def parse_c_arg(c_arg):
c_arg = c_arg.replace("const", "").replace("volatile", "")
# process and remove declarator
import re
decl_re = re.compile(r"(\**)\s*([_a-zA-Z0-9]+)(\s*\[[ 0-9]*\])*\s*$")
decl_match = decl_re.search(c_arg)
if decl_match is None:
raise ValueError("couldn't parse C declarator '%s'" % c_arg)
name = decl_match.group(2)
if decl_match.group(1) or decl_match.group(3) is not None:
arg_class = VectorArg
else:
arg_class = ScalarArg
tp = c_arg[:decl_match.start()]
tp = " ".join(tp.split())
from pycuda.characterize import platform_bits
from sys import platform
if tp == "float": dtype = np.float32
elif tp == "double": dtype = np.float64
elif tp == "pycuda::complex<float>": dtype = np.complex64
elif tp == "pycuda::complex<double>": dtype = np.complex128
elif tp in ["int", "signed int"]: dtype = np.int32
elif tp in ["unsigned", "unsigned int"]: dtype = np.uint32
elif tp in ["long", "long int"]:
if platform_bits() == 64 and 'win32' not in platform:
dtype = np.int64
else:
dtype = np.int32
elif tp in ["unsigned long", "unsigned long int"]:
if platform_bits() == 64 and 'win32' not in platform:
dtype = np.uint64
else:
dtype = np.uint32
elif tp in ["short", "short int"]:
dtype = np.int16
elif tp in ["unsigned short", "unsigned short int"]:
dtype = np.uint16
elif tp in ["char", "signed char"]:
dtype = np.int8
elif tp in ["unsigned char"]:
dtype = np.uint8
elif tp in ["bool"]:
dtype = np.bool
else:
import pycuda.gpuarray as gpuarray
try:
return gpuarray.vec._c_name_to_dtype[tp]
except KeyError:
raise ValueError("unknown type '%s'" % tp)
return arg_class(dtype, name)
def parse_c_arg(c_arg):
c_arg = c_arg.replace("const", "").replace("volatile", "")
# process and remove declarator
import re
decl_re = re.compile(r"(\**)\s*([_a-zA-Z0-9]+)(\s*\[[ 0-9]*\])*\s*$")
decl_match = decl_re.search(c_arg)
if decl_match is None:
raise ValueError("couldn't parse C declarator '%s'" % c_arg)
name = decl_match.group(2)
if decl_match.group(1) or decl_match.group(3) is not None:
arg_class = VectorArg
else:
arg_class = ScalarArg
tp = c_arg[:decl_match.start()]
tp = " ".join(tp.split())
from pycuda.characterize import platform_bits
from sys import platform
if tp == "float": dtype = np.float32
elif tp == "double": dtype = np.float64
elif tp == "pycuda::complex<float>": dtype = np.complex64
elif tp == "pycuda::complex<double>": dtype = np.complex128
elif tp in ["int", "signed int"]: dtype = np.int32
elif tp in ["unsigned", "unsigned int"]: dtype = np.uint32
elif tp in ["long", "long int"]:
if platform_bits() == 64 and 'win32' not in platform:
dtype = np.int64
else:
dtype = np.int32
elif tp in ["unsigned long", "unsigned long int", "long unsigned int"]:
if platform_bits() == 64 and 'win32' not in platform:
dtype = np.uint64
else:
dtype = np.uint32
elif tp in ["short", "short int"]: dtype = np.int16
elif tp in ["unsigned short", "unsigned short int", "short unsigned int"]:
dtype = np.uint16
elif tp in ["char", "signed char"]: dtype = np.int8
elif tp in ["unsigned char"]: dtype = np.uint8
elif tp in ["bool"]: dtype = np.bool
else:
import pycuda.gpuarray as gpuarray
try:
dtype = gpuarray.vec._c_name_to_dtype[tp]
except KeyError:
raise ValueError("unknown type '%s'" % tp)
return arg_class(dtype, name)
def dtype_to_ctype(dtype, with_fp_tex_hack=False):
from pycuda.characterize import platform_bits
from sys import platform
if dtype is None:
raise ValueError("dtype may not be None")
dtype = np.dtype(dtype)
if dtype == np.int64 and platform_bits() == 64:
if 'win32' in platform:
return "long long"
else:
return "long"
elif dtype == np.uint64 and platform_bits() == 64:
if 'win32' in platform:
return "unsigned long long"
else:
return "unsigned long"
elif dtype == np.int32:
return "int"
elif dtype == np.uint32:
return "unsigned int"
elif dtype == np.int16:
