qkeras.qtools.quantized_operators.multiplier_impl

multiplier operation implementations.

Functions

assert_neither_input_and_weights_is_floating_point(...)

assert non float type.

Classes

Adder(weight_quantizer, input_quantizer, ...)	adder implementation.
AndGate(weight_quantizer, input_quantizer, ...)	and gate implementation.
FixedPointMultiplier(weight_quantizer, ...)	multiplier for fixed point.
FloatingPointMultiplier(weight_quantizer, ...)	multiplier for floating point.
IMultiplier(weight_quantizer, ...)	abstract class for multiplier.
Mux(weight_quantizer, input_quantizer, ...)	Use mux for the hardware implementation of multiplier.
Shifter(weight_quantizer, input_quantizer, ...)	shifter gate.
XorGate(weight_quantizer, input_quantizer, ...)	Use XorGate for hardware implementation of a multiplier.

class qkeras.qtools.quantized_operators.multiplier_impl.Adder(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

adder implementation.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.AndGate(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

and gate implementation.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.FixedPointMultiplier(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

multiplier for fixed point.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.FloatingPointMultiplier(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

multiplier for floating point.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.IMultiplier(weight_quantizer, input_quantizer, output_quantizer)

Bases: ABC

abstract class for multiplier.

This class is about how multiplier is implemented in hardware, which can be

mux gate, shifter, adder, etc.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

abstractmethod static implemented_as()

name()

Return type:

str

output_quantizer()

class qkeras.qtools.quantized_operators.multiplier_impl.Mux(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

Use mux for the hardware implementation of multiplier.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.Shifter(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

shifter gate.

po2*qbit is implemented as a shifter. output is qbits type.

determin number of bits in the output qbits type:

1. min_exp in po2: number of bits to be expanded on the

   right (decimal bits) in qbits for example, min_exp = -2 -> po2 =2^min_exp = 2^(-2) : this means, po2*qbit -> qbit value right shifted for 2 bits

2. max_exp in po2: number of bits to be expanded on

   the left (int_bits) in qbits

How to calculate min_exp and max_exp:

1.if po2 is_signed (quantized_po2)

one bit for sign for the entire po2 value; *exp has non_sign_bits = bits - 1 number of bits, *furthermore, 1 bit from non_sign_bits is used as sign bit in exp; *value range for exp is [-2 * (non_sign_bits - 1),

2 ** (non_sign_bits - 1) - 1]

2.if not_signed (quantized_relu_po2)

• 0 bit for the entire po2 value

• exp has non_sign_bits = bits

• rest is the same as above

determine sign bit in the output qbits:

1. qbits no_sign and po2 is_sign: since max_exp and min_exp

   are computed without sign bit

   we need to add 1 sign bit to the final result;

2. qbits is_sign: since qbits already has a sign bit,

   no extra sign bit needed

3. qbits no_sign and po2 no_sign: no extra sign bit needed

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

input

input_quantizer

weight

weight_quantizer

output

output_quantizer

gate_factor

relative energy comparing to an Adder

gate_bits

number of bits for energy calculation.

static implemented_as()

class qkeras.qtools.quantized_operators.multiplier_impl.XorGate(weight_quantizer, input_quantizer, output_quantizer)

Bases: IMultiplier

Use XorGate for hardware implementation of a multiplier.

Parameters:

• weight_quantizer (IQuantizer)

• input_quantizer (IQuantizer)

• output_quantizer (IQuantizer)

static implemented_as()

qkeras.qtools.quantized_operators.multiplier_impl.assert_neither_input_and_weights_is_floating_point(multiplier)

assert non float type.

Parameters:

multiplier (IMultiplier)