dpctl.utils

dpctl.utils.onetrace_enabled()

Enable onetrace collection for kernels executed in this context.

Note

Proper working of this utility assumes that Python interpreter has been launched by onetrace or unitrace tool from project intel/pti-gpu.

Example:

Launch the Python interpreter using onetrace tool:

$ onetrace --conditional-collection -v -t --demangle python app.py

Now using the context manager in the Python sessions enables data collection and its output for every offloaded kernel:

import use_kernel as eg

import dpctl.memory as dpmem
from dpctl.utils import onetrace_enabled

x_dev = dpmem.MemoryUSMDevice(1024)

# onetrace output reporting on execution of the kernel
# should be seen, starting with "Device Timeline"
with onetrace_enabled():
    eg.run_kernel(x_dev)

dpctl.utils.intel_device_info(sycl_device)

For Intel(R) GPU devices returns a dictionary with device architectural details, and an empty dictionary otherwise. The dictionary contains the following keys:

device_id:

32-bits device PCI identifier

gpu_eu_count:

Total number of execution units

gpu_hw_threads_per_eu:

Number of thread contexts in EU

gpu_eu_simd_width:

Physical SIMD width of EU

gpu_slices:

Total number of slices

gpu_subslices_per_slice:

Number of sub-slices per slice

gpu_eu_count_per_subslice:

Number of EUs in subslice

max_mem_bandwidth:

Maximum memory bandwidth in bytes/second

free_memory:

Global memory available on the device in units of bytes

Unsupported descriptors are omitted from the dictionary.

Descriptors other than the PCI identifier are supported only for SyclDevices with Level-Zero backend.

Note

Environment variable ZES_ENABLE_SYSMAN may need to be set to 1 for the "free_memory" key to be reported.

dpctl.utils.SequentialOrderManager

Thread-local object mapping each dpctl.SyclQueue to an order manager, used to ensure sequential ordering of offloaded tasks.

Record submitted tasks with add_event and use submitted_events as the dependency list of subsequent submissions. To keep Python objects referenced by a task alive until it completes, use dpctl.SyclQueue.keep_args_alive().

Record events that gate the release of objects used by a task with add_cleanup_event, and find them in cleanup_events. They are waited on, but never become dependencies of later tasks.

Waiting with wait also drops the references that dpctl.SyclQueue.keep_args_alive() took for tasks that have since completed, which dpctl otherwise does on a thread of its own.

Deprecated since version 0.23.0: add_event_pair, host_task_events and num_host_task_events are deprecated. Tasks are no longer paired with a host task event, so add_event takes the computational event alone, and cleanup is tracked by add_cleanup_event, cleanup_events and num_cleanup_events.