| 1 |
This program tests TSQR::DistTsqr, which implements the internode- |
| 2 |
parallel part of TSQR (TSQR::Tsqr). TSQR (Tall Skinny QR) computes |
| 3 |
the QR factorization of a tall and skinny matrix (with many more rows |
| 4 |
than columns) distributed in a block row layout across one or more |
| 5 |
processes. |
| 6 |
|
| 7 |
By default, TSQR::DistTsqr will only be tested with real arithmetic |
| 8 |
Scalar types (currently, float and double, corresponding to LAPACK's |
| 9 |
"S" resp. "D" data types). If you build Trilinos with complex |
| 10 |
arithmetic support (Teuchos_ENABLE_COMPLEX), and invoke this program |
| 11 |
with the "--complex" option, complex arithmetic Scalar types will also |
| 12 |
be tested (currently, std::complex<float> and std::complex<double>, |
| 13 |
corresponding to LAPACK's "C" resp. "Z" data types). |
| 14 |
|
| 15 |
This program tests DistTsqr on a test matrix $A$, consisting of a |
| 16 |
stack of square numCols by numCols upper triangular matrices, |
| 17 |
distributed with one on each MPI process in MPI_COMM_WORLD. It does |
| 18 |
the following: |
| 19 |
|
| 20 |
1. Compute the QR factorization of A, with the Q factor stored |
| 21 |
implicitly |
| 22 |
2. Compute the explicit form of the Q factor, by applying the |
| 23 |
implicitly stored Q factor to the first numCols columns of the |
| 24 |
identity matrix (distributed across processes in the same manner |
| 25 |
as A) |
| 26 |
|
| 27 |
That means this program is only testing applying $Q$, and not $Q^T$ |
| 28 |
(or $Q^H$, the conjugate transpose, in the complex-arithmetic case). |
| 29 |
This exercises the typical use case of TSQR in iterative methods, in |
| 30 |
which the explicit $Q$ factor is desired in order to compute a rank- |
| 31 |
revealing decomposition and possibly also replace the null space basis |
| 32 |
vectors with random data. |
| 33 |
|
| 34 |
This program can test accuracy ("--verify") or performance |
| 35 |
("--benchmark"). For accuracy tests, it computes both the |
| 36 |
orthogonality $\| I - Q^* Q \|_F$ and the residual $\| A - Q R \|_F$. |
| 37 |
For performance tests, it repeats the test with the same data for a |
| 38 |
number of trials (specified by the "--ntrials=<n>" command-line |
| 39 |
option). This ensures that the test is performed with warm cache, so |
| 40 |
that kernel performance rather than memory bandwidth is being |
| 41 |
measured. |