Brad,
Thanks for sharing this email. Not an approach I had considered. I will follow
up with a GitHub issue.
- Apan
On Jul 10, 2017, at 11:52 AM, Brad Chamberlain
<[email protected]<mailto:[email protected]>> wrote:
Hi Apan --
Completely unrelated to this thread with you, I received the following mail
from Nikhil Padmanabhan (community user/developer and Yale prof), who said it
was OK for me to reflect it on this thread (he's not on this mailing list,
suggesting that it might be useful/interesting to capture this discussion on a
GitHub issue). Editing his mail down a bit, he offered the following. Hope
this is helpful/of interest.
-Brad
Nikhil writes:
Here is an oldie that I recently came back to in working on the PAW code... how
does one map an SOA to AOS?
Now, zip gets you most of this, but we'd ideally not want to keep writing zip
and remembering structure names etc... But it seems like zip wrapped in an
iterator does the trick:
use BlockDist;
const Space = {0.. #10};
const D : domain(1) dmapped Block(boundingBox=Space) = Space;
var x,y,z: [D] int;
iter Coords() ref {
for tup in zip(x,y,z) {
yield tup;
}
}
iter Coords(param tag:iterKind) ref where tag==iterKind.standalone {
forall tup in zip(x,y,z) {
yield tup;
}
}
forall (x1,y1,z1) in Coords() {
if (x1.locale.id != here.id) then writeln("ERROR!");
if (y1.locale.id != here.id) then writeln("ERROR!");
if (z1.locale.id != here.id) then writeln("ERROR!");
x1 = 1 + here.id;
y1 = 2 + here.id;
z1 = 3 + here.id;
}
writeln(x);
writeln(y);
writeln(z);
On Fri, 7 Jul 2017, Brad Chamberlain wrote:
Hi Apan --
I'm happy that you're thinking about this challenge, as it's a longstanding one
that users bring up and one for which we've kicked ideas around, yet without
any ideal solution yet.
My question is, is there a way to determine which field is being requested in
an AoS reference (A[i].x) from within the domain map code?
Unfortunately not -- because the domain map code is only implementing the array
itself, it can only get the [i] access as an argument and know that the element
type is a(n apparent) record.
I'll mention that when this challenge was first levelled at us (by some
potential Chapel users from DOE labs), I proposed a domain map-based solution,
albeit possibly one different than the one you are pursuing. I'd imagined
having the domain map be parameterized by the number/type of elements in the
fields and having it pre-allocate the data for them. I recall that there was
something about this solution that they found off-putting, but can't quite
recall what it was -- it seems like it was related to an extra indirection or
multiplication, but I can't reconstruct the conversation just now.
>From there, my thoughts turned to approaches that would rely on the facts that
>Chapel types can (1) implement their own indexing functions and (2) support
>methods that omit parenthesis to get pseudo-fields. That is:
-----
record R {
proc this(i: real) {
return 2*i;
}
proc x {
return 3.14;
}
}
var myR: R;
writeln(R[2.4]);
writeln(R.x);
-----
The last time I was kicking this problem around, I was wondering whether
creating a record/class that wrapped an array, or group of arrays, could lean
on these mechanisms to provide AoS vs. SoA transparency. Offhand, I can't
recall whether I got stymied or never spent the time on it to see it through.
I'll also mention that I'm (personally) open to modifying the language to
"solve" the AoS vs. SoA challenge, though I'm hesitant to take the approach of
giving domain maps more context for the expression that follows the indexing
expression as you were asking about, at least on the surface (but perhaps could
be convinced, given a strong enough proposal and rationale).
-Brad
On Fri, 7 Jul 2017, Qasem, Apan M wrote:
Hello,
I am implementing data layout transformations in Chapel that favor
GPU/heterogenous memory hierarchy.
One of the transformations I am looking at is the AoS-to-SoA conversion
(critical for heterogenous applications). I am hoping to implement this using a
domain map. For example,
var N = 1..100;
record img {
var r: int;
var g: int;
var b: int;
var x: int;
}
var domAoS : domain(1) = {N};
var domSoA : domain(1) dmapped SoA() = {N};
var A : [domAoS] img;
var B : [domSoA] img;
A[3].g = 17; // A accessed as AoS
B[3].g = 17; // B accessed as SoA
I have a partial implementation where I am able to map an AoS index to an SoA
index in the SoA() domain map (as long as the field access information is hard
coded in the module code).
My question is, is there a way to determine which field is being requested in
an AoS reference (A[i].x) from within the domain map code?
- Apan
Apan Qasem, PhD
Visiting Scholar, AMD Research
Associate Professor, Dept of Computer Science
Texas State University
https://na01.safelinks.protection.outlook.com/?url=http:%2F%2Fwww.cs.txstate.edu%2F~aq10&data=01%7C01%7Capan%40txstate.edu%7C8f964b54ac394ddbaa6308d4c7b406af%7Cb19c134a14c94d4caf65c420f94c8cbb%7C0&sdata=5frsBV5jw8umfiGCX0s6ZJgIo7FewZl8OIaolOVbNDI%3D&reserved=0
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Chapel-developers mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/chapel-developers
