POWER6/7, compilers and more

By: Anil Maliyekkel (a.delete@this.a.edu), October 3, 2007 2:12 pm
Room: Moderated Discussions
Linus Torvalds (torvalds@osdl.org) on 10/3/07 wrote:
---------------------------
>Anil Maliyekkel (amaliy1@nospam.edu) on 10/3/07 wrote:
>>
>>They have already said that it was in order before.
>>Apparently the only OOO feature is that non-div/sqrt FP
>>instructions can execute in FP unit and finish ahead of
>>an executing div or sqrt.
>
>I thought they were also out-of-order in the sense that
>FP and integer are queued separately, as are memory ops.
>No?
>IOW, they are "mostly in-order", in the sense that most
>of the pipelines are in-order, but different pipes will
>be out-of-order wrt each other. So, for example, they
>have a separate pipeline for fdiv/sqrt, which is why the
>other FP ops can schedule independently of fdiv/sqrt.
>
From my understanding all dependencies have to be resolved at group dispatch. The reason why FP operations can go ahead of fdiv/sqrt is that IBM decides that stages not utilized during certain cycles during divide and sqrt operations could be used to continue executing FP operations. I have no idea if divide/sqrt instruction are implemented as a series of micro operations with other instructions interleaved, or whether subsequent instructions can just finish executing ahead of the divide/sqrt instruction.

>And I suspect the memory pipe is probably aggressively
>out-of-order at least after it misses in the L1, so I
>think there's some out-of-order even within the pipes,
>even if that is probably limited to just the memory side.

Prefetches are initiated speculatively in look ahead mode to bring data into the L1 after an initial L1 or ERAT (the 1st level TLB) miss. I'm guessing any subsequent L1 misses are ignored so as to allow look ahead mode to continue rapidly. I do wonder if look ahead execution back tracks to any load which missed the L1 if the data is retrieved before the original load that missed.
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
POWER6/7, compilers and moreDavid Kanter2007/10/02 09:36 PM
  POWER6/7, compilers and moreGabriele Svelto2007/10/02 11:32 PM
    POWER6/7, compilers and moreAnil Maliyekkel2007/10/03 08:23 AM
      POWER6/7, compilers and moreanon2007/10/03 08:28 AM
        POWER6/7, compilers and moreAnil Maliyekkel2007/10/03 02:22 PM
      Can you email me the PPT? (NT)David Kanter2007/10/03 08:56 AM
        Me too if you don't mind Anil (NT)Euronymous2007/10/03 12:33 PM
      POWER6/7, compilers and moreLinus Torvalds2007/10/03 09:34 AM
        POWER6/7, compilers and moreAnil Maliyekkel2007/10/03 02:12 PM
          POWER6/7, compilers and moreAnil Maliyekkel2007/10/03 04:06 PM
  POWER6/7, compilers and moreJesper frimann2007/10/03 12:18 AM
  POWER6/7, compilers and moreMr. Camel2007/10/03 02:33 AM
    POWER6/7, compilers and moreasdf2007/10/03 02:08 PM
  POWER6/7, compilers and moreanon2007/10/03 08:09 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?