#!/usr/bin/env python
# ******************************************************************************
# Copyright 2022 Brainchip Holdings Ltd.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ******************************************************************************
import tensorflow as tf
from .qtensor import QTensor, saturate, round_through, floor_through, pow2, ceil_log2
from ..debugging import assert_equal, assert_none_equal, assert_less, assert_less_equal
[docs]class FixedPoint(QTensor):
r"""A Tensor of integer values representing fixed-point numbers
The value_bits parameter sets the maximum integer values that can be stored:
.. math:: int\_max = 2^{bits} - 1.
When a FixedPoint is created, its values are clipped to [-int_max-1, int_max].
Args:
values (tf.Tensor): a tensor of integer values
value_bits (int): the number of value bits.
frac_bits (tf.Tensor): an integer tensor of fractional bits.
"""
values: tf.Tensor = 1.0
value_bits: int = 7
frac_bits: tf.Tensor = 0.
def __init__(self, values, value_bits, frac_bits):
# We store integer values in a float tensor to speed up calculations
if isinstance(values, tf.Tensor):
values = tf.cast(values, tf.float32)
else:
values = tf.convert_to_tensor(values, dtype=tf.float32)
# Clamp to fixed-point boundaries
self.values = saturate(values, tf.cast(value_bits, tf.float32))
# We store fractional bits in a float tensor to speed up calculations
if isinstance(frac_bits, tf.Tensor):
self.frac_bits = tf.cast(frac_bits, tf.float32)
else:
self.frac_bits = tf.convert_to_tensor(frac_bits, tf.float32)
self.value_bits = value_bits
self.shape = self.values.shape
@property
def name(self):
return "no-name" if not hasattr(self.values, "name") else self.values.name
@property
def per_tensor(self):
return self.frac_bits.shape.ndims == 0 or self.frac_bits.shape == (1,)
[docs] @staticmethod
def max_frac_bits(value_bits, ranges, clamp=True):
"""Evaluate the maximum fractional bit index for the quantization ranges.
This method evaluates the minimum number of integer bits required to cover the specified
quantization ranges (this can be a negative number if the ranges are strictly lower than
0.5).
From that it deduces the rightmost fractional bit indices.
The resulting frac_bits can be a negative number if the ranges are higher than the biggest
integer than can be represented with the specified value bits).
If specified, the maximum fractional bits are clamped to the available value_bits.
Args:
value_bits (int): the number of value bits.
ranges (tf:Tensor): a tensor of float quantization ranges.
clamp (bool, optional): clamp the results to self.value_bits. Defaults to True.
Returns:
tf:Tensor: a tensor of fractional bits.
"""
# Get the powers of two containing the quantization ranges
int_bits = ceil_log2(ranges)
# Evaluate the rightmost fractional bits (they can be negative)
value_bits = tf.cast(value_bits, tf.float32)
frac_bits = value_bits - int_bits
if clamp:
frac_bits = tf.minimum(frac_bits, value_bits)
return frac_bits
[docs] @staticmethod
def quantize(x, value_bits, frac_bits=None):
r"""Converts a float Tensor to a FixedPoint
It converts the original float values into integer values so that:
.. math:: {x_{int}} = round(x * 2^{frac\_bits})
Note: :math:`2^{-frac\_bits}` represents the FixedPoint precision.
Before returning, the resulting integer values are clipped to the
maximum integer values that can be stored for the specified value bits:
.. math:: [-2^{value\_bits}, 2^{value\_bits} - 1]
If frac_bits is not specified, the method starts by evaluating the number
of bits to dedicate to represent the integer part of the float tensor,
clipped to value_bits:
.. math:: int\_bits = ceil(log2(x))
Note: this number can be negative when x < 0.5.
It then evaluates the offset of the least significant bit of the fractional
part of the float starting from zero. This represents the fractional bits:
.. math:: frac\_bits = value\_bits - int\_bits
Args:
x (tf.Tensor): a tensor of float values.
value_bits (int): the number of value bits
frac_bits (tf.Tensor, optional): an integer tensor of fractional bits.
Defaults to None.
