By: S. Rao (sonny.delete@this.burdell.org), June 21, 2008 4:50 pm
Room: Moderated Discussions
Linus Torvalds (torvalds@linux-foundation.org) on 6/21/08 wrote:
---------------------------
>S. Rao (sonny@burdell.org) on 6/21/08 wrote:
>>
>>(Maybe I'm way off here, but here goes)
>>
>>Remember that thing called register renaming ?
>>
>>Why does the architected register File have anything to
>>do with it?
>
>You want to rename into a larger set of physical
>registers, not into a smaller one. The latter is possible
>but very very inconvenient (it is what you effectively
>have to do if you are emulating a large register file
>architecture on a CPU with a smaller register file: you
>map the large register file into cache or some other
>"secondary" register file, and then rename it all into the
>smaller actual register file).
>
>So even with renaming, you do care about the size of the
>architected register file. A smaller architected register
>file allows you more flexibility in choosing the size of
>the actual hardware register file.
I think the implementation will make a difference here,
if you keep your architected state in the physical reg
file then yeah more threads == fewer renames. I think
IBM does it this way but not sure if Intel does it
this way necessarily. But then I would say you have
to weight it againt the penalty you're paying for having
fewer regs in the first place.... it's not clear that it
isn't a wash when it comes to complexity/timing
---------------------------
>S. Rao (sonny@burdell.org) on 6/21/08 wrote:
>>
>>(Maybe I'm way off here, but here goes)
>>
>>Remember that thing called register renaming ?
>>
>>Why does the architected register File have anything to
>>do with it?
>
>You want to rename into a larger set of physical
>registers, not into a smaller one. The latter is possible
>but very very inconvenient (it is what you effectively
>have to do if you are emulating a large register file
>architecture on a CPU with a smaller register file: you
>map the large register file into cache or some other
>"secondary" register file, and then rename it all into the
>smaller actual register file).
>
>So even with renaming, you do care about the size of the
>architected register file. A smaller architected register
>file allows you more flexibility in choosing the size of
>the actual hardware register file.
I think the implementation will make a difference here,
if you keep your architected state in the physical reg
file then yeah more threads == fewer renames. I think
IBM does it this way but not sure if Intel does it
this way necessarily. But then I would say you have
to weight it againt the penalty you're paying for having
fewer regs in the first place.... it's not clear that it
isn't a wash when it comes to complexity/timing
Topic | Posted By | Date |
---|---|---|
Intel AVX kills AMD SSE5 | Agner | 2008/06/17 08:14 AM |
Intel AVX kills AMD SSE5 | a reader | 2008/06/17 09:03 AM |
Bulldozer? | David Kanter | 2008/06/19 04:23 PM |
Bulldozer? | EduardoS | 2008/06/19 06:11 PM |
Bulldozer? | Max | 2008/06/19 06:16 PM |
Bulldozer? | Goose | 2008/06/21 02:23 AM |
Bulldozer? | David Kanter | 2008/06/21 07:37 AM |
Bulldozer? | someone | 2008/06/21 07:55 AM |
Bulldozer? | David Kanter | 2008/06/21 08:07 AM |
Bulldozer? | S. Rao | 2008/06/21 11:08 AM |
Regfiles | Peter | 2008/06/21 11:49 AM |
Bulldozer? | Linus Torvalds | 2008/06/21 12:23 PM |
Bulldozer? | S. Rao | 2008/06/21 04:50 PM |
unified physical register file nullifies x86 advantage | Michael S | 2008/06/22 12:24 AM |
unified physical register file nullifies x86 advantage | David Kanter | 2008/06/22 09:35 AM |
unified physical register file nullifies x86 advantage | hobold | 2008/06/22 01:03 PM |
Reg file vs. forwarding network | David Kanter | 2008/06/22 10:36 AM |
Reg file vs. forwarding network | hobold | 2008/06/22 12:39 PM |
Reg file vs. forwarding network | Peter | 2008/06/22 02:48 PM |
Reg file vs. forwarding network | David Kanter | 2008/06/22 08:54 PM |
Reg file vs. forwarding network | Peter | 2008/06/23 03:44 AM |
Reg file vs. forwarding network | savantu | 2008/06/23 04:41 AM |
Reg file vs. forwarding network | Peter | 2008/06/23 07:35 AM |
Reg file vs. forwarding network | Anders Jensen | 2008/06/23 11:05 AM |
Reg file vs. forwarding network | left nutz | 2008/06/27 07:31 AM |
Intel AVX kills AMD SSE5 | nobat | 2008/06/21 11:23 AM |
Intel AVX kills AMD SSE5 | Agner | 2008/06/21 10:01 PM |
So... | Dean Kent | 2008/06/22 07:35 AM |
SSE5 has a great chance to succeed. | mpx | 2008/06/22 12:25 AM |
SSE5 has a great chance to succeed. | Michael S | 2008/06/22 01:42 AM |
SSE5 has a great chance for fiasco | Agner | 2008/06/22 03:32 AM |
SSE5 has a great chance for fiasco | Ian Ameline | 2008/06/22 08:37 AM |
SSE5 has a great chance for fiasco | anonymous | 2008/06/22 09:02 AM |
SSE5 has a great chance for fiasco | hobold | 2008/06/22 12:59 PM |
SSE5 has a great chance for fiasco | Howard Chu | 2008/06/22 04:38 PM |
SSE5 has a great chance to succeed. | hobold | 2008/06/22 12:52 PM |
SSE5 has a great chance to succeed. | Michael S | 2008/06/22 01:46 PM |
SSE5 has a great chance to succeed. | Hannes | 2008/06/24 08:49 AM |
SSE5 has a great chance to succeed. | anonymous | 2008/06/24 10:46 AM |
SSE5 has a great chance to succeed. | Ian Ollmann | 2008/06/24 10:12 PM |