RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
Vladimir Kozlov
vladimir.kozlov at oracle.com
Fri Oct 16 04:48:41 UTC 2015
On 10/16/15 9:12 AM, Civlin, Jan wrote:
> Vladimir,
>
> Michael Berg is about to add this new function for filtering based on vlen, so I use it now for my case and Michael will add more cases soon.
>
> const bool Matcher::match_rule_supported_vlen(int opcode, int vlen)
For your changes I would simple add next to the check in
VectorNode::implemented():
&& (vopc != Op_CMoveD || vlen == 4);
For changes which Michael is working on, you have to add it to all
platforms .ad (same as match_rule_supported()).
I would suggest to have match_rule_supported_vector(int opcode, int
vlen) which you will call from VectorNode::implemented() instead of
match_rule_supported(). And move all current vector opcodes check in
match_rule_supported() to new one.
Thanks,
Vladimir
>
>
> The updated webrev is here:
>
> http://cr.openjdk.java.net/~iveresov/vector-cmove/webrev.02/
>
>
> Thank you,
>
> Jan.
>
> -----Original Message-----
> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
> Sent: Wednesday, October 14, 2015 2:30 AM
> To: Civlin, Jan
> Cc: hotspot compiler
> Subject: Re: RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
>
> I said before that condition in .ad file in Matcher::match_rule_supported():
>
> case Op_CMoveVD:
> if (UseAVX > 2) <<<<< (UseAVX < 1 || UseAVX > 2)
> ret_value = false;
>
> should match predicates:
>
> predicate(UseAVX > 0 && UseAVX < 3 && n->as_Vector()->length() == 4);
>
> Also why only length 4? There is no length 2? I don't see any code in superword.cpp which limits size of CMoveVD vector.
>
> Otherwise it seems fine.
>
> Thanks,
> Vladimir
>
> On 10/14/15 1:57 AM, Civlin, Jan wrote:
>> Vladimir,
>>
>> the updated webrev is at
>>
>> http://cr.openjdk.java.net/~iveresov/vector-cmove/webrev.01/
>>
>> Thank you, Igor.
>>
>> Best,
>>
>> Jan.
>>
>> -----Original Message-----
>> From: Igor Veresov [mailto:igor.veresov at oracle.com]
>> Sent: Tuesday, October 13, 2015 10:51 AM
>> To: Civlin, Jan
>> Cc: Vladimir Kozlov
>> Subject: Re: RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
>>
>> Done.
>>
>>> On Oct 13, 2015, at 9:51 AM, Civlin, Jan <jan.civlin at intel.com> wrote:
>>>
>>> Igor,
>>>
>>> I got a response from the moderator that the previous message (attached below) has been rejected due to the size limits.
>>> Please upload the attached webrev (I guess it will be
>>> http://cr.openjdk.java.net/~iveresov/vector-cmove/webrev.01/)
>>>
>>> And I'll send the link to the group.
>>>
>>> Thank you,
>>>
>>> Jan.
>>>
>>> -----Original Message-----
>>> From: Civlin, Jan
>>> Sent: Monday, October 12, 2015 10:19 PM
>>> To: Vladimir Kozlov
>>> Cc: Igor Veresov; hotspot compiler; Civlin, Jan
>>> Subject: RE: RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
>>>
>>> Vladimir,
>>>
>>> Here is a webrev built from a single changeset.
>>> Please use it for further reviewing.
>>>
>>> Thank you,
>>>
>>> Jan.
>>>
>>> -----Original Message-----
>>> From: Civlin, Jan
>>> Sent: Saturday, October 10, 2015 12:14 PM
>>> To: Vladimir Kozlov
>>> Cc: hotspot compiler
>>> Subject: RE: RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
>>>
>>> Vladimir,
>>>
>>> Thank you for the quick review.
>>>
>>>
>>> I thought you are patching your repo using hotspot.patch which is clean of empty files or .hgtag and you use webrev only for facilitation of visual code review.
