Re: [PATCH 3/6 -tip] perf_counter: Add Generalized Hardware vectored co-processor support for AMD

[Ingo Molnar <mingo@elte.hu>]

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

> $ ./perf stat -e add -e multiply -e divide -e vec-idle-cycles -e vec-stall-cycles -e vec-ops -- /usr/bin/vlc ~jaswinder/Videos/Linus_Torvalds_interview_with_Charlie_Rose_Part_1.flv
> 
>  Performance counter stats for '/usr/bin/vlc /home/jaswinder/Videos/Linus_Torvalds_interview_with_Charlie_Rose_Part_1.flv':
> 
>     20177177044  vec-adds                  (scaled from 66.63%)
>     34101687027  vec-muls                  (scaled from 66.64%)
>      3984060862  vec-divs                  (scaled from 66.71%)
>     26349684710  vec-idle-cycles           (scaled from 66.65%)
>      9052001905  vec-stall-cycles          (scaled from 66.66%)
>     76440734242  vec-ops                   (scaled from 66.71%)
> 
>   272.523058097  seconds time elapsed

Ok, this looks very nice now - a highly generic and still very 
useful looking categorization of FPU/MMX/SSE related co-processor hw 
events.

I'm still waiting for feedback from Paulus, BenH and Anton, whether 
this kind of generic enumeration fits PowerPC well enough.

I think from a pure logic/math/physics POV this categorization is 
pretty complete: a modern co-processor has three fundamental states 
we are interested in: idle, busy and busy-stalled. It has an 'ops' 
metric that counts instructions, plus the main operations are add, 
mul and div.

Cell is i guess a complication to be solved, as there the various 
vector units have separate decoders and separate thread state. This 
above abstraction only covers the portion of CPU designs where there 
are vector operations in the main ALU decoder stream of instructions

One thing that might be worth exposing is vectored loads/stores in 
general. But we dont have those in the generic ALU enumeration yet 
and if then it should be done together.

Also, the Nehalem bits need to be tested, i'll try to find time for 
that.

Good stuff.

	Ingo