return "short int"
elif dtype == np.uint16:
return "short unsigned int"
elif dtype == np.int8:
return "signed char"
elif dtype == np.uint8:
return "unsigned char"
elif dtype == np.bool:
return "bool"
elif dtype == np.float32:
if with_fp_tex_hack:
return "fp_tex_float"
else:
return "float"
elif dtype == np.float64:
if with_fp_tex_hack:
return "fp_tex_double"
else:
return "double"
elif dtype == np.complex64:
return "pycuda::complex<float>"
elif dtype == np.complex128:
return "pycuda::complex<double>"
else:
import pycuda.gpuarray as gpuarray
try:
return gpuarray.vec._dtype_to_c_name[dtype]
except KeyError:
raise ValueError, "unable to map dtype '%s'" % dtype
def dtype_to_ctype(dtype, with_fp_tex_hack=False):
from pycuda.characterize import platform_bits
from sys import platform
if dtype is None:
raise ValueError("dtype may not be None")
dtype = np.dtype(dtype)
if dtype == np.int64 and platform_bits() == 64:
if 'win32' in platform:
return "long long"
else:
return "long"
elif dtype == np.uint64 and platform_bits() == 64:
if 'win32' in platform:
return "unsigned long long"
else:
return "unsigned long"
elif dtype == np.int32:
return "int"
elif dtype == np.uint32:
return "unsigned int"
elif dtype == np.int16:
return "short int"
elif dtype == np.uint16:
return "short unsigned int"
elif dtype == np.int8:
return "signed char"
elif dtype == np.uint8:
return "unsigned char"
elif dtype == np.bool:
return "bool"
elif dtype == np.float32:
if with_fp_tex_hack:
return "fp_tex_float"
else:
return "float"
elif dtype == np.float64:
if with_fp_tex_hack:
return "fp_tex_double"
else:
return "double"
elif dtype == np.complex64:
return "pycuda::complex<float>"
elif dtype == np.complex128:
return "pycuda::complex<double>"
else:
import pycuda.gpuarray as gpuarray
try:
return gpuarray.vec._dtype_to_c_name[dtype]
except KeyError:
raise ValueError, "unable to map dtype '%s'" % dtype
def compile_plain(source, options, keep, nvcc, cache_dir):
from os.path import join
if cache_dir:
checksum = _new_md5()
checksum.update(source)
for option in options:
checksum.update(option)
checksum.update(get_nvcc_version(nvcc))
from pycuda.characterize import platform_bits
checksum.update(str(platform_bits()))
cache_file = checksum.hexdigest()
cache_path = join(cache_dir, cache_file + ".cubin")
try:
return open(cache_path, "rb").read()
except:
pass
from tempfile import mkdtemp
file_dir = mkdtemp()
file_root = "kernel"
cu_file_name = file_root + ".cu"
cu_file_path = join(file_dir, cu_file_name)
outf = open(cu_file_path, "w")
outf.write(str(source))
outf.close()
if keep:
options = options[:]
options.append("--keep")
print "*** compiler output in %s" % file_dir
cmdline = [nvcc, "--cubin"] + options + [cu_file_name]
try:
from pytools.prefork import call_capture_output
except ImportError:
from pytools.prefork import call
try:
result = call(cmdline, cwd=file_dir)
except OSError, e:
raise OSError("%s was not found (is it on the PATH?) [%s]"
% (nvcc, str(e)))
stdout = None
stderr = None
def _create_vector_types():
name_to_dtype = {}
dtype_to_name = {}
from pycuda.characterize import platform_bits
if platform_bits() == 32:
long_dtype = np.int32
ulong_dtype = np.uint32
else:
long_dtype = np.int64
ulong_dtype = np.uint64
field_names = ["x", "y", "z", "w"]
for base_name, base_type, counts in [
('char', np.int8, [1, 2, 3, 4]),
('uchar', np.uint8, [1, 2, 3, 4]),
('short', np.int16, [1, 2, 3, 4]),
('ushort', np.uint16, [1, 2, 3, 4]),
('int', np.uint32, [1, 2, 3, 4]),
('uint', np.uint32, [1, 2, 3, 4]),
('long', long_dtype, [1, 2, 3, 4]),
('ulong', ulong_dtype, [1, 2, 3, 4]),
('longlong', np.int64, [1, 2]),
('ulonglong', np.uint64, [1, 2]),
('float', np.float32, [1, 2, 3, 4]),
('ulonglong', np.float64, [1, 2]),
]:
for count in counts:
name = "%s%d" % (base_name, count)
