Decoding-Free Two-Input Arithmetic for Low-Precision Real Numbers

In this work, we present a novel method for directly computing functions of two real numbers using logic circuits without decoding; the real numbers are mapped to a particularly-chosen set of integer numbers. We theoretically prove that this mapping always exists and that we can implement any kind of binary operation between real numbers regardless of the encoding format. While the real numbers in the set can be arbitrary (rational, irrational, transcendental), we find practical applications to low-precision posit™ number arithmetic. We finally provide examples for decoding-free 4-bit Posit arithmetic operations, showing a reduction in gate count up to a factor of \(7.6\times \) (and never below \(4.4\times \)) compared to a standard two-dimensional tabulation.

Work partially supported by H2020 projects EPI2 (grant no. 101036168, https://www.european-processor-initiative.eu/ and TextaRossa (grant no. 956831, https://textarossa.eu/).

Authors and Affiliations

1. Arizona State University, Tempe, AZ, USA

   John L. Gustafson

2. University of Pisa, Pisa, Italy

   Marco Cococcioni, Federico Rossi & Sergio Saponara

3. MMI s.p.a, Pisa, Italy

   Emanuele Ruffaldi

Corresponding author

Correspondence to Federico Rossi .

Editors and Affiliations

1. Arizona State University, Tempe, AZ, USA

   John Gustafson

2. Swiss National Supercomputing Centre, ETH Zurich, Lugano, Switzerland

   Siew Hoon Leong

3. National Supercomputing Centre, Singapore, Singapore

   Marek Michalewicz

Appendix: How to Build an Initial Feasible Solution

An initial feasible solution (useful to speedup Matlab intlinprog function) can be constructed as shown below. We will focus on the positive values in X, different both from NaR (notice that we are excluding the zero as well). Let us call this set \(\mathcal {X}\). Let us indicate with \(x^{*}_i \in \mathcal {X}\) the corresponding bistring (in the next we will refer to the \(Posit\left\langle 4,0\right\rangle \) case, as an example). Therefore, the bistrings will be the ones of \(Posit\left\langle 4,0\right\rangle \) without its most significant bit (see Table 1).

• Each \(x_i \in \mathcal {X}\) is mapped to the natural number \(L^x_i = x_i^* \cdot 2^n,\) n being the maximum number of bits needed for representing the \(x_i\) (in the case of \(Posit\left\langle 4,0\right\rangle \), \(n = 3\) )

• Each \(y_j \in \mathcal {Y}\) is mapped to the natural number \(L^y_j = y_j^*\)

Therefore, we obtain the \(L^x, L^y\) sets: \(L^x: \{x_1^* \cdot 2^n, \dots , x_{|\mathcal {X}|}^* \cdot 2^n \}\) and \(L^y: \{ y_1^*, \dots , y_{|\mathcal {Y}|}^* \}\). Each \(L_{i,j}^z \in L^z\) is obtained by the concatenation of the bit strings \(x_i^*, y_j^*\) (or equivalently, as \(L_{i,j}^z = L_i^x + L_j^y\), as shown in Table 12).

We now prove that this solution satisfies the constraint given in Eq. (1):

• Since there are no conflicting encodings of the real numbers in \(\mathcal {X}\) and \(\mathcal {Y}\), we can guarantee that different real numbers have different bit-strings that digitally encode them. Therefore, \(L_i^x \ne L_j^x, \forall i \ne j\) and \(L_p^y \ne L_l^y, \forall k \ne l\).

• Since all the encodings in \(L^x, L^y\) are different from each other, also the concatenation of any pair \(L_i^x, L_j^y\) is unique. Therefore, \(L_{i,j}^z \ne L_{k,q}^z, \) if \(x_i \triangledown y_j \ne X_j \triangledown Y_q \) (with \(\triangledown \) we indicate the generic operation for which we are finding the mapping).

• Being the values \(L_{i,j}^z\) unique (no duplicates), we can easily obtain the ordered set \(L^z\), by sorting them.

• the k-element vector \(\boldsymbol{w}\) can be trivially obtained as cast\((x_i \triangledown y_j)\), when \(i\cdot 2^n + j = k\).

An example of feasible solution for the multiplication of \(Posit\left\langle 4,0\right\rangle \) numbers is reported in Table 12.

Reprints and permissions

© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG

Policies and ethics