1 |
On Wed, 9 Sep 2015, Stefan G. Weichinger wrote: |
2 |
|
3 |
> On 09.09.2015 20:01, Stefan G. Weichinger wrote: |
4 |
> > Digging up that thread now somewhere ... |
5 |
> |
6 |
> Ah, I even participated then ;-) |
7 |
> |
8 |
> AFAI understand gcc-5 should compile faster? |
9 |
> And generate faster code in some cases? |
10 |
|
11 |
No, and yes. Compilation is not affected in any way and runtime |
12 |
performance can only be improved _if_ this stuff is explicitly used within |
13 |
the code. |
14 |
|
15 |
Meaning you would feel a difference in no less then 5 years when gcc-6 is |
16 |
widely used and accelerator support is not restricted to intel MIC and |
17 |
nvidia gpus. James is getting a bit ahead of himself calling this a |
18 |
"game changer" - yeaaaaah... not really right now. |
19 |
|
20 |
Right now this functionality is a toy for the HPC community and will stay |
21 |
that way. To use it you have to build a separate offloading compiler, need |
22 |
custom code used by few, and expensive hardware. The tree ebuild doesn't |
23 |
even provide a way for enabling the accelerator support. |
24 |
|
25 |
|
26 |
> |
27 |
> ... does it use this new stuff anyway, do we need a specific USE-flag |
28 |
> enabled (I can't spot it, looking for something like "acc" or "rdma" |
29 |
> ;-)), do we need specific CFLAGS .. ? |
30 |
> |
31 |
> just curious. |
32 |
|
33 |
I can't speak for RDMA (can't find any mention of it in gcc) because |
34 |
that's an even more exotic thing than plain old accelerator support |
35 |
(unless you run infiniband at home...), but the flags are: |
36 |
|
37 |
-fopenmp |
38 |
-foffload |
39 |
-fopenacc |
40 |
|
41 |
However enabling them is as useful as having CFLAGS=-fopenmp currently. It |
42 |
changes __nothing__ unless an application has openmp annotations, and the |
43 |
ones that do should already provide a means of doing so in the build |
44 |
system. |
45 |
|
46 |
tldr: don't buy a dedicated gpu just because you read something on a |
47 |
mailing list ;) |