dtype = np.dtype([(field_names[i], base_type)
for i in range(count)])
name_to_dtype[name] = dtype
dtype_to_name[dtype] = name
setattr(vec, name, dtype)
my_field_names = ",".join(field_names[:count])
setattr(
vec, "make_" + name,
staticmethod(
eval(
"lambda %s: array((%s), dtype=my_dtype)" %
(my_field_names, my_field_names),
dict(array=np.array, my_dtype=dtype))))
vec._dtype_to_c_name = dtype_to_name
vec._c_name_to_dtype = name_to_dtype
def _create_vector_types():
name_to_dtype = {}
dtype_to_name = {}
from pycuda.characterize import platform_bits
if platform_bits() == 32:
long_dtype = np.int32
ulong_dtype = np.uint32
else:
long_dtype = np.int64
ulong_dtype = np.uint64
field_names = ["x", "y", "z", "w"]
for base_name, base_type, counts in [
('char', np.int8, [1,2,3,4]),
('uchar', np.uint8, [1,2,3,4]),
('short', np.int16, [1,2,3,4]),
('ushort', np.uint16, [1,2,3,4]),
('int', np.int32, [1,2,3,4]),
('uint', np.uint32, [1,2,3,4]),
('long', long_dtype, [1,2,3,4]),
('ulong', ulong_dtype, [1,2,3,4]),
('longlong', np.int64, [1,2]),
('ulonglong', np.uint64, [1,2]),
('float', np.float32, [1,2,3,4]),
('ulonglong', np.float64, [1,2]),
]:
for count in counts:
name = "%s%d" % (base_name, count)
dtype = np.dtype([
(field_names[i], base_type)
for i in range(count)])
name_to_dtype[name] = dtype
dtype_to_name[dtype] = name
setattr(vec, name, dtype)
my_field_names = ",".join(field_names[:count])
setattr(vec, "make_"+name,
staticmethod(eval(
"lambda %s: array((%s), dtype=my_dtype)"
% (my_field_names, my_field_names),
dict(array=np.array, my_dtype=dtype))))
vec._dtype_to_c_name = dtype_to_name
vec._c_name_to_dtype = name_to_dtype
def _create_vector_types():
from pycuda.characterize import platform_bits
if platform_bits() == 32:
long_dtype = np.int32
ulong_dtype = np.uint32
else:
long_dtype = np.int64
ulong_dtype = np.uint64
field_names = ["x", "y", "z", "w"]
from pycuda.tools import register_dtype
for base_name, base_type, counts in [
("char", np.int8, [1, 2, 3, 4]),
("uchar", np.uint8, [1, 2, 3, 4]),
("short", np.int16, [1, 2, 3, 4]),
("ushort", np.uint16, [1, 2, 3, 4]),
("int", np.int32, [1, 2, 3, 4]),
("uint", np.uint32, [1, 2, 3, 4]),
("long", long_dtype, [1, 2, 3, 4]),
("ulong", ulong_dtype, [1, 2, 3, 4]),
("longlong", np.int64, [1, 2]),
("ulonglong", np.uint64, [1, 2]),
("float", np.float32, [1, 2, 3, 4]),
("double", np.float64, [1, 2]),
]:
for count in counts:
name = "%s%d" % (base_name, count)
dtype = np.dtype([(field_names[i], base_type) for i in range(count)])
register_dtype(dtype, name, alias_ok=True)
setattr(vec, name, dtype)
my_field_names = ",".join(field_names[:count])
setattr(
vec,
"make_" + name,
staticmethod(
eval(
"lambda %s: array((%s), dtype=my_dtype)" % (my_field_names, my_field_names),
dict(array=np.array, my_dtype=dtype),
)
),
)
def compile_plain(source, options, keep, nvcc, cache_dir, target="cubin"):
from os.path import join
assert target in ["cubin", "ptx", "fatbin"]
if cache_dir:
checksum = _new_md5()
if '#include' in source:
checksum.update(
preprocess_source(source, options, nvcc).encode("utf-8"))
else:
checksum.update(source.encode("utf-8"))
for option in options:
checksum.update(option.encode("utf-8"))
checksum.update(get_nvcc_version(nvcc).encode("utf-8"))
from pycuda.characterize import platform_bits
checksum.update(str(platform_bits()).encode("utf-8"))
cache_file = checksum.hexdigest()
cache_path = join(cache_dir, cache_file + "." + target)
try:
cache_file = open(cache_path, "rb")
try:
return cache_file.read()
finally:
cache_file.close()
except:
pass
from tempfile import mkdtemp
file_dir = mkdtemp()
file_root = "kernel"
cu_file_name = file_root + ".cu"
cu_file_path = join(file_dir, cu_file_name)
outf = open(cu_file_path, "w")
outf.write(str(source))
outf.close()
if keep:
options = options[:]
options.append("--keep")
