Source code for brevitas.core.scaling.int_scaling

# Copyright (C) 2023, Advanced Micro Devices, Inc. All rights reserved.
# SPDX-License-Identifier: BSD-3-Clause

from typing import Optional
from typing import Union

import torch
from torch import nn
from torch import Tensor

import brevitas
from brevitas.function.ops import max_int
from brevitas.function.ops import min_int


class IntScaling(brevitas.jit.ScriptModule):
    __constants__ = ['signed', 'narrow_range']

    def __init__(
            self,
            narrow_range: bool,
            signed: Optional[bool] = None,
            dtype: Optional[torch.dtype] = None):
        super(IntScaling, self).__init__()
        self.signed = signed
        self.narrow_range = narrow_range
        self.dtype = dtype

    @brevitas.jit.script_method
    def forward(self, bit_width: Tensor, signed: Optional[Union[bool, Tensor]] = None) -> Tensor:
        is_signed = signed if signed is not None else self.signed
        if is_signed is None:
            raise ValueError(f"signed is not defined, signed={is_signed}")
            is_signed = True  # Workaround for JIT type inference
        # Workaround: required for compatibility with the JIT for PT=2.2.2
        is_signed = bool(is_signed.item()) if isinstance(is_signed, Tensor) else is_signed
        if is_signed:
            return -min_int(is_signed, self.narrow_range, bit_width).to(dtype=self.dtype)
        else:
            return max_int(is_signed, self.narrow_range, bit_width).to(dtype=self.dtype)


class PowerOfTwoIntScaling(brevitas.jit.ScriptModule):
    __constants__ = ['signed']

    def __init__(self, signed: Optional[bool] = None):
        super(PowerOfTwoIntScaling, self).__init__()
        self.signed = signed

    @brevitas.jit.script_method
    def forward(self, bit_width: Tensor, signed: Optional[Union[bool, Tensor]] = None) -> Tensor:
        is_signed = signed if signed is not None else self.signed
        if is_signed is None:
            raise ValueError(f"signed is not defined, signed={is_signed}")
            is_signed = True  # Workaround for JIT type inference
        # Workaround: required for compatibility with the JIT for PT=2.2.2
        is_signed = bool(is_signed.item()) if isinstance(is_signed, Tensor) else is_signed
        return max_int(is_signed, False, bit_width) + 1