WARP11 Datasheet, PDF (7/19 Pages) STMicroelectronics – WEIGHT ASSOCIATIVE RULE PROCESSOR

W.A.R.P.1.1
provided by W.A.R.P. by means of A0-A9 and
EPA0-EPA2 output pins.
Data must be loaded 8 bit a time in the data bus
and can be read from an external non volatile
memory or loaded by an host processor.
ON-LINE MODE
In On-line mode W.A.R.P. is enabled to elaborate
input values and calculate outputs according to the
fuzzy rules stored into the microprogram. W.A.R.P.
reads the input values one a time in the input data
bus when all the inputs are given, a NP signal is
pulled high to indicate that the computation is start-
ing. The computational phase is divided in two main
parts. During the first one the input values are read
and the corresponding ALPHA values (activation
levels) are extracted from the internal memories. In
the second part the computation of the fuzzy rules
and the defuzzification are implemented.
The block diagram shown in figure 3 describes the
structure of W.A.R.P..
Antecedent Memory. It is formed by 4 benchs
each one containing one to four fuzzy sets bonded
to the input variables.
Consequent Memory. It is formed by one bench
where the fuzzy sets bondedto the output variables
are stored .
Program Memory. It is formed by a single bench.
Each line contains an operating code to execute
the computation of a rule. This code selects the
antecedentweights (ALPHA) involved in a rule, and
connects them by the programmed connective op-
erators (AND,OR).
Input Router. This internal block performs the
input data routing. Data are read one byte a time
from the input data bus, stored in 4 different buffers
and, thanks to a pipeline process, sent together to
4 indipendent modules to be processed in parallel
according to the chosen set-up configuration. Input
data resolution is decided by the user (MAX 128
points) according to the available configurations,
as shown in table 5.
The cycle starts when a positive pulse is applied at
FIN for a time no lower than an entire clock period
and continues until a new FIN (after NP low) or a
PRST signal is given.
Fuzzifier. This block generates the addresses of
the antecedentmemories where the ALPHA values
for each sampled input value are stored. It reads
the first four input values and calculates the corre-
sponding antecedent memories addresses. After-
wards it reads other four inputs values and
simultaneously sends, thanks to a pipeline proc-
ess, the previous four ALPHA values into internal
registers. These ALPHAvalues are then sent to the
Inference Unit. W.A.R.P. stores all ALPHA values
comprising a term set, which is formed by the MFs
connected to the IF-part of a rule, in successive
memory locations of the same memory word (see
figure 4). The vectors characterizing the MFs of a
term set are stored so that the ALPHAs of different
MFs corresponding to the same universe of dis-
course point (for the same input) are stored se-
quentially. So W.A.R.P. retrieves all the alpha
values of a term set using the crisp input value to
calculate the memory word address in the used
fuzzy memory device.The Fuzzifier Unit is driven
Figure 5. Antecedent Memory Organization
7/19