Re: [PATCH v3 1/2] x86/fpu: track AVX-512 usage of tasks

[Li, Aubrey]

On 2018/11/18 22:03, Samuel Neves wrote:
> On 11/17/18 12:36 AM, Li, Aubrey wrote:
>> On 2018/11/17 7:10, Dave Hansen wrote:
>>> Just to be clear: there are 3 AVX-512 XSAVE states:
>>>
>>>          XFEATURE_OPMASK,
>>>          XFEATURE_ZMM_Hi256,
>>>          XFEATURE_Hi16_ZMM,
>>>
>>> I honestly don't know what XFEATURE_OPMASK does.  It does not appear to
>>> be affected by VZEROUPPER (although VZEROUPPER's SDM documentation isn't
>>> looking too great).
> 
> XFEATURE_OPMASK refers to the additional 8 mask registers used in
> AVX512. These are more similar to general purpose registers than
> vector registers, and should not be too relevant here.
> 
>>>
>>> But, XFEATURE_ZMM_Hi256 is used for the upper 256 bits of the
>>> registers ZMM0-ZMM15.  Those are AVX-512-only registers.  The only way
>>> to get data into XFEATURE_ZMM_Hi256 state is by using AVX512 instructions.
>>>
>>> XFEATURE_Hi16_ZMM is the same.  The only way to get state in there is
>>> with AVX512 instructions.
>>>
>>> So, first of all, I think you *MUST* check XFEATURE_ZMM_Hi256 and
>>> XFEATURE_Hi16_ZMM.  That's without question.
>>
>> No, XFEATURE_ZMM_Hi256 does not request turbo license 2, so it's less
>> interested to us.
>>
> 
> I think Dave is right, and it's easy enough to check this. See the
> attached program. For the "high current" instruction vpmuludq
> operating on zmm0--zmm3 registers, we have (on a Skylake-SP Xeon Gold
> 5120)
> 
>            175,097      core_power.lvl0_turbo_license:u
>                      ( +-  2.18% )
>             41,185      core_power.lvl1_turbo_license:u
>                      ( +-  1.55% )
>         83,928,648      core_power.lvl2_turbo_license:u
>                      ( +-  0.00% )
> 
> while for the same code operating on zmm28--zmm31 registers, we have
> 
>            163,507      core_power.lvl0_turbo_license:u
>                      ( +-  6.85% )
>             47,390      core_power.lvl1_turbo_license:u
>                      ( +- 12.25% )
>         83,927,735      core_power.lvl2_turbo_license:u
>                      ( +-  0.00% )
> 
> In other words, the register index does not seem to matter at all for
> turbo license purposes (this makes sense, considering these chips have
> 168 vector registers internally; zmm15--zmm31 are simply newly exposed
> architectural registers).
> 
> We can also see that XFEATURE_Hi16_ZMM does not imply license 1 or 2;
> we may be using xmm15--xmm31 purely for the convenient extra register
> space. For example, cases 4 and 5 of the sample program:
> 
>         84,064,239      core_power.lvl0_turbo_license:u
>                      ( +-  0.00% )
>                  0      core_power.lvl1_turbo_license:u
>                  0      core_power.lvl2_turbo_license:u
> 
>         84,060,625      core_power.lvl0_turbo_license:u
>                      ( +-  0.00% )
>                  0      core_power.lvl1_turbo_license:u
>                  0      core_power.lvl2_turbo_license:u
> 

Thanks for your program, Samuel, it's very helpful. But I saw a different
output on my side, May I have your glibc version?

Thanks,
-Aubrey

> So what's most important is the width of the vectors being used, not
> the instruction set or the register index. Second to that is the
> instruction type, namely whether those are "heavy" instructions.
> Neither of these things can be accurately captured by the XSAVE state.
> 
>>>
>>> It's probably *possible* to run AVX512 instructions by loading state
>>> into the YMM register and then executing AVX512 instructions that only
>>> write to memory and never to register state.  That *might* allow
>>> XFEATURE_Hi16_ZMM and XFEATURE_ZMM_Hi256 to stay in the init state, but
>>> for the frequency to be affected since AVX512 instructions _are_
>>> executing.  But, there's no way to detect this situation from XSAVE
>>> states themselves.
>>>
>>
>> Andi should have more details on this. FWICT, not all AVX512 instructions
>> has high current, those only touching memory do not cause notable frequency
>> drop.
> 
> According to section 15.26 of the Intel optimization reference manual,
> "heavy" instructions consist of floating-point and integer
> multiplication. Moves, adds, logical operations, etc, will request at
> most turbo license 1 when operating on zmm registers.
> 
>>
>> Thanks,
>> -Aubrey
>>