>>> If you use hotspot.patch it does not propagate to your repo empty or already integrated in 8136725 files or my repo tags, since none of them are in the patch.
>>> I thought you are simply skipping unchanged files if you see them in webrev html pages.
>>>
>>> My repo is a consistent and long living clone of your repo. It is a single place that includes the history of all my changes, which let me go ahead in development way far of what is ready for going out to you. I can create an additional repo and filter out my tags on way out, but it will create more possibilities for accidental errors.
>>>
>>>
>>> x86.ad: it is indeed not (yet) AVX3, since AVX3 will have different ISA. Once we have AVX3 in the house and can debug it, AVX3 case will be added.
>>>
>>> c2_globals.hpp:
>>> I will rename the flag to UseCMoveUnconditionally, but I see no point of having another flag for the vector: once SuperWord detects CMovD, it merely vectorize it.
>>> In the opposite direction: one cannot ask for vectorization of CMovD and not to get them in the scalar case.
>>>
>>> compile.cpp:
>>> 1114: will add "do_vector_loop() &&" and change to NOT_PRODUCT(if
>>> (do_vector_loop() && Verbose&& && has_method())
>>> {tty->print("Compile::Init: use CMove without profitability tests for
>>> method %s\n", method()->name()->as_quoted_ascii());})
>>>
>>> loopopts.cpp:
>>> 601: will add "&& cmp_op == Op_CmpD" and change to if
>>> (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going
>>>
>>>
>>> Thank you,
>>>
>>> Jan
>>>
>>>
>>> -----Original Message-----
>>> From: Vladimir Kozlov [mailto:vladimir.kozlov at oracle.com]
>>> Sent: Saturday, October 10, 2015 5:52 AM
>>> To: Civlin, Jan
>>> Cc: hotspot compiler
>>> Subject: Re: RFR: 8139340: SuperWord enhancement to support vector conditional move (CMovVD ) on Intel AVX cpu.
>>>
>>> Changes contains also 8136725 (loop reserve copy) changes. Please, clean up.
>>>
>>> x86.ad WAIT! According to next conditions (BTW, support condition should match predicate) this is only supported with AVX1 and AVX2 and not AVX3. Really?
>>>
>>> I think it should be reversed condition (< 3):
>>> + case Op_CMoveVD:
>>> + if (UseAVX > 2)
>>> + ret_value = false;
>>>
>>> + instruct vcmov4D_reg(vecY dst, vecY src1, vecY src2, immI8 cop,
>>> cmpOp_vcmppd copnd) %{
>>> + predicate(UseAVX > 0 && UseAVX < 3 && n->as_Vector()->length() ==
>>> + 4);
>>>
>>>
>>> x86_64.ad changes are empty (could be only spacing change) so remove it from change.
>>>
>>> c2_globals.hpp DoReserveCopyInSuperWord is defined in 8136725 changes.
>>> UseCMov should be UseCMoveUnconditionally I think (note 'move'
>>> instead of 'mov'). Also I am not sure that you should mix 2 things
>>> into this
>>> changes: one is vectorizing cmoveD (for doubles) and always generate cmoves. I think it is different issues and this change should concentrate on vectorizing only double cmoves. Then flag could be VectorizeCMoveD.
>>>
>>> compile.cpp print not under UseCmov flag:
>>> + NOT_PRODUCT(if (Verbose && has_method()) {tty->print("Compile::Init:
>>> use CMove without profitability tests for method %s\n",
>>> method()->name()->as_quoted_ascii());})
>>>
>>> loopopts.cpp changes inconsistent. In one place you check use_cmove only for doubles and in other for all.
>>>
>>> I will look on superword changes after you cleanup changes: remove
>>> 8136725 and also create history from fresh repo - I don't want to see 50 revisions descriptions in webrev.
>>>
>>> Thanks,
>>> Vladimir
>>>
>>> On 10/10/15 7:04 AM, Civlin, Jan wrote:
>>>> Please review this patch.
