[Mlir-commits] [mlir] [MLIR][Presburger] Implement function to evaluate the number of terms in a generating function. (PR #78078)
Arjun P
llvmlistbot at llvm.org
Sun Jan 14 11:22:29 PST 2024
================
@@ -245,3 +245,236 @@ QuasiPolynomial mlir::presburger::detail::getCoefficientInRationalFunction(
}
return coefficients[power].simplify();
}
+
+static std::vector<Fraction> convolution(std::vector<Fraction> a,
+ std::vector<Fraction> b) {
+ // The length of the convolution is the maximum of the lengths
+ // of the two sequences. We pad the shorter one with zeroes.
+ unsigned convlen = std::max(a.size(), b.size());
+ for (unsigned k = a.size(); k < convlen; ++k)
+ a.push_back(0);
+ for (unsigned k = b.size(); k < convlen; ++k)
+ b.push_back(0);
+
+ std::vector<Fraction> convolution;
+ convolution.reserve(convlen);
+ convolution.clear();
+ for (unsigned k = 0; k < convlen; ++k) {
+ Fraction sum(0, 1);
+ for (unsigned l = 0; l <= k; ++l)
+ sum = sum + a[l] * b[k - l];
+ convolution.push_back(sum);
+ }
+ return convolution;
+}
+
+/// Substitute x_i = (s+1)^μ_i in one term of a generating function,
+/// returning
+/// a quasipolynomial which represents the exponent of the numerator
+/// of the result, and
+/// a vector which represents the exponents of the denominator of the
+/// result.
+std::pair<QuasiPolynomial, std::vector<Fraction>>
+substituteMuInTerm(unsigned numParams, ParamPoint v, std::vector<Point> ds,
+ Point mu) {
+ unsigned numDims = mu.size();
+ // First, the exponent in the numerator becomes
+ // - (μ • u_1) * (floor(first col of v))
+ // - (μ • u_2) * (floor(second col of v)) - ...
+ // - (μ • u_d) * (floor(d'th col of v))
+ // So we store the negation of the dot products.
+
+ // We have d terms, each of whose coefficient is the negative dot product,
+ SmallVector<Fraction> coefficients;
+ coefficients.reserve(numDims);
+ for (const Point &d : ds)
+ coefficients.push_back(-dotProduct(mu, d));
+
+ // Then, the affine function is a single floor expression, given by the
+ // corresponding column of v.
+ ParamPoint vTranspose = v.transpose();
+ std::vector<std::vector<SmallVector<Fraction>>> affine;
+ affine.reserve(numDims);
+ for (unsigned j = 0; j < numDims; ++j)
+ affine.push_back({SmallVector<Fraction>(vTranspose.getRow(j))});
+
+ QuasiPolynomial num(numParams, coefficients, affine);
+ num = num.simplify();
+
+ std::vector<Fraction> dens;
+ dens.reserve(ds.size());
+ // Similarly, each term in the denominator has exponent
+ // given by the dot product of μ with u_i.
+ for (const Point &d : ds)
+ dens.push_back(dotProduct(d, mu));
+ // This term in the denominator is
+ // (1 - (s+1)^dens.back())
+
+ return std::make_pair(num, dens);
+}
+
+void normalizeDenominatorExponents(int &sign, QuasiPolynomial &num,
+ std::vector<Fraction> &dens) {
+ // We track the number of exponents that are negative in the
+ // denominator, and convert them to their absolute values.
+ unsigned numNegExps = 0;
+ Fraction sumNegExps(0, 1);
+ for (unsigned j = 0, e = dens.size(); j < e; ++j) {
+ if (dens[j] < 0) {
+ numNegExps += 1;
+ sumNegExps += dens[j];
+ }
+ // All exponents will be made positive; then reduce
+ // (1 - (s+1)^x)
+ // to
+ // (-s)*(Σ_{x-1} (s+1)^j) because x > 0
+ dens[j] = abs(dens[j]) - 1;
+ }
+
+ // If we have (1 - (s+1)^-c) in the denominator,
+ // multiply and divide by (s+1)^c.
+ // We convert all negative-exponent terms at once; therefore
+ // we multiply and divide by (s+1)^sumNegExps.
+ // Then we get
+ // -(1 - (s+1)^c) in the denominator,
+ // increase the numerator by c, and
+ // flip the sign.
+ if (numNegExps % 2 == 1)
+ sign = -sign;
+ num = num - QuasiPolynomial(num.getNumInputs(), sumNegExps);
+
+ // Take all the (-s) out.
+ unsigned r = dens.size();
+ if (r % 2 == 1)
+ sign = -sign;
+}
+
+/// We have a generating function of the form
+/// f_p(x) = \sum_i sign_i * (x^n_i(p)) / (\prod_j (1 - x^d_{ij})
+///
+/// where sign_i is ±1,
+/// n_i \in Q^p -> Q^d is the sum of the vectors d_{ij}, weighted by the
+/// floors of d affine functions.
