+ warn_about_arg_count_bug = True
+
+ # }}}
+
+ cl_arg_idx = 0
+
+ from pytools.py_codegen import PythonCodeGenerator
+ gen = PythonCodeGenerator()
+
+ for arg_idx, arg_dtype in enumerate(scalar_arg_dtypes):
+ gen("# process argument {arg_idx}".format(arg_idx=arg_idx))
+ gen("")
+ gen("current_arg = {arg_idx}".format(arg_idx=arg_idx))
+ arg_var = "arg%d" % arg_idx
+
+ if arg_dtype is None:
+ self._generate_generic_arg_handler(gen, cl_arg_idx, arg_var)
+ cl_arg_idx += 1
+ gen("")
+ continue
+
+ arg_dtype = np.dtype(arg_dtype)
- self._arg_type_chars = arg_type_chars
+ if arg_dtype.char == "V":
+ self._generate_generic_arg_handler(gen, cl_arg_idx, arg_var)
+ cl_arg_idx += 1
- def kernel_set_args(self, *args):
- assert len(args) == self.num_args, (
+ elif arg_dtype.kind == "c":
+ if warn_about_arg_count_bug:
+ 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=self.function_name, stacklevel=2))
+
+ 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:
+ gen(
+ "buf = pack('{arg_char}', {arg_var}.real)"
+ .format(arg_char=arg_char, arg_var=arg_var))
+ self._generate_buffer_arg_setter(gen, cl_arg_idx, "buf")
+ cl_arg_idx += 1
+ gen(
+ "buf = pack('{arg_char}', {arg_var}.imag)"
+ .format(arg_char=arg_char, arg_var=arg_var))
+ self._generate_buffer_arg_setter(gen, cl_arg_idx, "buf")
+ cl_arg_idx += 1
+ else:
+ gen(
+ "buf = pack('{arg_char}{arg_char}', "
+ "{arg_var}.real, {arg_var}.imag)"
+ .format(arg_char=arg_char, arg_var=arg_var))
+ self._generate_buffer_arg_setter(gen, cl_arg_idx, "buf")
+ cl_arg_idx += 1
+
+ elif arg_dtype.char in "IL" and _CPY26:
+ # Prevent SystemError: ../Objects/longobject.c:336: bad
+ # argument to internal function
+
+ gen(
+ "buf = pack('{arg_char}', long({arg_var})"
+ .format(arg_char=arg_dtype.char, arg_var=arg_var))
+ self._generate_buffer_arg_setter(gen, cl_arg_idx, "buf")
+ cl_arg_idx += 1
+
+ else:
+ arg_char = arg_dtype.char
+ arg_char = _type_char_map.get(arg_char, arg_char)
+ gen(
+ "buf = pack('{arg_char}', {arg_var})"
+ .format(
+ arg_char=arg_char,
+ arg_var=arg_var))
+ self._generate_buffer_arg_setter(gen, cl_arg_idx, "buf")
+ cl_arg_idx += 1
+
+ gen("")
+
+ if cl_arg_idx != self.num_args:
+ raise TypeError(
"length of argument list (%d) and "
"CL-generated number of arguments (%d) do not agree"
- % (len(args), self.num_args))
+ % (cl_arg_idx, self.num_args))
- i = None
- try:
- try:
- arg_type_chars = self.__dict__["_arg_type_chars"]
- except KeyError:
- for i, arg in enumerate(args):
- self.set_arg(i, arg)
- else:
- from pyopencl._pvt_struct import pack
+ self._set_set_args_body(gen, len(scalar_arg_dtypes))
- for i, (arg, arg_type_char) in enumerate(
- zip(args, arg_type_chars)):
- if arg_type_char and arg_type_char != "V":
- self.set_arg(i, pack(arg_type_char, arg))
- else:
- self.set_arg(i, arg)
- except TypeError as e:
- if i is not None:
- advice = ""
- from pyopencl.array import Array
- if isinstance(args[i], Array):
- advice = " (perhaps you meant to pass 'array.data' " \
- "instead of the array itself?)"
-
- raise LogicError(
- "when processing argument #%d (1-based): %s%s"
- % (i+1, str(e), advice))
- else:
- raise
+ # }}}
+
+ def kernel_set_args(self, *args, **kwargs):
+ # Need to dupicate the 'self' argument for dynamically generated method
+ return self._set_args(self, *args, **kwargs)
+
+ def kernel_call(self, queue, global_size, local_size, *args, **kwargs):
+ # __call__ can't be overridden directly, so we need this
+ # trampoline hack.
