On Thursday, 22 August 2013 at 10:34:58 UTC, John Colvin wrote:
On Thursday, 22 August 2013 at 02:06:13 UTC, Tyler Jameson
Little wrote:
- array operations (int[] a; int[]b; auto c = a * b;)
- I don't think these are automagically SIMD'd, but there's
always hope =D
That isn't allowed. The memory for c must be pre-allocated, and
the expression then becomes c[] = a[] * b[];
Oops, that was what I meant.
Is it SIMD'd?
It depends. There is a whole load of hand-written assembler for
simple-ish expressions on builtin types, on x86. x86_64 is only
supported with 32bit integer types because I haven't finished
writing the rest yet...
However, I'm not inclined to do so at the moment as we need a
complete overhaul of that system anyway as it's currently a
monster*. It needs to be re-implemented as a template
instantiated by the compiler, using core.simd. Unfortunately
it's not a priority for anyone right now AFAIK.
That's fine. I was under the impression that it didn't SIMD at
all, and that SIMD only works if explicitly stated.
I assume this is something that can be done at runtime:
int[] a = [1, 2, 3];
int[] b = [2, 2, 2];
auto c = a[] * b[]; // dynamically allocates on the stack;
computes w/SIMD
writeln(c); // prints [2, 4, 6]
I haven't yet needed this, but it would be really nice... btw, it
seems D does not have dynamic allocation. I know C99 does, so I
know this is technically possible. Is this something we could
get? If so, I'll start a thread about it.
*
hand-written asm loops. If fully fleshed out there would be:
((aligned + unaligned + legacy mmx) * (x86 + x64) + fallback
loop)
* number of supported expressions * number of different types
of them. Then there's unrolling considerations. See
druntime/src/rt/arrayInt.d