# -*- coding: utf-8 -*-
# (C) Copyright 2020, 2021, 2022, 2023, 2024 IBM. All Rights Reserved.
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""Configurations for resistive processing units."""
# pylint: disable=too-few-public-methods
from dataclasses import dataclass, field
from typing import ClassVar, Type, Optional, Union, Any
from aihwkit.simulator.parameters.pre_post import PrePostProcessingRPU
from aihwkit.simulator.parameters.mapping import MappableRPU
from aihwkit.simulator.parameters.helpers import tile_parameters_to_bindings
from aihwkit.simulator.parameters.runtime import RuntimeParameter
from aihwkit.simulator.configs.devices import (
ConstantStepDevice,
FloatingPointDevice,
IdealDevice,
PulsedDevice,
)
from aihwkit.simulator.configs.compounds import DigitalRankUpdateCell, UnitCell, TransferCompound
from aihwkit.simulator.parameters import (
IOParameters,
IOParametersIRDropT,
PulseType,
UpdateParameters,
WeightClipParameter,
WeightModifierParameter,
WeightRemapParameter,
)
from aihwkit.inference import (
BaseDriftCompensation,
BaseNoiseModel,
GlobalDriftCompensation,
PCMLikeNoiseModel,
)
from aihwkit.simulator.tiles import AnalogTile, FloatingPointTile, InferenceTile, TorchInferenceTile
from aihwkit.simulator.tiles.torch_tile import TorchSimulatorTile
from aihwkit.simulator.tiles.torch_tile_irdrop_t import TorchSimulatorTileIRDropT
from aihwkit.simulator.tiles.array import TileModuleArray
[docs]@dataclass
class FloatingPointRPUConfig(MappableRPU, PrePostProcessingRPU):
"""Configuration for a floating point resistive processing unit."""
tile_class: Type = FloatingPointTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
device: FloatingPointDevice = field(default_factory=FloatingPointDevice)
"""Parameter that modify the behavior of the pulsed device."""
[docs]@dataclass
class IOManagedRPUConfig(MappableRPU, PrePostProcessingRPU):
"""Configuration for an analog (pulsed device) resistive processing unit."""
bindings_class: ClassVar[Optional[Union[Type, str]]] = "AnalogTileParameter"
bindings_module: ClassVar[Optional[str]] = "devices"
forward: IOParameters = field(
default_factory=IOParameters, metadata=dict(bindings_include=True)
)
"""Input-output parameter setting for the forward direction."""
backward: IOParameters = field(
default_factory=IOParameters, metadata=dict(bindings_include=True)
)
"""Input-output parameter setting for the backward direction."""
update: UpdateParameters = field(
default_factory=UpdateParameters, metadata=dict(bindings_include=True)
)
"""Parameter for the update behavior."""
[docs] def as_bindings(self) -> Any:
"""Return a representation of this instance as a simulator bindings object."""
if not hasattr(self, "runtime"):
# legacy
self.runtime = RuntimeParameter()
return tile_parameters_to_bindings(self, self.runtime.data_type)
[docs]@dataclass
class SingleRPUConfig(IOManagedRPUConfig):
"""Configuration for an analog (pulsed device) resistive processing unit."""
tile_class: Type = AnalogTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
device: PulsedDevice = field(default_factory=ConstantStepDevice)
"""Parameter that modify the behavior of the pulsed device."""
[docs]@dataclass
class UnitCellRPUConfig(IOManagedRPUConfig):
"""Configuration for an analog (unit cell) resistive processing unit."""
tile_class: Type = AnalogTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
device: Union[UnitCell, TransferCompound] = field(default_factory=UnitCell)
"""Parameter that modify the behavior of the pulsed device."""
[docs]@dataclass
class DigitalRankUpdateRPUConfig(IOManagedRPUConfig):
"""Configuration for an analog (unit cell) resistive processing unit
where the rank update is done in digital.
Note that for forward and backward, an analog crossbar is still
used, and during update the digitally computed rank update is
transferred to the analog crossbar using pulses.
"""
tile_class: Type = AnalogTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
device: DigitalRankUpdateCell = field(default_factory=DigitalRankUpdateCell)
"""Parameter that modify the behavior of the pulsed device."""
[docs]@dataclass
class InferenceRPUConfig(IOManagedRPUConfig):
"""Configuration for an analog tile that is used only for inference.
Training is done in *hardware-aware* manner, thus using only the
non-idealities of the forward-pass, but backward and update passes
are ideal.
During inference, statistical models of programming, drift
and read noise can be used.