+ return self._enqueue(self, queue, global_size, local_size, *args, **kwargs)
def kernel_capture_call(self, filename, queue, global_size, local_size,
*args, **kwargs):
@@ -683,9 +894,13 @@ def _add_functionality():
*args, **kwargs)
Kernel.__init__ = kernel_init
- Kernel.__call__ = kernel_call
+ Kernel._set_set_args_body = kernel__set_set_args_body
+ Kernel._generate_buffer_arg_setter = kernel__generate_buffer_arg_setter
+ Kernel._generate_generic_arg_handler = kernel__generate_generic_arg_handler
+ Kernel._generate_naive_call = kernel__generate_naive_call
Kernel.set_scalar_arg_dtypes = kernel_set_scalar_arg_dtypes
Kernel.set_args = kernel_set_args
+ Kernel.__call__ = kernel_call
Kernel.capture_call = kernel_capture_call
# }}}
@@ -842,11 +1057,20 @@ def _add_functionality():
except AttributeError:
return str(val)
else:
- result = "%s failed: %s" % (val.routine(),
- status_code.to_string(val.code(), "<unknown error %d>")
- .lower().replace("_", " "))
- if val.what():
- result += " - " + val.what()
+ result = ""
+ if val.code() != status_code.SUCCESS:
+ result = status_code.to_string(
+ val.code(), "<unknown error %d>")
+ routine = val.routine()
+ if routine:
+ result = "%s failed: %s" % (
+ routine.lower().replace("_", " "),
+ result)
+ what = val.what()
+ if what:
+ if result:
+ result += " - "
+ result += what
return result
def error_code(self):
diff --git a/pyopencl/cffi_cl.py b/pyopencl/cffi_cl.py
index 2f92250f..fbd2a0cc 100644
--- a/pyopencl/cffi_cl.py
+++ b/pyopencl/cffi_cl.py
@@ -37,6 +37,11 @@ from .compyte.array import f_contiguous_strides, c_contiguous_strides
_lib = _ffi.dlopen(None)
+
+class _CLKernelArg(object):
+ pass
+
+
# {{{ hook up connections between the wrapper and the interperter
import gc
@@ -713,7 +718,7 @@ class cffi_array(np.ndarray): # noqa
return self.__base
-class LocalMemory(object):
+class LocalMemory(_CLKernelArg):
__slots__ = ('_size',)
def __init__(self, size):
@@ -724,7 +729,7 @@ class LocalMemory(object):
return self._size
-class MemoryObjectHolder(_Common):
+class MemoryObjectHolder(_Common, _CLKernelArg):
def get_host_array(self, shape, dtype, order="C"):
dtype, shape, strides = _norm_shape_dtype(
shape, dtype, order, None, 'MemoryObjectHolder.get_host_array')
@@ -1030,17 +1035,20 @@ class Kernel(_Common):
self.ptr = ptr_kernel[0]
def set_arg(self, arg_index, arg):
+ # If you change this, also change the kernel call generation logic.
