On Apr 1, 2013, at 11:01 AM, Jed Brown <[email protected]> wrote:
Many uses of PetscStackCall() are O(1) calls per iteration or communication, thus I think acceptable in debug mode. But some of them are called O(n) times, such as in gamg/agg.c:
/* QR */ ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr); PetscStackCall("LAPACKgeqrf",LAPACKgeqrf_(&Mdata, &N, qqc, &LDA, TAU, WORK, &LWORK, &INFO)); ierr = PetscFPTrapPop();CHKERRQ(ierr);
PetscStackCall("LAPACKgeqrf",LAPACKgeqrf_(&Mdata, &N, qqc, &LDA, TAU, WORK, &LWORK, &INFO));
and in blockinvert.h. Presumably we need two versions of PetscCallStack, one that is used to call "safe" functions or those that are only called O(1) times, and one that is used to call "unsafe" or performance-sensitive functions.
Is the only issue that PetscStackCall() has a CHKMEMQ() within it? Otherwise it seems PetscStackCall() costs no more than any other PETSc function call with PetscFunctionBegin/Return(). Perhaps it could take a flag argument indicating if the memory check should be performed? For BLAS/LAPACK assuming we've allocated enough work space the memory check is likely not needed, should we have a PetscStackCallBLASLAPACK() (that possibly even handles the info flag when it exists?) which does not do a CHKMEMQ? Barry