Boots – shoes – and leggings
Patent
1986-06-16
1989-01-17
Harkcom, Gary V.
Boots, shoes, and leggings
364754, G06F 752
Patent
active
047991828
DESCRIPTION:
BRIEF SUMMARY
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to a floating-point serial-pipelined multiplier. More particularly, the present invention allows both addition and multiplication to be performed in one array of cells by delaying the Y operand, whereas the exponent propagates through the delay unchanged in the addition mode. This allows re-ordering of the mantissa digits for the multiply mode.
2. Description of the Related Art
Architectures well-matched to the computationally intensive task of real-time SONAR signal processing have been proposed by Whitehouse and SPEISER (1981) based on the systolic techniques of KUNG (1979). These architectures are characterised by arrays of identical, limited capability systolic processing elements (SPE's) with minimum complexity nearest neighbour interconnections which allow a large number of parallel arithmetic operations to occur. However, direct handling of the large-order matrices typical of SONAR applications is difficult. Physically realisable systolic arrays are currently of limited size and partitioning of the matrix computation is mandatory. This partitioning diminishes the benefits of the systolic approach.
A design methodology which minimises the partitioning problem is the bit-serial approach proposed by LYON (1981). It is based on minimum area serial data paths and fixed-point serial processing elements and allows many more processing elements to be included in a systolic array. For SONAR, however, the use of fixed-point arithmetic imposes unacceptable constraints. In particular, the large dynamic ranges typical of the application can only be met indirectly by increasing the precision of the representation (i.e. the number of bits). This solution slows the achievable processing rate and incurs area overheads of O(n) where n is the number of bits in the representation. Further, number overflows must be prevented by periodic rescaling which increases quantum noise in intermediate values. The net effect is a degradation of processor throughput.
To optimise system performance a floating-point format is required. This permits both precision and dynamic range to be independently optimised, and it has been shown that it can be implemented with a methodology related to the work of LYON. Bit serial data paths and processing elements are used and an associated mode bit differentiates between mantissa and exponent.
Each multiply/accumulate SPE as defined by WHITEHOUSE and SPEISER is constructed from several registers and a multiplier and accumulator. The major element is the multiplier. Current work on floating-point multipliers for systolic processor applications such as NASH (1984) uses parallel algorithms to implement the multiplication. The problem of implementing minimum complexity units is addressed as an application of this invention.
A study of the work on serial fixed-point multipliers, originally proposed by JACKSON et al. (1968), reveals many references to their inherent capability to perform concurrently the operation X.times.Y+Z. The present invention demonstrates a use for the addition operator which is applicable to the direct handling of time division multiplexed exponents in a serial floating-point data format. A canonic multiplication cell is described which when used in a linear array implements a serial-pipelined floating-point multiplier. For an m-bit mantissa and e-bit exponent, the area of the multiplier is O(m), independent of the exponent length. Further, the exponent length is arbitrary and allows the dynamic range to be varied during task execution.
The implications of this low complexity VLSI implementation optimised for both precision and dynamic range for real-time signal processors with computation rates in excess of 1000 million floating-point operations per second are significant.
Serial-pipelined multipliers of the general type to which the present invention is directed have attracted continuing attention for digital signal processing applications and reference may be had to a report of JACKSON, L. B., KAIS
REFERENCES:
patent: 3725649 (1973-04-01), Deerfield
patent: 3871578 (1975-03-01), Van De Goor et al.
patent: 4013879 (1977-03-01), Bornmann et al.
patent: 4283770 (1981-08-01), Stewart
patent: 4493048 (1985-01-01), Kung et al.
patent: 4639857 (1987-01-01), McCanny et al.
The Radio and Electronic Engineer, Vol. 49, No. 11, issued Nov. 1979 (UK) Pekmestzi et al., "Cellular two's Complement Serial-Pipeline Multipliers", pp. 575-580.
Harkcom Gary V.
Mai Tan V.
The Commonwealth of Australia
LandOfFree
Cellular floating-point serial pipelined multiplier does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Cellular floating-point serial pipelined multiplier, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Cellular floating-point serial pipelined multiplier will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2415964