if arg is None:
_handle_error(_lib.kernel__set_arg_null(self.ptr, arg_index))
- elif isinstance(arg, MemoryObjectHolder):
- _handle_error(_lib.kernel__set_arg_mem(self.ptr, arg_index, arg.ptr))
- elif isinstance(arg, Sampler):
- _handle_error(_lib.kernel__set_arg_sampler(self.ptr, arg_index,
- arg.ptr))
- elif isinstance(arg, LocalMemory):
- _handle_error(_lib.kernel__set_arg_buf(self.ptr, arg_index,
- _ffi.NULL, arg.size))
+ elif isinstance(arg, _CLKernelArg):
+ if isinstance(arg, MemoryObjectHolder):
+ _handle_error(_lib.kernel__set_arg_mem(self.ptr, arg_index, arg.ptr))
+ elif isinstance(arg, Sampler):
+ _handle_error(_lib.kernel__set_arg_sampler(self.ptr, arg_index,
+ arg.ptr))
+ elif isinstance(arg, LocalMemory):
+ _handle_error(_lib.kernel__set_arg_buf(self.ptr, arg_index,
+ _ffi.NULL, arg.size))
elif _CPY2 and isinstance(arg, np.generic):
+ # https://github.com/numpy/numpy/issues/5381
c_buf, size, _ = _c_buffer_from_obj(np.getbuffer(arg))
_handle_error(_lib.kernel__set_arg_buf(self.ptr, arg_index,
c_buf, size))
@@ -1869,7 +1877,7 @@ class Image(MemoryObject):
# {{{ Sampler
-class Sampler(_Common):
+class Sampler(_Common, _CLKernelArg):
_id = 'sampler'
def __init__(self, context, normalized_coords, addressing_mode, filter_mode):
diff --git a/setup.py b/setup.py
index 0c19b0f7..4ae6a052 100644
--- a/setup.py
+++ b/setup.py
@@ -207,7 +207,7 @@ def main():
install_requires=[
"numpy",
- "pytools>=2014.2",
+ "pytools>=2015.1.1",
"pytest>=2",
"decorator>=3.2.0",
"cffi>=1.1.0",
diff --git a/test/test_algorithm.py b/test/test_algorithm.py
index 5518d508..b55c850e 100644
--- a/test/test_algorithm.py
+++ b/test/test_algorithm.py
@@ -328,9 +328,6 @@ def test_dot(ctx_factory):
queue = cl.CommandQueue(context)
dev = context.devices[0]
- from pyopencl.characterize import has_struct_arg_count_bug
- if has_struct_arg_count_bug(dev):
- pytest.xfail("device has struct arg counting bug")
dtypes = [np.float32, np.complex64]
if has_double_support(dev):
diff --git a/test/test_array.py b/test/test_array.py
index adb2f744..ecfd3ba9 100644
--- a/test/test_array.py
+++ b/test/test_array.py
@@ -95,11 +95,6 @@ def test_mix_complex(ctx_factory):
context = ctx_factory()
queue = cl.CommandQueue(context)
- dev = context.devices[0]
- from pyopencl.characterize import has_struct_arg_count_bug
- if has_struct_arg_count_bug(dev):
- pytest.xfail("device has struct arg counting bug")
-
size = 10
dtypes = [
@@ -174,9 +169,6 @@ def test_pow_neg1_vs_inv(ctx_factory):
if not has_double_support(device):
from pytest import skip
skip("double precision not supported on %s" % device)
- from pyopencl.characterize import has_struct_arg_count_bug
- if has_struct_arg_count_bug(device):
- pytest.xfail("device has struct arg counting bug")
a_dev = make_random_array(queue, np.complex128, 20000)
diff --git a/test/test_clmath.py b/test/test_clmath.py
index e0e07764..4dea8f91 100644
--- a/test/test_clmath.py
+++ b/test/test_clmath.py
@@ -76,11 +76,6 @@ def make_unary_function_test(name, limits=(0, 1), threshold=0, use_complex=False
gpu_func = getattr(clmath, name)
cpu_func = getattr(np, numpy_func_names.get(name, name))
- dev = context.devices[0]
- from pyopencl.characterize import has_struct_arg_count_bug
- if use_complex and has_struct_arg_count_bug(dev):
- pytest.xfail("device has struct arg counting bug")
-
if has_double_support(context.devices[0]):
if use_complex:
dtypes = [np.float32, np.float64, np.complex64, np.complex128]
@@ -132,8 +127,8 @@ if have_cl():
(-math.pi/2 + 0.1, math.pi/2 - 0.1), 4e-5, use_complex=True)
test_atan = make_unary_function_test("atan", (-10, 10), 2e-7)
- test_sinh = make_unary_function_test("sinh", (-3, 3), 2e-6, use_complex=2e-3)
- test_cosh = make_unary_function_test("cosh", (-3, 3), 2e-6, use_complex=2e-3)
+ test_sinh = make_unary_function_test("sinh", (-3, 3), 3e-6, use_complex=2e-3)
+ test_cosh = make_unary_function_test("cosh", (-3, 3), 3e-6, use_complex=2e-3)
test_tanh = make_unary_function_test("tanh", (-3, 3), 2e-6, use_complex=True)
@@ -228,10 +223,6 @@ def test_frexp(ctx_factory):
context = ctx_factory()
queue = cl.CommandQueue(context)
- if context.devices[0].platform.name == "Portable Computing Language":
- # https://github.com/pocl/pocl/issues/202
- pytest.xfail("POCL's frexp seems to have issues")
-
for s in sizes:
a = cl_array.arange(queue, s, dtype=np.float32)/10
significands, exponents = clmath.frexp(a)
--
GitLab