>>>>
>>>> Hi All,
>>>>
>>>>
>>>> We would like to contribute the SuperWord enhancement to support
>>>> vector conditional move (CMovVD ) on Intel AVX cpu.
>>>>
>>>>
>>>> The contribution Bug ID: 8139340.
>>>>
>>>> Please review this patch:
>>>>
>>>> Bug-id: https://bugs.openjdk.java.net/browse/JDK-8139340
>>>>
>>>> webrev: http://cr.openjdk.java.net/~iveresov/vector-cmove/webrev.00/
>>>>
>>>> Description:
>>>>
>>>> SuperWord enhancement to support vector conditional move (CMovVD)
>>>> on Intel AVX cpu.
>>>>
>>>> The SuperWord optimization bails out on counted loops that contain
>>>> any conditional statement other than the loop exit, and this
>>>> prevents vectorization of many compute bound loops.
>>>>
>>>> The proposed enhancement enables generation of CMovD on demand
>>>> (-XX:+UseCMov), and further vectorization of CMovD (into CMovVD )
>>>> in SuperWord optimization.
>>>>
>>>> The performance gain observed on a simplified Monte Carlo Option
>>>> Calculation was up to 2x speed-up.
>>>>
>>>> Thank you,
>>>>
>>>> Jan.
>>>>
>>>> *From:*Igor Veresov [mailto:igor.veresov at oracle.com]
>>>> *Sent:* Friday, October 9, 2015 3:22 PM
>>>> *To:* Civlin, Jan
>>>> *Cc:* hotspot compiler; Vladimir Kozlov
>>>> *Subject:* Re: SuperWord enhancement to support vector conditional
>>>> move (CMovVD ) on Intel AVX cpu.
>>>>
>>>> Here it is: https://bugs.openjdk.java.net/browse/JDK-8139340
>>>>
>>>> igor
>>>>
>>>> On Oct 9, 2015, at 2:57 PM, Civlin, Jan <jan.civlin at intel.com
>>>> <mailto:jan.civlin at intel.com>> wrote:
>>>>
>>>> Thank you, Igor.
>>>>
>>>> What is the RFR for this?
>>>>
>>>> *From:*Igor Veresov [mailto:igor.veresov at oracle.com]
>>>> *Sent:*Friday, October 9, 2015 2:53 PM
>>>> *To:*Civlin, Jan
>>>> *Cc:*hotspot compiler; Vladimir Kozlov
>>>> *Subject:*Re: SuperWord enhancement to support vector conditional
>>>> move (CMovVD ) on Intel AVX cpu.
>>>>
>>>> Here the webrev:
>>>> http://cr.openjdk.java.net/~iveresov/vector-cmove/webrev.00/
>>>>
>>>> igor
>>>>
>>>> On Oct 9, 2015, at 1:15 PM, Civlin, Jan <jan.civlin at intel.com
>>>> <mailto:jan.civlin at intel.com>> wrote:
>>>>
>>>> Igor,
>>>>
>>>> Please create RFR and upload this patch. You may need to rename
>>>> ancnav.js.remove_this_extention back to ancnav.js (I have to
>>>> rename it for passing the mail server filters).
>>>>
>>>> Description:
>>>>
>>>> SuperWord enhancement to support vector conditional move
>>>> (CMovVD) on Intel AVX cpu.
>>>>
>>>> The SuperWord optimization bails out on counted loops that
>>>> contain any conditional statement other than the loop exit, and
>>>> this prevents vectorization of many compute bound loops.
>>>>
>>>> The proposed enhancement enables generation of CMovD on demand
>>>> (-XX:+UseCMov), and further vectorization of CMovD (into CMovVD
>>>> ) in SuperWord optimization.
>>>>
>>>> The performance gain observed on a simplified Monte Carlo Option
>>>> Calculation was up to 2x speed-up.
>>>>
>>>> Thank you,
>>>>
>>>> Jan.
>>>>
>>>> <webrev-r9162-9157.tar.bz2>
>>>>
>>> <webrev-r9050.tar.bz2>
>>
More information about the hotspot-compiler-dev
mailing list