home *** CD-ROM | disk | FTP | other *** search
-
-
-
- - 1 -
-
-
-
- 4. _K_n_o_w_n__I_s_s_u_e_s__a_n_d__W_o_r_k_a_r_o_u_n_d_s__f_o_r__I_R_I_X__s_y_s_t_e_m_s
-
- This chapter describes the known problems and workarounds
- for SCSL on IRIX/Mips systems.
-
-
- +o In addition to the passed in work array, the FFT
- routines also dynamically allocate scratch space
- from the stack. The amount of space allocated can
- be slightly bigger than the size of the largest
- processor cache. For single processor runs, the
- default stack size is large enough that these
- allocations generally cause no problems. But for
- parallel runs, you need to ensure that the stack
- size of slave threads is big enough to hold this
- scratch space. Failure to reserve sufficient stack
- space will cause programs to dump core due to
- stack overflows. The stack size of MP library
- slave threads is controlled via the
- MP_SLAVE_STACKSIZE environment variable or the
- mp_set_slave_stacksize() library routine. See the
- mp(3C), mp(3F) and pe_environ(5) reference pages
- for more information on controlling the slave
- stack size. For pthreads applications, the
- thread's stack size is specified as one of many
- creation attributes provided in the pthread_attr_t
- argument to pthread_create(3P). The stacksize
- attribute should be set explicitly to a non-
- default value using the
- pthread_attr_setstacksize(3P) call, described in
- the pthread_attr_init(3P) man page.
-
- +o A bug in libmp can cause a segmentation fault if
- the stacksize on the machine is unlimited. To
- avoid this use the limit command to set the
- stacksize to a finite number (e.g. limit stacksize
- 1024).
-
- +o The FFT routines that initialize the table of
- factors and trigonometric weights require more
- workspace than documented when the isys parameter
- is set to 0. This will cause a small memory leak,
- and this can become significant when the array is
- initialized repeatedly. A set of routines are
- available to the user to free this space. Please
- see the FFT_FREE(3S) man page for details.
-
- +o The LAPACK manpages for SSYEVD, DSYEVD, CHEEVD and
- ZHEEVD in SCSL 1.2 stated
-
- If JOBZ = 'V' and N > 1, LIWORK must be at
-
-
-
-
-
-
-
-
-
-
-
- - 2 -
-
-
-
- least 2 + 5*N.
-
- This requirement is not sufficient and has been
- corrected in SCSL 1.3 and newer, it must say
-
- If JOBZ = 'V' and N > 1, LIWORK must be at
- least 3 + 5*N.
-
-
- +o LAPACK in SCSL 1.3 and newer has an updated
- version of DHSEQR. In this version the 13th
- argument LWORK is now required vs the version in
- SCSL 1.2. Please see the DHSEQR(3S) man page for
- details.
-
-
- +o In the standard LAPACK test suite the following
- tests failed to pass the threshold:
-
- csep.out: CST drivers: 1 out of 11664
- csep.out: CST drivers: 1 out of 11664
- csvd.out: CBD: 1 out of 4085
- ctest.out: CPB: 12 out of 3458
- ctest.out: CPB drivers: 4 out of 4750
- ctest.out: CLS drivers: 1 out of 65268
- dgd.out: DXV drivers: 200 out of 5000
- dsvd.out: DBD: 1 out of 5510
- sgd.out: SXV drivers: 37 out of 5000
- ssep.out: SST drivers: 1 out of 14256
- ssep.out: SST drivers: 1 out of 14256
- stest.out: SPB: 1 out of 3458
- stest.out: SLS drivers: 3 out of 65268
-
-
- +o The following LAPACK routines were found by the
- standard LAPACK test suite to have abnormally high
- roundoff error:
-
- cgebak dgebak sgebak zgebak
- cggbak dggbak sggbak zggbak
- cggbal dggbal sggbal zggbal
-
-
-
- Feedback, comments and problems about SCSL are
- welcome at scsl-bugs@postofc.corp.sgi.com.
-
-
-
-
-
-
-
-
-
-
-
-
-
-