OpenBLAS/relapack/test

• ReLAPACK Test Suite
• • Implementation

ReLAPACK Test Suite

This test suite compares ReLAPACK's recursive routines with LAPACK's compute
routines in terms of accuracy: For each test-case, we execute both ReLAPACK's
and LAPACK's routine on the same data and consider the numerical difference
between the two solutions.

This difference is computed as the maximum error across all elements of the
routine's outputs, where the error for each element is the minimum of the
absolute error and the relative error (with LAPACK as the reference). If the
error is below the error bound configured in config.h (default: 1e-5 for
single precision and 1e-14 for double precision) the test-case is considered as
passed.

For each routine the test-cases cover a variety of input argument combinations
to ensure that ReLAPACK's routines match the functionality of LAPACK for all use
cases.

The matrix size for all experiments (default: 100) can also be specified in
config.h.

Implementation

test.h provides the framework for our tests: It provides macros that allow to
generalize the tests for each operation in one file covering all data-types.
Such a file is structured as follows:

• All matrices required by the test-cases are declared globally. For each
  matrix, an array of two pointers is declared; one for the matrix copy passed
  to ReLAPACK and one passed to LAPACK.

• tests() contains the main control flow: it allocates (and later frees) the
  copies of the globally declared matrices. It then defines the macro
  ROUTINE to contain the name of the currently tested routine.
  It then uses the macro TEST to perform the test-cases.
  It receives the arguments of the routine, where matrices of which ReLAPACK
  and LAPACK receive a copy are index with i. (Example: TEST("L", &n, A[i], &n, info);)

• The macro TEST first calls pre(), which initializes all relevant
  matrices, then executes the ReLAPACK algorithm on the matrices with i = 0
  and then the LAPACK counter part with i = 1. It then calls post(),
  which computes the difference between the results, storing it in error.
  Finally, the error is printed out and compared to the error bound.

If all test-cases pass the error bound test, the program will have a 0 return
value, otherwise it is 1, indicating an error.