Re: [PATCH 1/2 -tip] perf_counter: Add generalized hardware vectored co-processor support for AMD and Intel Corei7/Nehalem

[Ingo Molnar <mingo@elte.hu>]

* Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:

> On Fri, 2009-07-03 at 18:19 +0530, Jaswinder Singh Rajput wrote:
> > On Fri, 2009-07-03 at 17:25 +0530, Jaswinder Singh Rajput wrote:
> > > On Fri, 2009-07-03 at 12:29 +0200, Ingo Molnar wrote:
> > > > * Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:
> > > > 
> > > > >  Performance counter stats for '/usr/bin/rhythmbox /home/jaswinder/Music/singhiskinng.mp3':
> > > > > 
> > > > >        17552264  vec-adds                  (scaled from 66.28%)
> > > > >        19715258  vec-muls                  (scaled from 66.63%)
> > > > >        15862733  vec-divs                  (scaled from 66.82%)
> > > > >     23735187095  vec-idle-cycles           (scaled from 66.89%)
> > > > >        11353159  vec-stall-cycles          (scaled from 66.90%)
> > > > >        36628571  vec-ops                   (scaled from 66.48%)
> > > > 
> > > > Is stall-cycles equivalent to busy-cycles? 
> > > 
> > > 
> > > hmm, normally we can use these terms interchangeably. But they can be
> > > different some times.
> > > 
> > > busy means it is already executing some instructions so it will not take
> > > another instruction.
> > > 
> > > stall can be busy(executing) or non-executing may be it is waiting for
> > > some operands due to cache miss.
> > > 
> > > 
> > > > I.e. do we have this 
> > > > general relationship to the cycle event:
> > > > 
> > > > 	cycles = vec-stall-cycles + vec-idle-cycles
> > > > 
> > > > ?
> > 
> > Like on AMD :
> > 
> >     13390918485  vec-adds                  (scaled from 57.07%)
> >     22465091289  vec-muls                  (scaled from 57.22%)
> >      2643789384  vec-divs                  (scaled from 57.21%)
> >     17922784596  vec-idle-cycles           (scaled from 57.23%)
> >      6402888606  vec-stall-cycles          (scaled from 57.17%)
> >     55823491597  cycles                    (scaled from 57.05%)
> >     51035264218  vec-ops                   (scaled from 57.05%)
> > 
> >   187.494664172  seconds time elapsed
> > 
> > vec-idle-cycles + vec-stall-cycles = 24325673202
> > 
> > so cycles = 2.29 * (vec-idle-cycles + vec-stall-cycles)

that equation is entirely bogus.

> > 
> > On AMD I used : EventSelect 0D7h Dispatch Stall for FPU Full The 
> > number of processor cycles the decoder is stalled because the 
> > scheduler for the Floating Point Unit is full. This condition 
> > can be caused by a lack of parallelism in FP-intensive code, or 
> > by cache misses on FP operand loads (which could also show up as 
> > EventSelect 0D8h instead, depending on the nature of the 
> > instruction sequences). May occur simultaneously with certain 
> > other stall conditions; see EventSelect 0D1h
> > 
> > So stall is due to lack of parallelism and cache misses. If we 
> > keep on increasing the size of FP units and cache may at some 
> > point be we can get vec-stall-cycles = zero.
> > 
> 
> I mean, So stall is majorly due to lack of parallelism and cache 
> misses. If we keep on increasing the size of FP units and cache 
> then stall time will keep on decreasing (ofcourse it will be never 
> Zero ;)
> 
> And same thing will be happen for Intel.
> 
> So stall is not equal to busy.
> 
> Please let me know what is next, should I remove busy term from 
> alias.

What is needed is for you to understand these events and provide a 
generalization around them that makes sense. Or to declare it 
honestly when you dont.

The numbers simply dont add up:

> >     13390918485  vec-adds                  (scaled from 57.07%)
> >     22465091289  vec-muls                  (scaled from 57.22%)
> >      2643789384  vec-divs                  (scaled from 57.21%)
> >     17922784596  vec-idle-cycles           (scaled from 57.23%)
> >      6402888606  vec-stall-cycles          (scaled from 57.17%)
> >     55823491597  cycles                    (scaled from 57.05%)
> >     51035264218  vec-ops                   (scaled from 57.05%)

vec-idle-cycles + vec-stall-cycles does not add up to cycles - 
because a stall is not an 'interchangeable' term with 'busy' as you 
claimed before, but a special state of the pipeline, a subset of 
busy.

I prefer to apply patches from people who understand what they are 
doing - and more importantly, who express and declare their own 
limits properly when they _dont_ understand something and are 
guessing.

Frankly, your patches dont give me this impression and you are also 
babbling way too much about things you clearly dont understand, and 
thus you hinder the discussions with noise.

It's not bad at all to not understand something (we all are at 
various stages of a big and constantly refreshing learning curves), 
but it's very bad to pretend you understand something while you 
clearly dont. What we need in lkml discussions is an honest laying 
down of facts, opinions and doubts.

Why the heck didnt you say:

 " I dont know much about PMUs or vector units yet, but I have found
   these blurbs in the Intel and AMD docs and what do you think 
   about structuring these events the following way. Someone who 
   knows this stuff should review this first, it is quite likely 
   incomplete. "

	Ingo