Re: [PATCH -tip] perf_counter: Add Generalized Hardware FPU support for AMD

[Jaswinder Singh Rajput <jaswinder@kernel.org>]

On Tue, 2009-06-30 at 18:50 +0530, Jaswinder Singh Rajput wrote:
> On Tue, 2009-06-30 at 12:11 +0200, Ingo Molnar wrote:
> > * Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:
> > 
> > >  $./perf stat -e add -e multiply -e fpu-store -e fpu-empty -e fpu-busy -e x87 -e mmx-3dnow -e sse-sse2 -- ls -lR /usr/include/ > /dev/null
> > > 
> > >  Performance counter stats for 'ls -lR /usr/include/':
> > > 
> > >            7335  add                       (   2.00x scaled)
> > >            8012  multiply                  (   1.99x scaled)
> > >            5229  fpu-store                 (   2.00x scaled)
> > >       793097355  fpu-empty                 (   2.00x scaled)
> > >             182  fpu-busy                  (   2.00x scaled)
> > >               6  x87                       (   2.01x scaled)
> > >               4  mmx-3dnow                 (   2.00x scaled)
> > >            8933  sse-sse2                  (   2.00x scaled)
> > > 
> > >     0.393548820  seconds time elapsed
> > > 
> > >  $./perf stat -e add -e multiply -e fpu-store -e fpu-empty -e fpu-busy -e x87 -e mmx-3dnow -e sse-sse2 -- /usr/bin/rhythmbox ~jaswinder/Music/singhiskinng.mp3
> > > 
> > >  Performance counter stats for '/usr/bin/rhythmbox /home/jaswinder/Music/singhiskinng.mp3':
> > > 
> > >        19583739  add                       (   2.01x scaled)
> > >        20856051  multiply                  (   2.01x scaled)
> > >        18669503  fpu-store                 (   2.00x scaled)
> > >     25100224054  fpu-empty                 (   1.99x scaled)
> > >        12540131  fpu-busy                  (   1.99x scaled)
> > >          207228  x87                       (   1.99x scaled)
> > >         1768418  mmx-3dnow                 (   2.00x scaled)
> > >        42286702  sse-sse2                  (   2.01x scaled)
> > > 
> > >   302.698647617  seconds time elapsed
> > > 
> > >  $./perf stat -e add -e multiply -e fpu-store -e fpu-empty -e fpu-busy -e x87 -e mmx-3dnow -e sse-sse2 -- /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':
> > > 
> > >      6572682335  add                       (   2.00x scaled)
> > >     11131555181  multiply                  (   2.00x scaled)
> > >      1317520699  fpu-store                 (   2.00x scaled)
> > >      9089415134  fpu-empty                 (   1.99x scaled)
> > >      2902772713  fpu-busy                  (   2.00x scaled)
> > >           26047  x87                       (   2.00x scaled)
> > >     24850978532  mmx-3dnow                 (   2.00x scaled)
> > >       262276117  sse-sse2                  (   2.01x scaled)
> > > 
> > >    96.169312358  seconds time elapsed
> > > 
> > > Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
> > > ---
> > >  arch/x86/kernel/cpu/perf_counter.c |   34 ++++++++++++++++++++++++++++++
> > >  include/linux/perf_counter.h       |   17 +++++++++++++++
> > >  kernel/perf_counter.c              |    1 +
> > >  tools/perf/util/parse-events.c     |   40 ++++++++++++++++++++++++++++++++++++
> > >  4 files changed, 92 insertions(+), 0 deletions(-)
> > > 
> > > diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
> > > index b83474b..4417edf 100644
> > > --- a/arch/x86/kernel/cpu/perf_counter.c
> > > +++ b/arch/x86/kernel/cpu/perf_counter.c
> > > @@ -372,6 +372,12 @@ static const u64 atom_hw_cache_event_ids
> > >   },
> > >  };
> > >  
> > > +/*
> > > + * Generalized hw fpu event table
> > > + */
> > > +
> > > +static u64 __read_mostly hw_fpu_event_ids[PERF_COUNT_HW_FPU_MAX];
> > 
> > ok, this looks genuinely useful, but there are some gaps. Where's 
> > the divides? 
> 
> I was also surprised divide is not available for AMD. Thats why I did
> not included it. You are right it should be there.
> 

In AMD FPU operations include add, multiple and store.
Can I use store as divide for AMD, samples I shown above seems like they
are divide.

> > Plus things like mmx-3dnow are AMD specific, sse-sse2 
> > is x86 specific. We definitely want this general table, but the 
> > events should be truly general.
> > 
> 
> mmx and sse are available for both Intel and AMD. Thats why I added both
> of them. Is it OK.
> 

Is this looks :

 enum perf_hw_fpu_id {
       PERF_COUNT_HW_FPU_ADD                   = 0,
       PERF_COUNT_HW_FPU_MULTIPLY              = 1,
       PERF_COUNT_HW_FPU_DIVIDE                = 2,
       PERF_COUNT_HW_FPU_EMPTY                 = 3,
       PERF_COUNT_HW_FPU_STALL                 = 4,
       PERF_COUNT_HW_FPU_X87                   = 5,
       PERF_COUNT_HW_FPU_MMX                   = 6,
       PERF_COUNT_HW_FPU_SSE                   = 7,

       PERF_COUNT_HW_FPU_MAX,                  /* non-ABI */


> > Also, how would this look like on Intel, roughly?
> > 
> 
> Intel have almost all of them + divide.
> 
> As you know I work from home and I do not have any Intel machine which
> supports PMU.
> 
> Can you suggest your machine name so that I can prepare the FPU events
> list for your machine and you can verify it on your side.
> 

For Nehalem it will look like :

 static const u64 nehalem_hw_fpu_event_ids[] =
 {
  [PERF_COUNT_HW_FPU_ADD]               = 0x01B1, /* UOPS_EXECUTED.PORT0     */
  [PERF_COUNT_HW_FPU_MULTIPLY]          = 0x0214, /* ARITH.MUL               */
  [PERF_COUNT_HW_FPU_DIVIDE]            = 0x0114, /* ARITH.CYCLES_DIV_BUSY   */
  [PERF_COUNT_HW_FPU_EMPTY]             = 0x0,
  [PERF_COUNT_HW_FPU_STALL]             = 0x60A2, /* RESOURCE_STALLS.FPCW|MXCSR*/
  [PERF_COUNT_HW_FPU_X87]               = 0x0110, /* FP_COMP_OPS_EXE.X87     */
  [PERF_COUNT_HW_FPU_MMX]               = 0x0210, /* FP_COMP_OPS_EXE.MMX     */
  [PERF_COUNT_HW_FPU_SSE]               = 0x0410, /* FP_COMP_OPS_EXE.SSE_FP  */
};

Is these looks OK to you. Can I resend the patch based on these.

Thanks,
--
JSR