+/// d_{ij} \in Q^d are vectors.
+///
+/// We need to find the number of terms of the form x^t in the expansion of
+/// this function.
+/// However, direct substitution causes the denominator to become zero.
+///
+/// We therefore use the following procedure instead:
+/// 1. Substitute x_i = (s+1)^μ_i for some vector μ. This makes the generating
+/// function a function of a scalar s.
+/// 2. Write each term in this function as P(s)/Q(s), where P and Q are
+/// polynomials. P has coefficients as quasipolynomials in d parameters, while
+/// Q has coefficients as scalars.
+/// 3. Find the constant term in the expansion of each term P(s)/Q(s). This is
+/// equivalent to substituting s = 0.
+///
+/// Step (1) We need to find a μ_i such that we can substitute x_i =
+/// (s+1)^μ_i. After this substitution, the exponent of (s+1) in the
+/// denominator is (μ_i • d_{ij}) in each term. Clearly, this cannot become
+/// zero. Hence we find a vector μ that is not orthogonal to any of the
+/// d_{ij} and substitute x accordingly.
+///
+/// Step (2) We need to express the terms in the function as quotients of
+/// polynomials. Each term is now of the form
+/// sign_i * (s+1)^n'_i / (\prod_j (1 - (s+1)^d'_{ij}))
+/// For the i'th term, we first normalize the denominator to have only
+/// positive exponents. We convert all the d'_{ij} to their
+/// absolute values by multiplying and dividing by (s+1)^(-d'_{ij}) if it is
+/// negative. We change the sign accordingly to keep the denominator in the
+/// same form.
+/// Then, we replace each (1 - (s+1)^(d'_{ij})) with
+/// (-s)(\sum_{0 ≤ k < d'_{ij}} (s+1)^k).
+/// Thus the term overall has now the form
+/// sign'_i * (s+1)^n'_i / (s^r * \prod_j (\sum_k (s+1)^k)).
+/// This means that
+/// the numerator is a polynomial in s, with coefficients as
+/// quasipolynomials (given by binomial coefficients), and the denominator
+/// is polynomial in s, with fractional coefficients (given by taking the
+/// convolution over all j).
+///
+/// Step (3) We need to find the constant term in the expansion of each
+/// term. Since each term has s^r as a factor in the denominator, we avoid
+/// substituting s = 0 directly; instead, we find the coefficient of s^r in
+/// sign'_i * (s+1)^n'_i / (\prod_j (\sum_k (s+1)^k)),
+/// for which we use the `getCoefficientInRationalFunction()` function.
+///
+/// Verdoolaege, Sven, et al. "Counting integer points in parametric
+/// polytopes using Barvinok's rational functions." Algorithmica 48 (2007):
+/// 37-66.
+QuasiPolynomial
+mlir::presburger::detail::computeNumTerms(const GeneratingFunction &gf) {
+ std::vector<Point> allDenominators;
+ for (ArrayRef<Point> den : gf.getDenominators())
+ allDenominators.insert(allDenominators.end(), den.begin(), den.end());
+ Point mu = getNonOrthogonalVector(allDenominators);
+
+ unsigned numParams = gf.getNumParams();
+
+ const std::vector<std::vector<Point>> &ds = gf.getDenominators();
+ QuasiPolynomial totalTerm(numParams, 0);
+ for (unsigned i = 0, e = ds.size(); i < e; ++i) {
+ int sign = gf.getSigns()[i];
+
+ auto [numExp, dens] =
+ substituteMuInTerm(numParams, gf.getNumerators()[i], ds[i], mu);
+ // Now the numerator is (s+1)^numExp
+ // and the denominator is \prod_j (1 - (s+1)^dens[j]).
+
+ normalizeDenominatorExponents(sign, numExp, dens);
+
+ // Now the expression is
+ // (s+1)^num /
+ // (s^r * Π_(0 ≤ i < r) (Σ_{0 ≤ j ≤ dens[i]} (s+1)^j))
+
+ // Letting P(s) = (s+1)^num and Q(s) = Π_r (...),
+ // we need to find the coefficient of s^r in
+ // P(s)/Q(s).
+
+ unsigned r = dens.size();
+ // First, the coefficients of P(s), which are binomial coefficients.
+ // We need r+1 of these.
+ std::vector<QuasiPolynomial> numeratorCoefficients;
+ numeratorCoefficients.reserve(r + 1);
+ numeratorCoefficients.push_back(
+ QuasiPolynomial(numParams, 1)); // Coeff of s^0
+ for (unsigned j = 1; j <= r; ++j)
+ numeratorCoefficients.push_back(
+ (numeratorCoefficients[j - 1] *
+ (numExp - QuasiPolynomial(numParams, j - 1)) / Fraction(j, 1))
+ .simplify());
----------------
Superty wrote:
Write a comment saying this is the recursive formula for binomial coeffs
https://github.com/llvm/llvm-project/pull/78078
More information about the Mlir-commits
mailing list