Raul

indeed i can make up anything
using various b.
storing seems to be reliable
only as 8b characters
i think some 16b Unicode's
are illegal
so i cannot store and manipulate
16 bit floats that way
to do 12b floats
i might store pairs
in 3 bytes

i suppose one could
communicate with the many
DLL's that access the
BFloat16 hiding in every AVX machine
by sifting in and out of
that format
though bytes are rather inconvenient
due to the vagaries
of how sign, multiplier and multiplicand
bits are stored

one would prefer that that was all handled
and could check for positivity with
a simple >0
rather than make a DLL call
or another sift

~greg heil
https//picsrp.github.io

--

from: Raul Miller <[email protected]>
to: Programming forum <[email protected]>
date: Aug 2, 2021, 10:12 PM
subject: Re: [Jprogramming] Two implementation questions

https://www.jsoftware.com/help/dictionary/dbdotn.htm

   #~.(i.1024) 32 b. 1
64

>Or, on my machine, the bitwise operations give me 64 bits packed in a J atom.

>Constructing mechanisms which make an array of such things appear to be a bit 
>array of arbitrary dimensions should be straightforward for someone with your 
>background. (Though you could also use them "as is" with an implicit trailing 
>dimension of 64.)

Thanks,

--
Raul

--

from: greg heil <[email protected]>
to: Programming forum <[email protected]>
date: Aug 2, 2021, 8:33 PM
subject: Re: [Jprogramming] Two implementation questions

Henry

Hmm, how would one create an array of bits?
use packed 8-bit characters?
would one use a similar strategy
for 16-bit floats?
(which i am more interested in,
and is directly supported in most hardware
-for its utility in Deep Learning)

Seems a serious amount of
packing unpacking
indexing chopping and inverting
just for matrix multiplication
i could be wrong
but i do not see the algorithm

~greg heil
https//picsrp.github.io

--

from: Henry Rich <[email protected]>
to: [email protected]
date: Aug 2, 2021, 8:07 PM
subject: Re: [Jprogramming] Two implementation questions

>Rather than having a new datatype, you could create arrays of bits and use 
>bitwise booleans on them.

>Can you list a set of operations on bit arrays that would suffice to solve the 
>problems you were working on?

Henry Rich

--

from: greg heil <[email protected]>
to: Programming forum <[email protected]>
date: Aug 2, 2021, 7:07 PM
subject: Re: [Jprogramming] Two implementation questions



i did my undergraduate and doctoral thesis
on the categories of Graphs
with multiplication (and/or) as the operator
both were done using APL's bit-boolean's

much more than just multiplication
were done on these graphs
homomorphism, isomorphism
clustering cliquing
transitive/closures
all benefited from the simpler
bit-boolean's

as categories of graphs
run the full gamut
from 0-100% full
one cannot use
any form of linked list

a size multiplication by 8
would (at least in those days)
have caused a LOT of cache misses

even now caching would
be a lot happier
because the Social Networks
under investigation
(the object of the studies)
have grown a LOT bigger!-)

~greg heil
https//picsrp.github.io
.
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to