def _generate_file(self, ext: str):
# generate cubin/ptx by calling nvcc
global _test_cache_dir
nvcc = cupy.cuda.get_nvcc_path()
# split() is needed because nvcc could come from the env var NVCC
cmd = nvcc.split()
arch = '-gencode=arch=compute_{cc},code=sm_{cc}'.format(
cc=compiler._get_arch())
source = '{}/test_load_cubin.cu'.format(_test_cache_dir)
file_path = _test_cache_dir + 'test_load_cubin'
with open(source, 'w') as f:
f.write(_test_source5)
if ext == 'cubin':
file_path += '.cubin'
flag = '-cubin'
elif ext == 'ptx':
file_path += '.ptx'
flag = '-ptx'
else:
raise ValueError
cmd += [arch, flag, source, '-o', file_path]
compiler._run_nvcc(cmd, _test_cache_dir)
return file_path
def _generate_file(self, ext: str):
# generate cubin/ptx by calling nvcc/hipcc
if not cupy.cuda.runtime.is_hip:
cc = cupy.cuda.get_nvcc_path()
arch = '-gencode=arch=compute_{CC},code=sm_{CC}'.format(
CC=compiler._get_arch())
code = _test_source5
else:
# TODO(leofang): expose get_hipcc_path() to cupy.cuda?
cc = cupy._environment.get_hipcc_path()
arch = '-v' # dummy
code = compiler._convert_to_hip_source(_test_source5, None, False)
# split() is needed because nvcc could come from the env var NVCC
cmd = cc.split()
source = '{}/test_load_cubin.cu'.format(self.cache_dir)
file_path = self.cache_dir + 'test_load_cubin'
with open(source, 'w') as f:
f.write(code)
if ext == 'cubin':
file_path += '.cubin'
flag = '-cubin'
elif ext == 'ptx':
file_path += '.ptx'
flag = '-ptx'
elif ext == 'hsaco':
file_path += '.hsaco'
flag = '--genco'
else:
raise ValueError
cmd += [arch, flag, source, '-o', file_path]
cc = 'nvcc' if not cupy.cuda.runtime.is_hip else 'hipcc'
compiler._run_cc(cmd, self.cache_dir, cc)
return file_path
def _check_get_arch(self, device_cc, expected_arch):
with mock.patch('cupy.cuda.device.Device') as device_class:
device_class.return_value.compute_capability = device_cc
assert compiler._get_arch() == expected_arch