"""
# pylint: disable=too-many-instance-attributes
tile_class: Type = InferenceTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
forward: IOParameters = field(
default_factory=IOParameters, metadata=dict(bindings_include=True)
)
"""Input-output parameter setting for the forward direction.
This parameters govern the hardware definitions specifying analog
MVM non-idealities.
Note:
This forward pass is applied equally in training and
inference. In addition, materials effects such as drift and
programming noise can be enabled during inference by
specifying the ``noise_model``
"""
noise_model: BaseNoiseModel = field(default_factory=PCMLikeNoiseModel)
"""Statistical noise model to be used during (realistic) inference.
This noise models establishes a phenomenological model of the
material which is applied to the weights during inference only, when
``program_analog_weights`` or ``drift_analog_weights`` is called.
"""
drift_compensation: Optional[BaseDriftCompensation] = field(
default_factory=GlobalDriftCompensation
)
"""For compensating the drift during inference only."""
clip: WeightClipParameter = field(default_factory=WeightClipParameter)
"""Parameter for weight clip.
If a clipping type is set, the weights are clipped according to
the type specified.
Caution:
The clipping type is set to ``None`` by default, setting
parameters of the clipping will not be taken into account, if
the clipping type is not specified.
"""
remap: WeightRemapParameter = field(default_factory=WeightRemapParameter)
"""Parameter for remapping.
Remapping can be enabled by specifying a remap ``type``. If
enabled, it ensures that the weights are mapped maximally into the
conductance units during training. It will be called after each mini-batch.
"""
modifier: WeightModifierParameter = field(default_factory=WeightModifierParameter)
"""Parameter for weight modifier.
If a modifier type is set, it is called once per mini-match in the
``post_update_step`` and modifies the weight in forward and
backward direction for the next mini-batch during training, but
updates hidden reference weights. In eval mode, the reference
weights are used instead for forward.
The modifier is used to do hardware-aware training, so that the
model becomes more noise robust during inference (e.g. when the
``noise_model`` is employed).
"""
# The following fields are not included in `__init__`, and should be
# treated as read-only.
device: IdealDevice = field(default_factory=IdealDevice, init=False)
"""Parameter that modify the behavior of the pulsed device: ideal device."""
backward: IOParameters = field(
default_factory=lambda: IOParameters(is_perfect=True),
init=False,
metadata=dict(bindings_include=True),
)
"""Input-output parameter setting for the backward direction: perfect."""
update: UpdateParameters = field(
default_factory=lambda: UpdateParameters(pulse_type=PulseType.NONE),
init=False,
metadata=dict(bindings_include=True),
)
"""Parameter for the update behavior: ``NONE`` pulse type."""
[docs] def compatible_with(self, tile_class_name: str) -> bool:
if tile_class_name in ["TorchInferenceTile"]:
return True
return tile_class_name == self.tile_class.__name__
[docs]@dataclass
class TorchInferenceRPUConfig(InferenceRPUConfig):
"""TorchInference configuration.
This configuration defaults to a tile module implementation that
supported a subset of functions of the ``InferenceRPUConfig`` but
uses native torch instead of the RPUCuda library for simulating
the analog MVM.
The advantage is that autograd is more fully supported and
hardware aware training is more flexible to be modified. However,
some nonidealities are not supported.
Note:
For features that are not supported a ``NotImplementedError`` or a
``TorchTileConfigError`` is raised.
"""
simulator_tile_class: Type = TorchSimulatorTile
tile_class: Type = TorchInferenceTile
"""Tile class that corresponds to this RPUConfig."""
tile_array_class: Type = TileModuleArray
"""Tile class used for mapped logical tile arrays."""
[docs]@dataclass
class TorchInferenceRPUConfigIRDropT(TorchInferenceRPUConfig):
"""Inference configuration using time-dependent IR drop.
This configuration defaults to a tile module implementation that
supported a subset of functions of the ``InferenceRPUConfig`` but
uses native torch instead of the RPUCuda library for simulating
the analog MVM.
The advantage is that autograd is more fully supported and
hardware aware training is more flexible to be modified. However,
some nonidealities are not supported.
Note:
For features that are not supported a ``NotImplementedError`` or a
``TorchTileConfigError`` is raised.
"""
simulator_tile_class: Type = TorchSimulatorTileIRDropT
forward: IOParametersIRDropT = field(default_factory=IOParametersIRDropT)
"""Input-output parameter setting for the forward direction.
This parameters govern the hardware definitions specifying analog
MVM non-idealities.
Note:
This forward pass is applied equally in training and
inference. In addition, materials effects such as drift and
programming noise can be enabled during inference by
specifying the ``noise_model``
"""