From 98303ded5eb205e1d3fac68851b8ab357c72d2e4 Mon Sep 17 00:00:00 2001
From: arghdos <arghdos@gmail.com>
Date: Mon, 1 May 2017 15:28:42 -0400
Subject: [PATCH] runs simple C kernels
---
loopy/execution.py | 72 +++++-------------------------
loopy/target/c/c_execution.py | 11 ++---
loopy/target/pyopencl_execution.py | 28 +++++++-----
3 files changed, 32 insertions(+), 79 deletions(-)
diff --git a/loopy/execution.py b/loopy/execution.py
index ad33ae3e7..ee2666745 100644
--- a/loopy/execution.py
+++ b/loopy/execution.py
@@ -391,6 +391,9 @@ class ExecutionWrapperGeneratorBase(object):
# }}}
+ def get_arg_pass(self, arg):
+ return arg.name
+
# {{{ arg setup
def generate_arg_setup(
@@ -580,7 +583,7 @@ class ExecutionWrapperGeneratorBase(object):
gen("")
if arg.arg_class in [lp.GlobalArg, lp.ConstantArg]:
- args.append("%s.base_data" % arg.name)
+ args.append(self.get_arg_pass(arg))
else:
args.append("%s" % arg.name)
@@ -604,7 +607,7 @@ class ExecutionWrapperGeneratorBase(object):
gen.add_to_preamble("from loopy.target.c.compyte.array"
" import as_strided as _lpy_strided")
- def intialize_system_args(self, gen):
+ def initialize_system_args(self, gen):
"""
Override to intialize any default system args
"""
@@ -615,10 +618,8 @@ class ExecutionWrapperGeneratorBase(object):
def generate_invocation(self, gen, kernel_name, args):
gen("for knl in _lpy_c_kernels:")
with Indentation(gen):
- gen("{kernel_name}({args})"
- .format(
- kernel_name='knl.name',
- args=", ".join(args)))
+ gen('knl({args})'.format(
+ args=", ".join(args)))
# }}}
@@ -629,21 +630,6 @@ class ExecutionWrapperGeneratorBase(object):
from loopy.kernel.data import KernelArgument
- if not options.no_numpy:
- gen("if out_host is None and (_lpy_encountered_numpy "
- "and not _lpy_encountered_dev):")
- with Indentation(gen):
- gen("out_host = True")
-
- gen("if out_host:")
- with Indentation(gen):
- gen("pass") # if no outputs (?!)
- for arg in implemented_data_info:
- if not issubclass(arg.arg_class, KernelArgument):
- continue
-
- gen("")
-
if options.return_dict:
gen("return None, {%s}"
% ", ".join("\"%s\": %s" % (arg.name, arg.name)
@@ -663,6 +649,9 @@ class ExecutionWrapperGeneratorBase(object):
# }}}
+ def generate_host_code(self, gen, codegen_result):
+ pass
+
def __call__(self, kernel, codegen_result):
"""
Generates the wrapping python invoker for this execution target
@@ -694,7 +683,7 @@ class ExecutionWrapperGeneratorBase(object):
gen.add_to_preamble(host_code)
gen.add_to_preamble("")
- self.intialize_system_args(gen)
+ self.initialize_system_args(gen)
self.generate_integer_arg_finding_from_shapes(
gen, kernel, implemented_data_info)
@@ -712,7 +701,6 @@ class ExecutionWrapperGeneratorBase(object):
self.generate_output_handler(gen, options, kernel, implemented_data_info)
- import pdb; pdb.set_trace()
if options.write_wrapper:
output = gen.get()
if options.highlight_wrapper:
@@ -849,43 +837,7 @@ class KernelExecutorBase(object):
raise NotImplementedError()
def __call__(self, queue, **kwargs):
- """
- :arg allocator: a callable passed a byte count and returning
- a :class:`pyopencl.Buffer`. A :class:`pyopencl` allocator
- maybe.
- :arg wait_for: A list of :class:`pyopencl.Event` instances
- for which to wait.
- :arg out_host: :class:`bool`
- Decides whether output arguments (i.e. arguments
- written by the kernel) are to be returned as
- :mod:`numpy` arrays. *True* for yes, *False* for no.
-
- For the default value of *None*, if all (input) array
- arguments are :mod:`numpy` arrays, defaults to
- returning :mod:`numpy` arrays as well.
