On Fri, Oct 8, 2010 at 11:47 AM, Andrew P. Mullhaupt <[email protected]> wrote: > On 10/7/2010 8:21 PM, Robert Kern wrote: >> On Thu, Oct 7, 2010 at 18:46, Andrew P. Mullhaupt<[email protected]> >> wrote: > >> It appears that your answer is YES, we COULD have that, BUT it would be >> a lot of work. >> You misread him. The answer is "no." We will not change the memory >> model so incompatibly. There is not only too much code in numpy to >> rewrite, but a ton of other code that we do not maintain that relies >> on our current memory model. > > Why would the class need to use the memory model of numpy? It wouldn't > be getting it's memory from numpy if you used memory mapped files.
It does not need to strictly speaking, but it is certainly required for any kind of acceptable performances. Are you ready to give up the ufunc machinery, for example ? More precisely, a big issue with having real/imaginary not contiguously in memory is that it is incompatible with the C99/C++ complex model (which *requires* that a complex double has the same binary representation as double[2]). As far as I know, it is impossible to fake this using different mmap areas for real and complex. cheers, David _______________________________________________ NumPy-Discussion mailing list [email protected] http://mail.scipy.org/mailman/listinfo/numpy-discussion
