hysop.backend.device.opencl.opencl_kernel_launcher module

class hysop.backend.device.opencl.opencl_kernel_launcher.HostLauncherI(name, **kwds)

Bases: LauncherI

Create a OpenClKernelLauncher.

Parameters:

• name (str) – A name for this kernel launcher (for logging purposes).

• kwds (dict) – Base class arguments.

global_size_configured()

Return True is this kernel is ready to be enqueued without specifying global_work_size (local_work_size is always optional).

parameters()

class hysop.backend.device.opencl.opencl_kernel_launcher.LauncherI(name, profiler=None, **kwds)

Bases: object

Interface for any object that has the ability to be a launcher.

Create a OpenClKernelLauncher.

Parameters:

• name (str) – A name for this kernel launcher (for logging purposes).

• kwds (dict) – Base class arguments.

abstract __call__(queue=None, wait_for=None, **kwds)

Launch with a specific queue. Wait for wait_for events before computing. If queue has profiling enabled, events are pushed into a local list of events to compute kernel statistics when self.statistics is fetched.

property events

All events since the last call to update statistics.

abstract global_size_configured()

Return True is this kernel is ready to be enqueued without specifying global_work_size (local_work_size is always optional).

property name

Get the name of this kernel launcher.

property statistics

Compute statistics and clear events of kernels that finished.

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClIterativeKernelLauncher(parameters_map, args_list, iterated_parameters, **kwds)

Bases: OpenClParametrizedKernelLauncher

Wraps an OpenCL kernel ready to be iteratively enqueued with some extra kernel arguments. Extra kernel arguments may be splitted between:

*classic extra arguments like with OpenClParametrizedKernelLauncher *iterated arguments that automatically yielded for each __apply__

An OpenClIterativeKernelLauncher is acting like an OpenClKernelListLauncher without the need of generating the list of all kernels with fixed arguments.

Create a OpenClIterativeKernelLauncher.

Parameters:

• args_list (tuple) – Known arguments of the kernel, as a tuple of tuples (arg_index, arg_value).

• parameters_map (dict) – Mapping between unknown parameter names and (parameter_index, parameter_type).

• iterated_parameters (dict) – Mapping between iterated parameter names and OpenClKernelParameterGenerator. Iterated parameters should be included in parameters_map as well.

• kwds (dict) – Base class parameters.

__call__(queue=None, wait_for=None, enqueue_barrier=True, global_work_size=None, local_work_size=None, **kwds)

Launch kernel with a specific queue, global_work_size and local_work_size. Wait for wait_for events before computing. If queue has profiling enabled, events are pushed into a local list of events to compute kernel statistics when self.statistics is fetched.

If the queue is out of order, a barrie is enqueued unless enqueue_barrier is set to False. If enqueue_barrier is False, returned event is None.

iter_parameters()

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClKernelLauncher(name, kernel, args_list, default_global_work_size=None, default_local_work_size=None, default_queue=None, **kwds)

Bases: OpenClKernelLauncherI

Wraps an OpenCL kernel ready to be enqueued without extra arguments. Manage launching of one OpenCL kernel with fixed arguments.

Create a OpenClKernelLauncher.

Parameters:

• name (str) – A name for this kernel launcher (for logging purposes).

• kernel (cl.Program or cl.Kernel) – The precompiled program that contains the kernel to be launched. If a program is passed we can create a unique instance of a kernel that will have its arguments already set.

• args_list (tuple) – All arguments of the kernel, in the right order as a tuple or tuple of tuples (arg_index, arg_value).

• default_queue (cl.CommandQueue, optional) – Default queue to run the kernel.

• default_global_work_size (tuple of ints, optional) – Default global work size.

• default_local_work_size (tuple of ints, optional) – Default local work size.

• kwds (dict) – Base class arguments.

__call__(queue=None, wait_for=None, global_work_size=None, local_work_size=None, **kwds)

Launch kernel with a specific queue, global_work_size and local_work_size. Wait for wait_for events before computing. If queue has profiling enabled, events are pushed into a local list of events to compute kernel statistics when self.statistics is fetched.

property args_list

All arguments of the kernel as a tuple.

as_list_launcher(name)

Convert a OpenClKernelLauncher to a OpenClKernelListLauncher.

property default_queue

Default queue to launch the kernel.

global_size_configured()

Return True is this kernel is ready to be enqueued without specifying global_work_size (local_work_size is always optional).

property kernel

Get the precompiled kernel to be launched.

queue_configured()

Return True is this kernel is ready to be enqueued without specifying a queue.