Returns:
:obj:`FixedPoint`: the FixedPoint tensor
"""
if frac_bits is None:
if isinstance(x, int) or tf.reduce_all(x == tf.math.round(x)):
# The input does not need to be quantized
frac_bits = 0
else:
frac_bits = FixedPoint.max_frac_bits(value_bits, tf.math.reduce_max(tf.abs(x)))
# Project float into fixed-point representation space
x_scaled = tf.math.multiply(x, pow2(frac_bits), name="quantize")
# Round or ceil to obtain integer values
values = round_through(x_scaled)
return FixedPoint(values, value_bits, frac_bits)
@property
def sign(self):
"""Returns the sign of the FixedPoint
Returns:
:obj:`FixedPoint`: the sign as a FixedPoint.
"""
return FixedPoint(tf.math.sign(self.values), self.value_bits, 0)
[docs] def to_float(self):
return self.values / pow2(self.frac_bits)
[docs] def align(self, other, value_bits=None):
"""Align fractional bits
This returns an equivalent FixedPoint with a scalar fractional bit
corresponding to the maximum of the current and other FixedPoint on all
channels.
This is required before performing an operation that adds or subtracts
elements along the last dimension, to make sure all these elements are
in the same scale.
Args:
other (:obj:`FixedPoint`): a FixedPoint to align to
value_bits (int, optional): the target value bits. Defaults to None.
Returns:
tuple(:obj:`FixedPoint`, tf.Tensor): a new FixedPoint with aligned
fractional bits and the shift that was applied.
"""
if not isinstance(other, FixedPoint):
raise ValueError("Other must be a FixedPoint.")
max_frac_bits = tf.math.maximum(self.frac_bits, other.frac_bits)
return self.upscale(max_frac_bits, value_bits)
[docs] def downscale(self, frac_bits):
"""Reduce the precision of a FixedPoint
Args:
frac_bits (tf.Tensor): the target fractional bits
Returns:
:obj:`FixedPoint`: the downscaled FixedPoint
"""
frac_bits = tf.cast(frac_bits, tf.float32)
assert_less_equal(frac_bits, self.frac_bits,
f"Cannot reduce {self.name} precision because the target precision "
f"({frac_bits}) is higher than the current one ({self.frac_bits})")
return self.rescale(frac_bits)
[docs] def rescale(self, frac_bits, value_bits=None):
"""Rescale a FixedPoint to a specified precision and bitwidth
This primarily rescales the FixedPoint values to match the precision
specified by the target fractional bits.
Optionally, this adjusts the value bits to the specified bitwidth.
The rescaling operation is:
- a left shift of the values when their precision increases,
- a rounded right shift of the values when their precision decreases.
This method can be used to:
- compress a FixedPoint to a lower bitwidth after having reduced its precision,
- expand a FixedPoint to a larger bitwidth after having increased its precision.
Args:
frac_bits (tf.Tensor): the target fractional bits
value_bits (int, optional): the target value bits
Returns:
:obj:`FixedPoint`: the rescaled FixedPoint
"""
if value_bits is None:
value_bits = self.value_bits
frac_bits = tf.cast(frac_bits, tf.float32)
# Evaluate the shift to apply to reach the target precision
shift = frac_bits - self.frac_bits
# The shift can be positive (left-shift) or negative (rounded right-shift)
# A positive shift exceeding the target bitwidth always leads to a saturation
assert_less(shift, value_bits,
f"Cannot rescale {self.name} to {frac_bits} fractional bits as it will saturate"
f" its {value_bits} buffer")
# The integer operations are simulated in float arithmetics:
# - the shifts by multiplying by the positive/negative shift power of two,
# - the rounding by adding 0.5 and flooring (to workaround banker's rounding).
# For positive shift, the results being integer, the rounding has no effect.
# We therefore apply the same operations for both shifts.
values = tf.math.multiply(self.values, pow2(shift), name="rescale")
values = floor_through(values + 0.5)
# return a new FixedPoint with the target precision and bitwidth
return FixedPoint(values, value_bits, frac_bits), shift
[docs] def upscale(self, frac_bits, value_bits=None):
"""Align a FixedPoint to a specified precision
The target precision must be higher than the current one.