print("*** compiler output in %s" % file_dir)
cmdline = [nvcc, "--" + target] + options + [cu_file_name]
result, stdout, stderr = call_capture_output(cmdline,
cwd=file_dir,
error_on_nonzero=False)
try:
result_f = open(join(file_dir, file_root + "." + target), "rb")
except IOError:
no_output = True
else:
no_output = False
if result != 0 or (no_output and (stdout or stderr)):
if result == 0:
from warnings import warn
warn("PyCUDA: nvcc exited with status 0, but appears to have "
"encountered an error")
from pycuda.driver import CompileError
raise CompileError("nvcc compilation of %s failed" % cu_file_path,
cmdline,
stdout=stdout.decode("utf-8", "replace"),
stderr=stderr.decode("utf-8", "replace"))
if stdout or stderr:
lcase_err_text = (stdout + stderr).decode("utf-8", "replace").lower()
from warnings import warn
if "demoted" in lcase_err_text or "demoting" in lcase_err_text:
warn(
"nvcc said it demoted types in source code it "
"compiled--this is likely not what you want.",
stacklevel=4)
warn("The CUDA compiler succeeded, but said the following:\n" +
(stdout + stderr).decode("utf-8", "replace"),
stacklevel=4)
result_data = result_f.read()
result_f.close()
if cache_dir:
outf = open(cache_path, "wb")
outf.write(result_data)
outf.close()
if not keep:
from os import listdir, unlink, rmdir
for name in listdir(file_dir):
unlink(join(file_dir, name))
rmdir(file_dir)
return result_data
def compile_plain(source, options, keep, nvcc, cache_dir, target="cubin"):
from os.path import join
assert target in ["cubin", "ptx", "fatbin"]
if cache_dir:
checksum = _new_md5()
if '#include' in source:
checksum.update(preprocess_source(source, options, nvcc).encode("utf-8"))
else:
checksum.update(source.encode("utf-8"))
for option in options:
checksum.update(option.encode("utf-8"))
checksum.update(get_nvcc_version(nvcc).encode("utf-8"))
from pycuda.characterize import platform_bits
checksum.update(str(platform_bits()).encode("utf-8"))
cache_file = checksum.hexdigest()
cache_path = join(cache_dir, cache_file + "." + target)
try:
cache_file = open(cache_path, "rb")
try:
return cache_file.read()
finally:
cache_file.close()
except:
pass
from tempfile import mkdtemp
file_dir = mkdtemp()
file_root = "kernel"
cu_file_name = file_root + ".cu"
cu_file_path = join(file_dir, cu_file_name)
outf = open(cu_file_path, "w")
outf.write(str(source))
outf.close()
if keep:
options = options[:]
options.append("--keep")
print("*** compiler output in %s" % file_dir)
cmdline = [nvcc, "--" + target] + options + [cu_file_name]
result, stdout, stderr = call_capture_output(cmdline,
cwd=file_dir, error_on_nonzero=False)
try:
result_f = open(join(file_dir, file_root + "." + target), "rb")
except IOError:
no_output = True
else:
no_output = False
if result != 0 or (no_output and (stdout or stderr)):
if result == 0:
from warnings import warn
warn("PyCUDA: nvcc exited with status 0, but appears to have "
"encountered an error")
from pycuda.driver import CompileError
raise CompileError("nvcc compilation of %s failed" % cu_file_path,
cmdline, stdout=stdout.decode("utf-8", "replace"),
stderr=stderr.decode("utf-8", "replace"))
if stdout or stderr:
lcase_err_text = (stdout+stderr).decode("utf-8", "replace").lower()
from warnings import warn
if "demoted" in lcase_err_text or "demoting" in lcase_err_text:
warn("nvcc said it demoted types in source code it "
"compiled--this is likely not what you want.",
stacklevel=4)
warn("The CUDA compiler succeeded, but said the following:\n"
+ (stdout+stderr).decode("utf-8", "replace"), stacklevel=4)
result_data = result_f.read()
result_f.close()
if cache_dir:
outf = open(cache_path, "wb")
outf.write(result_data)
outf.close()
if not keep:
from os import listdir, unlink, rmdir
for name in listdir(file_dir):
unlink(join(file_dir, name))
rmdir(file_dir)
return result_data