[prev in list] [next in list] [prev in thread] [next in thread] 

List:       pypy-dev
Subject:    Re: [pypy-dev] Vectorizing pypy traces no3
From:       Maciej Fijalkowski <fijall () gmail ! com>
Date:       2015-04-14 11:39:49
Message-ID: CAK5idxSs2aimLmp8wQs6-UZbXF6WL460kob7_XaL=gmutOGM3Q () mail ! gmail ! com
[Download RAW message or body]

Hi Richard.

I read it. but I don't quite understand, want to discuss on IRC?

On Tue, Apr 14, 2015 at 1:16 PM, Richard Plangger <rich@pasra.at> wrote:
> Hi,
>
> I have recently managed to correctly transform a trace to a vectorized
> trace that includes a guard. I'm hoping that this might be merged into
> the code base of pypy (when it is finished), thus it would be nice to
> get feedback and iron out some problems I currently have. Of course this
> needs explanation (hope that does not lead to tl;dr):
>
> Consider the following trace:
> short version (pseudo syntax):
>
> ```
> label(...,i,...)
> store(c,i) = load(a,i) + load(b,i)
> j = i+1
> guard(j<MAX)
> jump(...,j,...)
> ```
> long version: http://pastebin.com/e24s1vZg
>
> By unrolling this short trace, it is _NOT_ possible to vectorize it. The
> guard prohibits the store operation to be executed after the guard. I
> solved this problem by introducing a new guard (called 'early-exit'). It
> saves the live variables at the beginning of the trace. By finding the
> index calculations + guards and moving them above the early exit the
> following is possible:
>
> short version (pseudo syntax):
>
> ```
> label(...,i,...)
> j = i + 1
> guard(j<MAX)
> k = j + 1
> guard(k<MAX)
> guard_early_exit() # will not be emitted
> va = vec_load(a,i,2)
> vb = vec_load(b,i,2)
> vc = vec_add(va,vb)
> vec_store(c, i, 2) = vc
> jump(...,k,...)
> ```
> long version http://pastebin.com/vc3HaZCn
>
> My assumptions: Any guard that fails before the early exit must guide
> blackhole to the original loop at instruction 0. Only pure operations
> and the guards protecting the index are allowed to move before early-exit.
>
> The previous and the use of the live variables of the early exit (at the
> guard instructions) preserve correctness.
>
> I'm not quite sure how to handle the following problems:
>
> 1) I had the problem that uneven iterations moved to the blackhole
> interpreter and executed the loop from the beginning. I fixed it by
> resetting the blackhole interpreter position to the jitcode index 0.
> Is this the right way to start from the beginning?
>
> 2) Is there a better way to tell the blackhole interpreter to resume
> from the beginning of the trace, or even do not blackhole and just jump
> into the normal interpreter?
>
> 3) Are there any objections to do it this way (guard early-exit)?
>
> Best,
> Richard
>
> _______________________________________________
> pypy-dev mailing list
> pypy-dev@python.org
> https://mail.python.org/mailman/listinfo/pypy-dev
>
_______________________________________________
pypy-dev mailing list
pypy-dev@python.org
https://mail.python.org/mailman/listinfo/pypy-dev
[prev in list] [next in list] [prev in thread] [next in thread]