import os
import platform
'''
Get and set environment variables that indicate to use CPU or GPU processing and
the appropriate device IDs
This allows configuring a Dask worker or inline executor to use CPU or a
particular GPU, which allows simultaneous use of all GPUs and CPUs on a
processing node without oversubscription.
Internally, this gets and sets the environment variables
:code:`LIBERTEM_USE_CUDA` and :code:`LIBERTEM_USE_CPU`.
See :ref:`udf cuda` for details on how to run UDFs with CUDA support and
:ref:`debugging udfs` for information on how to use these functions to control
where an :class:`~libertem.executor.inline.InlineJobExecutor` runs tasks.
'''
[docs]
def get_use_cuda() -> int | None:
'''
.. versionadded:: 0.6.0
Returns
-------
id : int or None
CUDA device ID to use.
'''
ret = os.environ.get("LIBERTEM_USE_CUDA")
opp = os.environ.get("LIBERTEM_USE_CPU")
if ret is not None:
if opp is not None:
raise RuntimeError(
"Both LIBERTEM_USE_CPU and LIBERTEM_USE_CUDA set, expecting at most one"
)
return int(ret)
else:
return None
[docs]
def set_use_cuda(cuda_device: int):
'''
This sets a CUDA device ID and unsets any CPU ID
.. versionadded:: 0.6.0
Parameters
----------
cuda_device : int
CUDA device ID to use
'''
os.environ["LIBERTEM_USE_CUDA"] = str(cuda_device)
os.environ.pop("LIBERTEM_USE_CPU", None)
[docs]
def get_use_cpu():
'''
.. versionadded:: 0.6.0
Returns
-------
id : int or None
CPU device ID to use. Currently there is no pinning, i.e. the value itself is
ignored. :code:`None` means "don't use CPU" and any integer means "use CPU".
Default is 0 if no settings were applied before.
'''
ret = os.environ.get("LIBERTEM_USE_CPU")
opp = os.environ.get("LIBERTEM_USE_CUDA")
if ret is not None:
if opp is not None:
raise RuntimeError(
"Both LIBERTEM_USE_CPU and LIBERTEM_USE_CUDA set, expecting at most one"
)
ret = int(ret)
elif get_use_cuda() is None:
# If no variable is set, return CPU 0
# For example inline executor or test code
ret = 0
return ret
[docs]
def set_use_cpu(cpu: int):
'''
This sets a CPU device ID and unsets any CUDA ID
.. versionadded:: 0.6.0
Parameters
----------
cpu : int
CPU to use. The value is currently ignored, i.e. any CPU is used without pinning
'''
os.environ.pop("LIBERTEM_USE_CUDA", None)
os.environ["LIBERTEM_USE_CPU"] = str(cpu)
[docs]
def get_device_class():
'''
.. versionadded:: 0.6.0
Returns
-------
class : str
Device class to use. Can be 'cpu' or 'cuda'. Default is 'cpu' if
no settings were applied before.
'''
cuda = get_use_cuda()
cpu = get_use_cpu()
if cpu is not None and cuda is not None:
raise RuntimeError(
"Both LIBERTEM_USE_CPU and LIBERTEM_USE_CUDA set, expecting at most one"
)
if cuda is not None:
return "cuda"
else:
return "cpu"
[docs]
def set_file_limit():
if platform.system() == "Windows":
# Windows has no problem opening huge numbers of files
return
import resource
_, hard_lim = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (hard_lim, hard_lim))