-
- :returns: ``(evt, output)`` where *evt* is a :class:`pyopencl.Event`
- associated with the execution of the kernel, and
- output is a tuple of output arguments (arguments that
- are written as part of the kernel). The order is given
- by the order of kernel arguments. If this order is unspecified
- (such as when kernel arguments are inferred automatically),
- enable :attr:`loopy.Options.return_dict` to make *output* a
- :class:`dict` instead, with keys of argument names and values
- of the returned arrays.
- """
-
- allocator = kwargs.pop("allocator", None)
- wait_for = kwargs.pop("wait_for", None)
- out_host = kwargs.pop("out_host", None)
-
- kwargs = self.packing_controller.unpack(kwargs)
-
- kernel_info = self.cl_kernel_info(self.arg_to_dtype_set(kwargs))
-
- return kernel_info.invoker(
- kernel_info.cl_kernels, queue, allocator, wait_for,
- out_host, **kwargs)
+ raise NotImplementedError()
# }}}
diff --git a/loopy/target/c/c_execution.py b/loopy/target/c/c_execution.py
index 618da0226..d819f347a 100644
--- a/loopy/target/c/c_execution.py
+++ b/loopy/target/c/c_execution.py
@@ -138,11 +138,10 @@ class CompiledCKernel(object):
self._fn.argtypes = [ctype for name, ctype in self._arg_info]
self._prepared_call_cache = weakref.WeakKeyDictionary()
- def __call__(self, **kwargs):
+ def __call__(self, *args):
"""Execute kernel with given args mapped to ctypes equivalents."""
args_ = []
- for knl_arg, arg_t in zip(self.knl.args, self._fn.argtypes):
- arg = kwargs[knl_arg.name]
+ for arg, arg_t in zip(args, self._fn.argtypes):
if hasattr(arg, 'ctypes'):
if arg.size == 0:
# TODO eliminate unused arguments from kernel
@@ -239,11 +238,9 @@ class CKernelExecutor(KernelExecutorBase):
from pytools import invoke_editor
dev_code = invoke_editor(dev_code, "code.cl")
- c_kernels = _Kernels()
+ c_kernels = []
for dp in codegen_result.device_programs:
- setattr(c_kernels, dp.name, CompiledCKernel(dp,
- self.kernel.target,
- self.compiler))
+ c_kernels.append(CompiledCKernel(dp, self.kernel.target, self.compiler))
return _KernelInfo(
kernel=kernel,
diff --git a/loopy/target/pyopencl_execution.py b/loopy/target/pyopencl_execution.py
index 021bc786f..e67c49c59 100644
--- a/loopy/target/pyopencl_execution.py
+++ b/loopy/target/pyopencl_execution.py
@@ -150,15 +150,13 @@ class PyOpenCLExecutionWrapperGenerator(ExecutionWrapperGeneratorBase):
# {{{ generate invocation
def generate_invocation(self, gen, kernel_name, args):
- gen("for knl in _lpy_cl_kernels:")
- with Indentation(gen):
- gen("_lpy_evt = {kernel_name}({args})"
- .format(
- kernel_name=kernel_name,
- args=", ".join(
- ["_lpy_cl_kernels", "queue"]
- + args
- + ["wait_for=wait_for"])))
+ gen("_lpy_evt = {kernel_name}({args})"
+ .format(
+ kernel_name=kernel_name,
+ args=", ".join(
+ ["_lpy_cl_kernels", "queue"]
+ + args
+ + ["wait_for=wait_for"])))
# }}}
@@ -185,7 +183,7 @@ class PyOpenCLExecutionWrapperGenerator(ExecutionWrapperGeneratorBase):
gen("")
if options.return_dict:
- gen("return None, {%s}"
+ gen("return _lpy_evt, {%s}"
% ", ".join("\"%s\": %s" % (arg.name, arg.name)
for arg in implemented_data_info
if issubclass(arg.arg_class, KernelArgument)
@@ -196,13 +194,19 @@ class PyOpenCLExecutionWrapperGenerator(ExecutionWrapperGeneratorBase):
if issubclass(arg.arg_class, KernelArgument)
if arg.base_name in kernel.get_written_variables()]
if out_args:
- gen("return None, (%s,)"
+ gen("return _lpy_evt, (%s,)"
% ", ".join(arg.name for arg in out_args))
else:
- gen("return None, ()")
+ gen("return _lpy_evt, ()")
# }}}
+ def generate_host_code(self, gen, codegen_result):
+ gen.add_to_preamble(codegen_result.host_code())
+
+ def get_arg_pass(self, arg):
+ return "%s.base_data" % arg.name
+
# }}}
--
GitLab