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClKernelLauncherI(name, profiler=None, **kwds)

Bases: LauncherI

Interface for any object that has the ability to enqueue a OpenCL kernel without extra arguments.

Create a OpenClKernelLauncher.

Parameters:

• name (str) – A name for this kernel launcher (for logging purposes).

• kwds (dict) – Base class arguments.

abstract __call__(queue=None, wait_for=None, global_work_size=None, local_work_size=None, **kwds)

Launch kernel with a specific queue. Wait for wait_for events before computing. If queue has profiling enabled, events are pushed into a local list of events to compute kernel statistics when self.statistics is fetched.

check_kernel_arg(arg, arg_id, arg_name, arg_type)

Check kernel argument type prior to setargs.

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClKernelListLauncher(name, profiler=None)

Bases: object

Wraps multiples OpenCL kernel ready to be enqueued without extra arguments. All kernels contained in a KernelListLauncher are called with the same queue (ie. OpenClKernelLauncher default_queue is not taken into account and all kernels

should at least have a pre-configured global_work_size).

Create a OpenClKernelListLauncher.

Parameters:

• name (str) – A name for this kernel list launcher (for logging purposes).

• kwds (dict) – Base class arguments.

__call__(queue, wait_for=None, **kwds)

Enqueue all kernels on the given queue in order. The first enqueued kernel will wait on the wait_for events. If this OpenClKernelListLauncher is empty, cl.wait_for_events will be called instead.

__iadd__(kernel)

Push a kernel into the list.

property name

Return the OpenClKernelLauncher name.

property parameters

Return parameters of OpenClParametrizedKernelLauncher. This is a mapping between the parameter names and parameter types and operator names.

push_copy_device_to_device(varname, src, dst, src_device_offset=None, dst_device_offset=None, byte_count=None)

Shortcut for OpenClCopyDevice2Device kernels creation.

push_copy_device_to_host(varname, src, dst, src_device_offset=None)

Shortcut for OpenClCopyDevice2Host kernels creation.

push_copy_host_device(varname, src, dst, src_device_offset=None, dst_device_offset=None, byte_count=None)

Shortcut for OpenClCopyBuffer kernels creation.

push_copy_host_to_device(varname, src, dst, dst_device_offset=None)

Shortcut for OpenClCopyHost2Device kernels creation.

push_kernels(*kernel_launchers)

Push OpenClKernelLaunchers into the list. None values are ignored for convenience.

property statistics

Compute statistics of each kernels and clear events of kernels that finished.

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClKernelParameterGenerator

Bases: object

Abstract base for opencl kernel parameter yielders.

abstract new_generator()

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClKernelParameterYielder(fn)

Bases: OpenClKernelParameterGenerator

Generate opencl parameters through an external iterator or generator factory.

Create a OpenClKernelParameterYielder.

Parameters:

fn (callable) –

Lambda, function or functor that takes no arguments which should return a Generator or an Iterator uppon call.

Example:

lambda: range(10)

new_generator()

class hysop.backend.device.opencl.opencl_kernel_launcher.OpenClParametrizedKernelLauncher(parameters_map, args_list, **kwds)

Bases: OpenClKernelLauncher

Wraps an OpenCL kernel ready to be enqueued with some extra kernel arguments.

Create a OpenClParametrizedKernelLauncher.

Parameters:

• args_list (tuple) – Known arguments of the kernel, as a tuple of tuples (arg_index, arg_value).

• parameters_map (dict) – Mapping between unknown parameter names and (parameter_index, parameter_type).

• kwds (dict) – Base class parameters.

property parameters_map

Mapping between parameter names and (parameter_index, parameter_type).

hysop.backend.device.opencl.opencl_kernel_launcher.profile_kernel(kernel, evt, kernel_msg=None, fprofiler=None)

hysop.backend.device.opencl.opencl_kernel_launcher.should_profile_kernel(kernel_msg)

hysop.backend.device.opencl.opencl_kernel_launcher.should_trace_kernel(kernel_msg)

hysop.backend.device.opencl.opencl_kernel_launcher.trace_kernel(kernel_msg)