US 11,416,218 B1Granted

Digital approximate squarer for machine learning

OwnerAli Tasdighi Far, Jerry M. Collingsinventors

Life of this patent

Estimating the term…

Loading patent…

Abstract

Digital approximate squarer (aSQR)s utilizing apparatuses, circuits, and methods are described in this disclosure. The disclosed aSQR methods can operate asynchronously and or synchronously. For applications where low precisions is acceptable, fewer interpolations can yield less precise square approximation, which can be computed faster and with lower power consumption. Conversely, for applications where higher precision are required, more interpolations steps can generate more precise square approximation. By utilizing the disclosed aSQR method, precision objectives of a squarer approximation function can be programmed real-time and on the fly, which enables optimizing for power consumption and speed of squaring, in addition to optimize for the approximate squarer's die size and cost.

Claims5

Dependent claims

An approximate digital squaring (aSQR) method in a digital state machine in an integrated circuit (IC), the digital state machine in the integrated circuit (IC) comprising at least one of an input port, an output port, a digital memory apparatus, a digital adder apparatus, a digital subtractor apparatus, a digital divider apparatus, a digital accumulator apparatus, and a digital maximum selector apparatus, the method comprising:

receiving a digital input word (D) at the input port, wherein the digital input word (D) has a digital value spanning from zero scale (ZS) to full scale (FS);
programming a number of digital interpolation clock steps (n) in the digital state machine;
initializing each digital word of a digital array O in the digital memory apparatus to zero, wherein the digital array O is comprised of a plurality of digital words Oj;
initializing each digital word of a digital array P in the digital memory apparatus to zero, wherein the digital array P is comprised of a plurality of digital words Pj;
programming the digital state machine beginning at a first digital interpolation clock step of j=1 and ending at a last digital interpolation clock step of j=n, generating and storing a sequence of digital words into the plurality of digital words Oj in the digital memory apparatus by, at each clock step j:
generating, in the digital divider apparatus, an array of offset values equal to full scale (FS) divided by 2j to generate a digital offset equal to
F
⁢
S
2
j
,
accumulating, in the digital accumulator apparatus, a sum of the plurality of digital words Pj and multiplying the sum of the plurality of digital words Pj by 2 to generate a digital accumulated sum equal to Σk=1j-1 2Pj;
adding, in the digital adder apparatus, the digital offset to the digital accumulated sum to generate a sum offset equal to
{
∑
k
=
1
j
-
1
⁢
2
⁢
P
j
+
F
⁢
S
2
j
}
;
subtracting, in the digital subtractor apparatus, the sum offset from the digital input word (D) to generate an output signal equal to
O
j
=
D
-
{
∑
k
=
1
j
-
1
⁢
2
⁢
P
j
+
F
⁢
S
2
j
}
;
and
generating, in the digital maximum selector apparatus, and storing a selected digital sequence into the plurality of digital words Pj, by respectively selecting a maximum from the plurality of digital words Oj and zero scale (ZS).

61 more elements

A method of generating an approximate square (aSQR) of input signals in an approximate squarer integrated circuit (IC), the approximate squarer integrated circuit (IC) comprising at least one of a subtracting apparatus, a summation apparatus, a maximum-selecting apparatus, and a scaling apparatus, the method comprising:

generating, during a first clock step (J1), a first offsetted output signal (O1) by:
generating, in the scaling apparatus, a one-half full scale signal (HS) by scaling a full scale signal (FS) by a factor of one-half;
subtracting, in the subtracting apparatus, the one-half full scale signal (HS) from an input signal (D), wherein the input signal (D) has a maximum range equal to the full-scale signal (FS);
selecting, during the first clock step (J1), in the maximum-selecting apparatus, a first positive P signal (P1) that is a maximum of the first offsetted output signal (O1) and a zero-scale signal (ZS), wherein the input signal (D) has a minimum range equal to the zero-scale signal (ZS);
generating, during a second clock step (J2), a second offsetted output signal (O2) by:
generating, in the scaling apparatus, a one-fourth full scale signal (QS) by scaling the full scale signal (FS) by a factor of one-fourth;
generating, in the scaling apparatus, a double first positive P signal (P1) by scaling the first positive P signal (P1) by a factor of two;
subtracting, in the subtracting apparatus, the one-fourth full scale signal (QS) and the double first positive P signal (P1) from the input signal (D);
selecting, during the second clock step (J2), in the maximum-selecting apparatus, a second positive P signal (P2) that is the maximum of the second offsetted output signal (O2) and the zero-scale signal (ZS);
and
generating, during the second clock step (J2), a second approximate square signal (S2) by summing, in the summing apparatus, the double first positive P signal (P1), and the second positive P signal (P2).

8 more elements

Dependencies are inferred from the claim text.