1 //===- LinearTransform.cpp - MLIR LinearTransform Class -------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "mlir/Analysis/Presburger/LinearTransform.h" 10 #include "mlir/Analysis/Presburger/IntegerPolyhedron.h" 11 12 namespace mlir { 13 14 LinearTransform::LinearTransform(Matrix &&oMatrix) : matrix(oMatrix) {} 15 LinearTransform::LinearTransform(const Matrix &oMatrix) : matrix(oMatrix) {} 16 17 // Set M(row, targetCol) to its remainder on division by M(row, sourceCol) 18 // by subtracting from column targetCol an appropriate integer multiple of 19 // sourceCol. This brings M(row, targetCol) to the range [0, M(row, sourceCol)). 20 // Apply the same column operation to otherMatrix, with the same integer 21 // multiple. 22 static void modEntryColumnOperation(Matrix &m, unsigned row, unsigned sourceCol, 23 unsigned targetCol, Matrix &otherMatrix) { 24 assert(m(row, sourceCol) != 0 && "Cannot divide by zero!"); 25 assert((m(row, sourceCol) > 0 && m(row, targetCol) > 0) && 26 "Operands must be positive!"); 27 int64_t ratio = m(row, targetCol) / m(row, sourceCol); 28 m.addToColumn(sourceCol, targetCol, -ratio); 29 otherMatrix.addToColumn(sourceCol, targetCol, -ratio); 30 } 31 32 std::pair<unsigned, LinearTransform> 33 LinearTransform::makeTransformToColumnEchelon(Matrix m) { 34 // We start with an identity result matrix and perform operations on m 35 // until m is in column echelon form. We apply the same sequence of operations 36 // on resultMatrix to obtain a transform that takes m to column echelon 37 // form. 38 Matrix resultMatrix = Matrix::identity(m.getNumColumns()); 39 40 unsigned echelonCol = 0; 41 // Invariant: in all rows above row, all columns from echelonCol onwards 42 // are all zero elements. In an iteration, if the curent row has any non-zero 43 // elements echelonCol onwards, we bring one to echelonCol and use it to 44 // make all elements echelonCol + 1 onwards zero. 45 for (unsigned row = 0; row < m.getNumRows(); ++row) { 46 // Search row for a non-empty entry, starting at echelonCol. 47 unsigned nonZeroCol = echelonCol; 48 for (unsigned e = m.getNumColumns(); nonZeroCol < e; ++nonZeroCol) { 49 if (m(row, nonZeroCol) == 0) 50 continue; 51 break; 52 } 53 54 // Continue to the next row with the same echelonCol if this row is all 55 // zeros from echelonCol onwards. 56 if (nonZeroCol == m.getNumColumns()) 57 continue; 58 59 // Bring the non-zero column to echelonCol. This doesn't affect rows 60 // above since they are all zero at these columns. 61 if (nonZeroCol != echelonCol) { 62 m.swapColumns(nonZeroCol, echelonCol); 63 resultMatrix.swapColumns(nonZeroCol, echelonCol); 64 } 65 66 // Make m(row, echelonCol) non-negative. 67 if (m(row, echelonCol) < 0) { 68 m.negateColumn(echelonCol); 69 resultMatrix.negateColumn(echelonCol); 70 } 71 72 // Make all the entries in row after echelonCol zero. 73 for (unsigned i = echelonCol + 1, e = m.getNumColumns(); i < e; ++i) { 74 // We make m(row, i) non-negative, and then apply the Euclidean GCD 75 // algorithm to (row, i) and (row, echelonCol). At the end, one of them 76 // has value equal to the gcd of the two entries, and the other is zero. 77 78 if (m(row, i) < 0) { 79 m.negateColumn(i); 80 resultMatrix.negateColumn(i); 81 } 82 83 unsigned targetCol = i, sourceCol = echelonCol; 84 // At every step, we set m(row, targetCol) %= m(row, sourceCol), and 85 // swap the indices sourceCol and targetCol. (not the columns themselves) 86 // This modulo is implemented as a subtraction 87 // m(row, targetCol) -= quotient * m(row, sourceCol), 88 // where quotient = floor(m(row, targetCol) / m(row, sourceCol)), 89 // which brings m(row, targetCol) to the range [0, m(row, sourceCol)). 90 // 91 // We are only allowed column operations; we perform the above 92 // for every row, i.e., the above subtraction is done as a column 93 // operation. This does not affect any rows above us since they are 94 // guaranteed to be zero at these columns. 95 while (m(row, targetCol) != 0 && m(row, sourceCol) != 0) { 96 modEntryColumnOperation(m, row, sourceCol, targetCol, resultMatrix); 97 std::swap(targetCol, sourceCol); 98 } 99 100 // One of (row, echelonCol) and (row, i) is zero and the other is the gcd. 101 // Make it so that (row, echelonCol) holds the non-zero value. 102 if (m(row, echelonCol) == 0) { 103 m.swapColumns(i, echelonCol); 104 resultMatrix.swapColumns(i, echelonCol); 105 } 106 } 107 108 ++echelonCol; 109 } 110 111 return {echelonCol, LinearTransform(std::move(resultMatrix))}; 112 } 113 114 SmallVector<int64_t, 8> 115 LinearTransform::preMultiplyWithRow(ArrayRef<int64_t> rowVec) const { 116 assert(rowVec.size() == matrix.getNumRows() && 117 "row vector dimension should match transform output dimension"); 118 119 SmallVector<int64_t, 8> result(matrix.getNumColumns(), 0); 120 for (unsigned col = 0, e = matrix.getNumColumns(); col < e; ++col) 121 for (unsigned i = 0, e = matrix.getNumRows(); i < e; ++i) 122 result[col] += rowVec[i] * matrix(i, col); 123 return result; 124 } 125 126 SmallVector<int64_t, 8> 127 LinearTransform::postMultiplyWithColumn(ArrayRef<int64_t> colVec) const { 128 assert(matrix.getNumColumns() == colVec.size() && 129 "column vector dimension should match transform input dimension"); 130 131 SmallVector<int64_t, 8> result(matrix.getNumRows(), 0); 132 for (unsigned row = 0, e = matrix.getNumRows(); row < e; row++) 133 for (unsigned i = 0, e = matrix.getNumColumns(); i < e; i++) 134 result[row] += matrix(row, i) * colVec[i]; 135 return result; 136 } 137 138 IntegerPolyhedron 139 LinearTransform::applyTo(const IntegerPolyhedron &poly) const { 140 IntegerPolyhedron result(poly.getNumIds()); 141 142 for (unsigned i = 0, e = poly.getNumEqualities(); i < e; ++i) { 143 ArrayRef<int64_t> eq = poly.getEquality(i); 144 145 int64_t c = eq.back(); 146 147 SmallVector<int64_t, 8> newEq = preMultiplyWithRow(eq.drop_back()); 148 newEq.push_back(c); 149 result.addEquality(newEq); 150 } 151 152 for (unsigned i = 0, e = poly.getNumInequalities(); i < e; ++i) { 153 ArrayRef<int64_t> ineq = poly.getInequality(i); 154 155 int64_t c = ineq.back(); 156 157 SmallVector<int64_t, 8> newIneq = preMultiplyWithRow(ineq.drop_back()); 158 newIneq.push_back(c); 159 result.addInequality(newIneq); 160 } 161 162 return result; 163 } 164 165 } // namespace mlir 166