The class SmallFpLogImpl is a very low level implementation class for fast
arithmetic in a small, prime finite field. It is not intended for use
by casual CoCoALib users, who should instead see the documentation in
QuotientRing (in particular the function NewZZmod), or possibly the
documentation in RingFp, RingFpLog, and RingFpDouble.
Compared to SmallFpImpl the only difference is an implementation
detail: multiplication and division are achieved using discrete log
tables -- this may be fractionally faster on some processors.
Note that the cost of construction of a SmallFpLogImpl(p) object for
larger primes may be quite considerable (linear in p), and the resulting
object may occupy quite a lot of space (e.g. probably about 6*p bytes).
All operations on values must be effected by calling member functions
of the SmallFpLogImpl class. Here is a brief summary.
SmallFpLogImpl::IsGoodCtorArg(p); // true iff ctor SmallFpLogImpl(p) will succeed
SmallFpLogImpl::ourMaxModulus(); // largest permitted modulus
SmallFpLogImpl ModP(p, convention); // create SmallFpLogImpl object
long n;
BigInt N;
BigRat q;
SmallFpImpl::value_t a, b, c;
ModP.myModulus(); // value of p (as a long)
ModP.myReduce(n); // reduce mod p
ModP.myReduce(N); // reduce mod p
ModP.myReduce(q); // reduce mod p
ModP.myExport(a); // returns a preimage (of type long) according to symm/non-neg convention.
ModP.myNegate(a); // -a mod p
ModP.myAdd(a, b); // (a+b)%p;
ModP.mySub(a, b); // (a-b)%p;
ModP.myMul(a, b); // (a*b)%p;
ModP.myDiv(a, b); // (a*inv(b))%p; where inv(b) is inverse of b
ModP.myPower(a, n); // (a^n)%p; where ^ means "to the power of"
ModP.myIsZeroAddMul(a,b,c) // a = (a+b*c)%p; result is (a==0)
For myExport the choice between least non-negative and symmetric
residues is determined by the convention specified when constructing
the SmallFpLogImpl object. This convention may be either
GlobalSettings::SymmResidues or
GlobalSettings::NonNegResidues.
The only clever bit is the economical construction of the log/exp
tables in the constructor where we exploit the fact that myRoot to
the power (p-1)/2 must be equal to -1.
This implementation uses discrete log/exp tables to effect multiplication and division quickly. Note that the residues themselves (i.e. the values of the ring elements) are held as machine integers whose value is the least non-negative representative of the residue class (i.e. in the range 0 to p-1). In particular, although log tables are used, we do NOT use a logarithmic representation for the field elements.
The log/exp tables are stored in C++ vectors: aside from their
construction during the RingFpLogImpl constructor, these vectors are
never modified, and are used only for table look-up. The C++ vectors
are resized in the body of the constructor to avoid large memory
requests when overly large characteristics are supplied as argument.
Besides these tables SmallFpLogImpl also remembers the characteristic in
myModulus; myRoot is the primitive root used to generate the log/exp
tables.
The members myResidueUPBValue and myIterLimit and myHalfNormalize
may be used for delayed normalization in loops: see the inner product example
in SmallFpImpl.
As the code currently stands, the modulus must also be small enough that it
can fit into an FpTableElem (an unsigned short), and that its
square can fit into a value_t. Using shorts in the tables gave
slightly better run-time performance in our tests. Furthermore, to permit
use of unnormalized products in some algorithms, twice the square of the
characteristic must fit into a value_t (i.e. myIterLimit must
be greater than zero). The constructor for a RingFpLogImpl checks the
size restrictions on the characteristic.
Note that the log table has a slot with index 0 which is never written to nor read from. The exp table is double size so that multiplication can be achieved more easily: the highest slot which could ever be used is that with index 2p-3 (in division), but the constructor fills two extra slots (as this makes the code simpler/neater).
The only slick part of the implementation is the filling of the tables in the constructor, where some effort is made to avoid doing more reductions modulo p than necessary. Note that the primitive root is always calculated (potentially costly!); there is no memorized global table of primitive roots anywhere.
It is not as fast as I hoped -- perhaps cache effects?