NumNN.jl
This package provides high-level Neural Network APIs deals with different number representations like Posit, Logarithmic Data Representations, Residual Number System (RNS), and -for sure- the conventional IEEE formats.
Since, the implementation and development process for testing novel number systems on different Deep Learning applications using the current available DP frameworks in easily feasible. An urgent need for an unconventional library that provides both the easiness and complexity of simulating and testing and evaluate new number systems before the hardware design complex process— was resurfaced.
Why Julia?
Julia provides in an unconventional way the ability to simulate new number systems and deploy this simulation to be used as high-level primitive type. Multiple Dispatch provides a unique ability to write a general code then specify the implementation based on the type.
Examples of Multiple Dispatch
julia> aInt = 1; #with Integer type
julia> bInt = 2; #with Integer type
julia> cInt = 1 + 2; #which is a shortcut for cInt = +(1,2)
julia> aFloat = 1.0; #with Float64 type
julia> bFloat = 2.0; #with Flot64 type
julia> aInt + bFloat #will use the method +(::Int64, ::Float64)
3.0Now let's do something more interesting with Posit (continue on the previous example)
julia> using SoftPosit
julia> aP = Posit16(1)
Posit16(0x4000)
julia> bP = Posit16(2)
Posit16(0x5000)
julia> aInt + aP #Note how the result is in Posit16 type
Posit16(0x5000)The output was of type Posit16 because in Julia you can define a Promote Rule which mean when the output can be in either type(s) of the input, promote the output to be of the specified type. Which can be defined as follows:
Base.promote_rule(::Type{Int64}, ::Type{Posit16}) = Posit16This means that the output of an operation on both Int64 and Posit16 should be converted to Posit16.