direct: diff from SciPy #615
Conversation
|
if there is interest in proceeding here, perhaps we could merge all of the trivial formatting changes, then take a closer look at how to deal with the reduced diff |
| #define MAXMEMORY 1073741824 | ||
| integer MAXFUNC = *maxf <= 0 ? 101000 : (*maxf + 1000 + *maxf / 2); | ||
| integer fixed_memory_dim = ((*n) * (sizeof(doublereal) + sizeof(integer)) + | ||
| (sizeof(doublereal) * 2 + sizeof(integer))); | ||
| MAXFUNC = MAXFUNC * fixed_memory_dim > MAXMEMORY ? MAXMEMORY/fixed_memory_dim : MAXFUNC; | ||
| c__ = (doublereal *) malloc(sizeof(doublereal) * (MAXFUNC * (*n))); | ||
| if (!(c__)) { | ||
| *ierror = -100; | ||
| goto cleanup; | ||
| } |
There was a problem hiding this comment.
This part was added as the algorithm uses absurd amounts of memory for large number of function evaluations. The MY_ALLOC macro did not help, we got a segmentation fault. Could this be of interest to nlopt @stevengj ?
|
A lot of these change change functions from |
Is there some way to keep the C library returning void, and add a wrapper layer to PyObject on top that could live in SciPy? |
|
CC @czgdp1807 as he knows most about the wrapping magic. A word of caution though: getting DIRECT into scipy was a substantial amount of work and we might just let it rest in its current state as I do not see open issues related to it (might also just be little adoption of it compared to SHGO and friends on the scipy side though). |
This draft PR shows the diff resulting from replacing the files of the direct algorithm here with what we have in SciPy. It looks like there is a mixture of (many) small formatting things, and (some) substantive changes.
Would there be interest in merging these sources into 'one true source'? This seems desirable, to de-duplicate work, ensuring any fixes made here end up in SciPy, and any contributions to SciPy end up here. We could set up a more formal vendoring situation to facilitate this once the sources are matched.
It may be that some of the changes here seem problematic, and we would rather change things in SciPy to match what is already here. That is also on the table, let me know what you think! In particular, it looks like we would want to split the
PyObjectparts out into separate bindings, either here on in SciPy.Cross reference scipy/scipy#21232