def fill_rand(ary):
from pyopencl.clrandom import fill_rand
if ary.dtype.kind == "c":
real_dtype = ary.dtype.type(0).real.dtype
real_ary = ary.view(real_dtype)
fill_rand(real_ary)
else:
fill_rand(ary)
def fill_rand(ary):
from pyopencl.clrandom import fill_rand
if ary.dtype.kind == "c":
real_dtype = ary.dtype.type(0).real.dtype
real_ary = ary.view(real_dtype)
fill_rand(real_ary)
else:
fill_rand(ary)
def make_ref_args(kernel, impl_arg_info, queue, parameters):
import pyopencl as cl
import pyopencl.array as cl_array
from loopy.kernel.data import ValueArg, ArrayArg, ImageArg, \
TemporaryVariable, ConstantArg
from pymbolic import evaluate
ref_args = {}
ref_arg_data = []
for arg in impl_arg_info:
kernel_arg = kernel.impl_arg_to_arg.get(arg.name)
if arg.arg_class is ValueArg:
if arg.offset_for_name:
continue
arg_value = parameters[arg.name]
try:
argv_dtype = arg_value.dtype
except AttributeError:
argv_dtype = None
if argv_dtype != arg.dtype:
arg_value = arg.dtype.numpy_dtype.type(arg_value)
ref_args[arg.name] = arg_value
ref_arg_data.append(None)
elif arg.arg_class is ArrayArg or arg.arg_class is ImageArg \
or arg.arg_class is ConstantArg:
if arg.shape is None or any(saxis is None for saxis in arg.shape):
raise LoopyError(
"array '%s' needs known shape to use automatic "
"testing" % arg.name)
shape = evaluate_shape(arg.unvec_shape, parameters)
dtype = kernel_arg.dtype
is_output = arg.base_name in kernel.get_written_variables()
if arg.arg_class is ImageArg:
storage_array = ary = cl_array.empty(queue,
shape,
dtype,
order="C")
numpy_strides = None
alloc_size = None
strides = None
else:
strides = evaluate(arg.unvec_strides, parameters)
alloc_size = sum(astrd * (alen - 1) if astrd != 0 else alen - 1
for alen, astrd in zip(shape, strides)) + 1
if dtype is None:
raise LoopyError("dtype for argument '%s' is not yet "
"known. Perhaps you want to use "
"loopy.add_dtypes "
"or loopy.infer_argument_dtypes?" %
arg.name)
itemsize = dtype.itemsize
numpy_strides = [itemsize * s for s in strides]
storage_array = cl_array.empty(queue, alloc_size, dtype)
if is_output and arg.arg_class is ImageArg:
raise LoopyError("write-mode images not supported in "
"automatic testing")
fill_rand(storage_array)
if arg.arg_class is ImageArg:
# must be contiguous
pre_run_ary = pre_run_storage_array = storage_array.copy()
ref_args[arg.name] = cl.image_from_array(
queue.context, ary.get())
else:
pre_run_storage_array = storage_array.copy()
ary = cl_array.as_strided(storage_array, shape, numpy_strides)
pre_run_ary = cl_array.as_strided(pre_run_storage_array, shape,
numpy_strides)
ref_args[arg.name] = ary
ref_arg_data.append(
TestArgInfo(name=arg.name,
ref_array=ary,
ref_storage_array=storage_array,
ref_pre_run_array=pre_run_ary,
ref_pre_run_storage_array=pre_run_storage_array,
ref_shape=shape,
ref_strides=strides,
ref_alloc_size=alloc_size,
ref_numpy_strides=numpy_strides,
needs_checking=is_output))
elif arg.arg_class is TemporaryVariable:
# global temporary, handled by invocation logic
pass
else:
raise LoopyError("arg type %s not understood" % type(arg))
return ref_args, ref_arg_data
def make_ref_args(kernel, impl_arg_info, queue, parameters):
import pyopencl as cl
import pyopencl.array as cl_array
from loopy.kernel.data import ValueArg, GlobalArg, ImageArg, TemporaryVariable
from pymbolic import evaluate
ref_args = {}
ref_arg_data = []
for arg in impl_arg_info:
kernel_arg = kernel.impl_arg_to_arg.get(arg.name)
if arg.arg_class is ValueArg:
if arg.offset_for_name:
continue
arg_value = parameters[arg.name]
try:
argv_dtype = arg_value.dtype
except AttributeError:
argv_dtype = None
if argv_dtype != arg.dtype:
arg_value = arg.dtype.numpy_dtype.type(arg_value)
ref_args[arg.name] = arg_value
ref_arg_data.append(None)
elif arg.arg_class is GlobalArg or arg.arg_class is ImageArg:
if arg.shape is None or any(saxis is None for saxis in arg.shape):
raise LoopyError("array '%s' needs known shape to use automatic "
"testing" % arg.name)
shape = evaluate_shape(arg.unvec_shape, parameters)
dtype = kernel_arg.dtype
is_output = arg.base_name in kernel.get_written_variables()
if arg.arg_class is ImageArg:
storage_array = ary = cl_array.empty(
queue, shape, dtype, order="C")
numpy_strides = None
alloc_size = None
strides = None
else:
strides = evaluate(arg.unvec_strides, parameters)
from pytools import all
assert all(s > 0 for s in strides)
alloc_size = sum(astrd*(alen-1)
for alen, astrd in zip(shape, strides)) + 1
if dtype is None:
raise LoopyError("dtype for argument '%s' is not yet "
"known. Perhaps you want to use "
"loopy.add_dtypes "
"or loopy.infer_argument_dtypes?"
% arg.name)
itemsize = dtype.itemsize
numpy_strides = [itemsize*s for s in strides]
storage_array = cl_array.empty(queue, alloc_size, dtype)
if is_output and arg.arg_class is ImageArg:
raise LoopyError("write-mode images not supported in "
"automatic testing")
fill_rand(storage_array)
if arg.arg_class is ImageArg:
# must be contiguous
pre_run_ary = pre_run_storage_array = storage_array.copy()
ref_args[arg.name] = cl.image_from_array(
queue.context, ary.get())
else:
pre_run_storage_array = storage_array.copy()
ary = cl_array.as_strided(storage_array, shape, numpy_strides)
pre_run_ary = cl_array.as_strided(
pre_run_storage_array, shape, numpy_strides)
ref_args[arg.name] = ary
ref_arg_data.append(
TestArgInfo(
name=arg.name,
ref_array=ary,
ref_storage_array=storage_array,
ref_pre_run_array=pre_run_ary,
ref_pre_run_storage_array=pre_run_storage_array,
ref_shape=shape,
ref_strides=strides,
ref_alloc_size=alloc_size,
ref_numpy_strides=numpy_strides,
needs_checking=is_output))
elif arg.arg_class is TemporaryVariable:
# global temporary, handled by invocation logic
pass
else:
raise LoopyError("arg type not understood")
return ref_args, ref_arg_data