Args:
frac_bits (tf.Tensor): the target fractional bits
value_bits (int, optional): the target value bits
Returns:
:obj:`FixedPoint`: the upscaled FixedPoint
"""
if value_bits is None:
value_bits = self.value_bits
frac_bits = tf.cast(frac_bits, tf.float32)
assert_less_equal(self.frac_bits, frac_bits,
f"Cannot increase {self.name} precision because the target precision "
f"({frac_bits}) is lower than the current one ({self.frac_bits})")
# Evaluate the shift to apply to reach the target precision
shift = frac_bits - self.frac_bits
assert_less(shift, value_bits,
f"Cannot upscale {self.name} to {frac_bits} fractional bits as it saturates its"
f" {value_bits}-bit buffer.")
# Apply shift
values = tf.math.multiply(self.values, pow2(shift), name="upscale")
# return a new FixedPoint with the target precision and bitwidth
return FixedPoint(values, value_bits, frac_bits), shift
[docs] def expand(self, value_bits):
"""Expand the FixedPoint to the specified bitwidth
This returns an equivalent FixedPoint with a higher or equal number of
value bits and a scalar fractional bit corresponding to the maximum of
the initial fractional bits on all channels.
This is mostly used to recover a per-tensor FixedPoint that has been
compressed to a lower number of value bits.
Args:
value_bits (int): the target value_bits
Returns:
tuple(:obj:`FixedPoint`, tf.Tensor): a new FixedPoint with expanded
fractional bits and the shift that was applied.
"""
if value_bits < self.value_bits:
raise ValueError(
f"Cannot reduce {self.name} bitwidth from {self.value_bits} to {value_bits}:"
" use a quantizer instead.")
max_frac_bits = tf.reduce_max(self.frac_bits)
return self.upscale(max_frac_bits, value_bits)
@staticmethod
def _rshift(values, shift):
return floor_through(values / pow2(shift))
@staticmethod
def _lshift(values, shift):
return tf.math.multiply(values, pow2(shift), name="lshift")
[docs] def shift(self, s):
"""Apply a tensor-wide left or right shift.
This takes a tensor of shift values and apply them on each item of the
FixedPoint values.
The shift values should positive or negative integer:
- if the value is positive, it is a left-shift,
- if the value is negative, it is a right-shift.
The resulting FixedPoint has the same value bits and fractional bits as
the source FixedPoint, which means that clipping is applied on
left-shift and flooring is applied on right-shift.
Args:
s (tf.Tensor): the shift values for each pixel.
Returns:
:obj:`FixedPoint`: the result as a FixedPoint
"""
values = tf.math.multiply(self.values, pow2(s), name="shift")
values = floor_through(values)
return FixedPoint(values, self.value_bits, self.frac_bits)
def __rshift__(self, shift):
"""Right shift the FixedPoint values
This operation has no direct equivalent in float arithmetics: it corresponds to a division
of the corresponding float by a power-of-two, then a flooring to the quantization interval.
Args:
shift (tf.Tensor): the power of 2 to divide by
Returns:
:obj:`FixedPoint`: the result as a FixedPoint
"""
shift = tf.cast(shift, tf.float32)
assert_less_equal(0, shift, "Shift must be all positive")
assert_equal(tf.rank(shift) <= tf.rank(self.frac_bits), True,
"The shift's rank must be less than or equal to the rank of frac_bits. "
f"Received {tf.rank(shift)} > {tf.rank(self.frac_bits)}."
"That means it is not possible to fold the shift into the FixedPoint. "
"Please use FixedPoint.shift instead of '>>'.")
# The shift can be folded into the fractional bits,
s_frac_bits = self.frac_bits + shift
# keeping the same values
s_values = self.values
# Return a new FixedPoint with updated fractional bits,
# which is equivalent in hardward without performing any operation
return FixedPoint(s_values, self.value_bits, s_frac_bits)
def __lshift__(self, shift):
"""Left shift the FixedPoint values
This operation has no direct equivalent in float arithmetics: it corresponds to a
multiplication of the corresponding float by a power-of-two, then a flooring to the
quantization interval.
