| bool batchlas::verify::eigenvalues | ( | int | n, |
| std::vector< D > & | a, | ||
| std::vector< double > & | w ) |
Ascending eigenvalues of the n x n Hermitian a (column-major, ld = n, both triangles valid; destroyed).
Definition at line 64 of file reference.hh.
| bool batchlas::verify::singular_values | ( | int | m, |
| int | n, | ||
| std::vector< D > & | a, | ||
| std::vector< double > & | s ) |
Descending singular values of the m x n a (column-major, ld = m; destroyed).
Definition at line 83 of file reference.hh.
| bool batchlas::verify::getrf_pivots | ( | int | m, |
| int | n, | ||
| std::vector< E > & | a, | ||
| std::vector< std::int32_t > & | ipiv ) |
1-based pivots of the m x n a (column-major, ld = m; overwritten by its LU factor): min(m, n) of them, from ?getrf in E's own precision (float, double, complex<float>, complex<double>).
a holds the data in its native precision (MatrixView::astype<E>(items), never a promoted copy, which sends float data to dgetrf). The view overload below does this for you. Definition at line 107 of file reference.hh.
| bool batchlas::verify::getrf_pivots | ( | const MatrixView< T, MatrixFormat::Dense > & | A, |
| int | item, | ||
| std::vector< std::int32_t > & | ipiv ) |
getrf_pivots of item item of a dense view, copied in the view's own precision.
Definition at line 136 of file reference.hh.
| bool batchlas::verify::geqrf_tau | ( | int | m, |
| int | n, | ||
| std::vector< D > & | a, | ||
| std::vector< D > & | tau ) |
LAPACK geqrf of the m x n a (column-major, ld = m; overwritten by R and the reflectors) and its min(m, n) tau.
Definition at line 143 of file reference.hh.
|
inline |
Eigenvalues of the symmetric tridiagonal (diagonal d, off-diagonal e): ascending, written to d.
Definition at line 162 of file reference.hh.
| bool batchlas::verify::eigenvalues | ( | const MatrixView< T, MatrixFormat::Dense > & | A, |
| int | item, | ||
| std::vector< double > & | w ) |
eigenvalues() of item item of the Hermitian A (both triangles valid), promoted.
Definition at line 174 of file reference.hh.
| bool batchlas::verify::singular_values | ( | const MatrixView< T, MatrixFormat::Dense > & | A, |
| int | item, | ||
| std::vector< double > & | s ) |
singular_values() of item item of A, promoted.
Definition at line 180 of file reference.hh.
| bool batchlas::verify::geqrf_tau | ( | const MatrixView< T, MatrixFormat::Dense > & | A, |
| int | item, | ||
| std::vector< typename MatrixView< T, MatrixFormat::Dense >::promoted_type > & | tau ) |
geqrf_tau() of item item of A, promoted; the factor is discarded.
Definition at line 187 of file reference.hh.
|
constexpr |
Unit roundoff of T's real type: 2^-24 for float, 2^-53 for double.
Definition at line 20 of file tolerance.hh.
|
inline |
max(a, b), NaN if either is NaN (std::max drops a NaN second argument, so a poisoned value would pass).
Definition at line 23 of file tolerance.hh.
| double batchlas::verify::bound | ( | Check | kind, |
| int | n ) |
c(kind) * f(n) * eps<T>(): f(n) = n (below 1 counts as 1), except blas, f(k) = k + 2 (k below 0 counts as 0).
Definition at line 117 of file tolerance.hh.
slack.factor * bound<T>(kind, n).
| batchlas::invalid_argument | for an empty reason or a factor <= 0. |
Definition at line 123 of file tolerance.hh.
| bool batchlas::verify::within | ( | Check | kind, |
| int | n, | ||
| double | value ) |
pass without recording: for predicates evaluated on results expected to be wrong.
Definition at line 129 of file tolerance.hh.
Definition at line 131 of file tolerance.hh.
| bool batchlas::verify::pass | ( | Check | kind, |
| int | n, | ||
| double | value ) |
True when value is not NaN and within bound<T>(kind, n).
With BATCHLAS_VERIFY_RECORD=<path> set, appends "kind dtype n value bound 1 -" to that file (calibration input).
Definition at line 137 of file tolerance.hh.
pass against the slacked bound; records the raw value, the kind's bound, the factor and the reason.
Definition at line 143 of file tolerance.hh.
| double batchlas::verify::pivot_ratio_bound | ( | ) |
Bound of pivot_ratio: 1 + 64 eps<T>() (32 machine epsilons), the slack of a cabs1 comparison in T.
Definition at line 150 of file tolerance.hh.
|
strong |
orthogonality: Q from Householder reflectors or a direct orthogonalization (geqrf, orgqr, form_q, ortho); orthogonality_rotations: eigen- and singular vectors from an iterative solver (Jacobi, QR iteration, divide and conquer, inverse iteration).
n per kind, as each was calibrated: orthogonality, the rows of Q (Householder error grows with the reflector length); orthogonality_rotations, the number of vectors checked (columns of Q). blas: n is the inner dimension k (the tuner's componentwise trsm check: k = the order); trsm_residual is normwise and judged under solve.
| Enumerator | |
|---|---|
| factorization | |
| solve | |
| blas | |
| orthogonality | |
| orthogonality_rotations | |
| eigen_residual | |
| values | |
Definition at line 34 of file tolerance.hh.
|
inlineconstexpr |
True when this build has host LAPACKE references; a reference that returns false here failed.
Definition at line 37 of file reference.hh.
Namespaces | |
| namespace | detail |
Classes | |
| struct | Slack |
| A per-site factor on a kind's bound, in either direction. More... | |
1.18.0