When two names are aliased to the same data area, do they share the same header block, or do they have different header blocks that point to the common data area?
If they have different header blocks, another solution might be an easier quick fix. If #A is composite, say equal to a*b, the computation could be written as +/ A +/@:*&((a,b)&$) B This is no better in the current system, but if just the data, not the pointer, is shared, it would be possible to change $ so that x $ y avoids copying the data when (*/$x) -: #y . Henry rich > -----Original Message----- > From: [EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED] On Behalf Of Roger Hui > Sent: Wednesday, May 10, 2006 11:53 AM > To: Programming forum > Subject: Re: [Jprogramming] Multiply&sum for two big arrays > > In example 1, the expression A * B takes different space than > C1=: A1 * B1 for two reasons: > - the first expression is shorter than the second > - the first expression does not do an assignment > Mapping or not mapping is not material. > > A f B creates a temp in memory (not mapped) even if A > and B are mapped, even if it is then immediately assigned > to a mapped name. > > What would improve the situation is if +/@:* (and in general > f/@:g) is supported by special code. This may be done for a > future release. It is probably too late for J601. > > > > ----- Original Message ----- > From: "Oleg Kobchenko" <[EMAIL PROTECTED]> > To: "Programming forum" <[email protected]> > Sent: Sunday, May 07, 2006 1:11 AM > Subject: Re: [Jprogramming] Multiply&sum for two big arrays > > Yes this is really strange: I thought operations > with MMF nouns should not take up physical memory. > Especially if the result is mapped, moreover > if there are no intermediate operations. > > But even simple operations, like C=: A*B all mapped > take even more memory than not mapped. > (Using left arg "JFL map_jmf_ y" does not help.) > > Example 1. > require 'jmf files dir' > ]P=: jpath'~temp' > c:\math\j601\temp > > A=: i.20 20 > B=: %>:i.-20 20 > > createjmf_jmf_ (P,'/A.jmf');*/20 20 8 > createjmf_jmf_ (P,'/B.jmf');*/20 20 8 > createjmf_jmf_ (P,'/C.jmf');*/20 20 8 > map_jmf_ 'A1';P,'/A.jmf' > map_jmf_ 'B1';P,'/B.jmf' > map_jmf_ 'C1';P,'/C.jmf' > A1=: i.20 20 > B1=: %>:i.-20 20 > > 7!:2 'A * B' > 4800 > 7!:2 'C1=: A1 * B1' > 4992 > > > Example 2. > (Choose total sizes exceeding your physical + swap file, > swap file should be fixed). > > createjmf_jmf_ (P,'/a.jmf'); 40e6 * 8 > createjmf_jmf_ (P,'/b.jmf'); 40e6 * 8 > createjmf_jmf_ (P,'/c.jmf'); 40e6 * 8 > > map_jmf_ 'A';P,'/a.jmf' > map_jmf_ 'B';P,'/b.jmf' > map_jmf_ 'C';P,'/c.jmf' > showmap_jmf_'' > +-------+-----------------------+--+---+---+----------+------- > ---+--+---------+----+ > |name |fn |sn|fh |mh |address |header > |ts|msize |refs| > +-------+-----------------------+--+---+---+----------+------- > ---+--+---------+----+ > |A_base_|c:\math\j601\temp/a.jmf| |224|228|539820032 > |539820032 | |320000000|2 | > +-------+-----------------------+--+---+---+----------+------- > ---+--+---------+----+ > |B_base_|c:\math\j601\temp/b.jmf| |232|236|859832320 > |859832320 | |320000000|2 | > +-------+-----------------------+--+---+---+----------+------- > ---+--+---------+----+ > |C_base_|c:\math\j601\temp/c.jmf| > |240|244|1179844608|1179844608| |320000000|2 | > +-------+-----------------------+--+---+---+----------+------- > ---+--+---------+----+ > > > A=: i.40e6 > B=: i.40e6 > B=: >:B > B=: %B > |out of memory > | B=: %B > > unmapall_jmf_'' NB. OK let's restart > 0 0 0 > map_jmf_ 'B';P,'/b.jmf' > B=: %B > unmapall_jmf_'' > 0 > map_jmf_ 'C';P,'/c.jmf' > C=: 40e6#0.1 NB. initialize to form of results > unmapall_jmf_'' > 0 0 0 > map_jmf_ 'A';P,'/a.jmf' > map_jmf_ 'B';P,'/b.jmf' > map_jmf_ 'C';P,'/c.jmf' > C=: A*B > |out of memory > | C=:A *B > dir P > 1.ijs 417 23-Apr-06 16:25:24 > 1.ijx 4509 07-May-06 03:38:21 > 2.ijs 1050 27-Apr-06 06:39:33 > a.jmf 320000284 07-May-06 03:40:21 > b.jmf 320000284 07-May-06 03:40:24 > c.jmf 320000284 07-May-06 03:40:27 > temp.txt 111 22-Apr-06 08:19:34 > > > > --- Anssi Seppälä <[EMAIL PROTECTED]> wrote: > > > I have two big 2-dimensional arrays in mapped files A and B. I need > > to calculate +/ A * B efficiently and without "out of memory". > > > > In the case A was a vector the . (dot) can make it +/ . * , but this > > does not work if A is a matrix. Is there a way to do this without a > > program loop? > > > ---------------------------------------------------------------------- > For information about J forums see > http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