Args:
shift (tf.Tensor): the power of 2 to multiply by
Returns:
:obj:`FixedPoint`: the result as a FixedPoint
"""
assert_less_equal(0, shift, "Shift must be all positive")
# Simply apply the shift on the values
s_values = FixedPoint._lshift(self.values, shift)
# Return a new FixedPoint with updated values
return FixedPoint(s_values, self.value_bits, self.frac_bits)
def _align_values(self, other):
# The sub fractional bits are the max of both terms
frac_bits = tf.math.maximum(self.frac_bits, other.frac_bits)
self_values = FixedPoint._lshift(
self.values, (frac_bits - self.frac_bits))
other_values = FixedPoint._lshift(
other.values, (frac_bits - other.frac_bits))
return frac_bits, self_values, other_values
def __add__(self, other):
if isinstance(other, int):
# Convert integer into a 32-bit fixed-point with no fractional bits,
# aligned with the current FixedPoint
return self + FixedPoint.quantize(other, 32, self.frac_bits)
elif isinstance(other, FixedPoint):
# Check that self and other are aligned
assert_equal(self.frac_bits, other.frac_bits,
message=f"{self.name} and {other.name} are not aligned")
# Return a new FixedPoint
return FixedPoint(self.values + other.values, self.value_bits, self.frac_bits)
raise TypeError(
f"Unsupported operand type(s) for +: 'FixedPoint' and '{type(other)}'")
def __sub__(self, other):
if isinstance(other, int):
# Convert integer into a 32-bit fixed-point with no fractional bits,
# aligned with the current FixedPoint
return self - FixedPoint.quantize(other, 32, self.frac_bits)
elif isinstance(other, FixedPoint):
# Check that self and other are aligned
assert_equal(self.frac_bits, other.frac_bits,
message=f"{self.name} and {other.name} are not aligned")
# Return a new FixedPoint
return FixedPoint(self.values - other.values, self.value_bits, self.frac_bits)
raise TypeError(
f"Unsupported operand type(s) for -: 'FixedPoint' and '{type(other)}'")
def __truediv__(self, other):
@tf.function
@tf.custom_gradient
def truncate(x):
# Remove decimal part from x, This is to obtain a value that
# matches the result estimated in C when working with ints.
rounded = tf.cast(tf.cast(x, tf.int32), tf.float32)
def grad(upstream):
return upstream
return rounded, grad
if isinstance(other, int):
return self / FixedPoint(other, 32, 0)
elif isinstance(other, FixedPoint):
assert_none_equal(other.values, 0.0, "Cannot divide by 0.")
# The division between fixed-point is straight-forward
d_values = truncate(self.values / other.values)
# Return a new FixedPoint whose frac bits from other is subtracted.
return FixedPoint(d_values, self.value_bits, self.frac_bits - other.frac_bits)
raise TypeError(
f"Unsupported operand type(s) for /: 'FixedPoint' and '{type(other)}'")
def __pow__(self, power):
if isinstance(power, int):
if power == 0:
return FixedPoint(tf.ones_like(self.values), self.value_bits, 0)
elif power == 1:
return FixedPoint(self.values, self.value_bits, self.frac_bits)
elif power > 1:
return self * self ** (power - 1)
else:
raise NotImplementedError(
"Negative powers are not implemented yet")
raise TypeError(
f"Unsupported operand type(s) for **: 'FixedPoint' and '{type(power)}'")
def __gt__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"Unsupported operand type(s) for >: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values > o_values
def __ge__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"Unsupported operand type(s) for >=: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values >= o_values
def __eq__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"Unsupported operand type(s) for ==: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values == o_values
def __ne__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"unsupported operand type(s) for !=: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values != o_values
def __lt__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"Unsupported operand type(s) for <: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values < o_values
def __le__(self, other):
if not isinstance(other, FixedPoint):
raise TypeError(
f"Unsupported operand type(s) for <=: 'FixedPoint' and '{type(other)}'")
_, s_values, o_values = self._align_values(other)
return s_values <= o_values
[docs] def abs(self):
"""Returns the absolute value of the FixedPoint
Returns:
:obj:`FixedPoint`: the absolute value.
"""
return FixedPoint(tf.math.abs(self.values), self.value_bits, self.frac_bits)
[docs] def floor(self):
"""Floors the FixedPoint
Returns:
tuple(:obj:`FixedPoint`, tf.Tensor): a new FixedPoint without
fractional bits and the shift that was applied.
"""
# Divide values to remove fractional bits
values = FixedPoint._rshift(self.values, self.frac_bits)
return FixedPoint(values, self.value_bits, 0), self.frac_bits