|
| 1 | + |
| 2 | +using Int = int; |
| 3 | + |
| 4 | +namespace hipo { |
| 5 | +Int denseFactTiny(Int n, Int k, double* A, double* B) { |
| 6 | + // =========================================================================== |
| 7 | + // Partial dense factorisation for supernodes with small size and front. |
| 8 | + // Assumes that both n and n-k are smaller than the block size. |
| 9 | + // Matrix A is in format FH |
| 10 | + // Matrix B is in format FH |
| 11 | + // BLAS calls: none |
| 12 | + // =========================================================================== |
| 13 | + |
| 14 | + if (n < 0 || k < 0 || !A || (k < n && !B)) return 1; |
| 15 | + if (n == 0) return 0; |
| 16 | + |
| 17 | + const Int ldb = n - k; |
| 18 | + |
| 19 | + // factorisation of top kxk block |
| 20 | + for (Int row = 0; row < k; ++row) { |
| 21 | + for (Int col = 0; col < row; ++col) { |
| 22 | + // off-diagonal |
| 23 | + for (Int l = 0; l < col; ++l) { |
| 24 | + A[col + row * k] -= A[l + row * k] * A[l + col * k] * A[l + l * k]; |
| 25 | + } |
| 26 | + double temp = A[col + row * k]; |
| 27 | + A[col + row * k] /= A[col + col * k]; |
| 28 | + |
| 29 | + // contribution to diagonal |
| 30 | + A[row + row * k] -= A[col + row * k] * temp; |
| 31 | + } |
| 32 | + } |
| 33 | + |
| 34 | + for (Int row = k; row < n; ++row) { |
| 35 | + // update rows below |
| 36 | + for (Int col = 0; col < k; ++col) { |
| 37 | + for (Int l = 0; l < col; ++l) { |
| 38 | + A[col + row * k] -= A[l + row * k] * A[l + col * k] * A[l + l * k]; |
| 39 | + } |
| 40 | + A[col + row * k] /= A[col + col * k]; |
| 41 | + } |
| 42 | + |
| 43 | + // update schur complement |
| 44 | + const Int rowB = row - k; |
| 45 | + for (Int col = k; col <= row; ++col) { |
| 46 | + const Int colB = col - k; |
| 47 | + for (Int l = 0; l < k; ++l) { |
| 48 | + B[colB + rowB * ldb] -= A[l + row * k] * A[l + col * k] * A[l + l * k]; |
| 49 | + } |
| 50 | + } |
| 51 | + } |
| 52 | + |
| 53 | + // store the reciprocal of the pivot |
| 54 | + for (Int row = k; row < n; ++row) { |
| 55 | + // A[row + k * row] = 1.0 / A[row + k * row]; |
| 56 | + } |
| 57 | +} |
| 58 | + |
| 59 | +} // namespace hipo |
0 commit comments