From: "Kewen.Lin" <linkw@linux.ibm.com>
To: Richard Biener <richard.guenther@gmail.com>,
Richard Sandiford <richard.sandiford@arm.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>,
Bill Schmidt <wschmidt@linux.ibm.com>,
Segher Boessenkool <segher@kernel.crashing.org>
Subject: Re: [PATCH v4] vect/rs6000: Support vector with length cost modeling
Date: Fri, 31 Jul 2020 20:37:03 +0800 [thread overview]
Message-ID: <4c7c935b-8eaa-1fd7-4357-72fd3239e710@linux.ibm.com> (raw)
In-Reply-To: <CAFiYyc0fwFsVzZP18ifSa6cpdiye8X7QBX7dFkB3=w87ib-Wdg@mail.gmail.com>
Hi Richards,
on 2020/7/31 下午7:20, Richard Biener wrote:
> On Fri, Jul 31, 2020 at 1:03 PM Richard Sandiford
> <richard.sandiford@arm.com> wrote:
>>
>> "Kewen.Lin" <linkw@linux.ibm.com> writes:
>>>>> + bool niters_known_p = LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo);
>>>>> + bool need_iterate_p
>>>>> + = (!LOOP_VINFO_EPILOGUE_P (loop_vinfo)
>>>>> + && !vect_known_niters_smaller_than_vf (loop_vinfo));
>>>>> +
>>>>> + /* Init min/max, shift and minus cost relative to single
>>>>> + scalar_stmt. For now we only use length-based partial vectors on
>>>>> + Power, target specific cost tweaking may be needed for other
>>>>> + ports in future. */
>>>>> + unsigned int min_max_cost = 2;
>>>>> + unsigned int shift_cost = 1, minus_cost = 1;
>>>>
>>>> Please instead add a scalar_min_max to vect_cost_for_stmt, and use
>>>> scalar_stmt for shift and minus. There shouldn't be any Power things
>>>> hard-coded into target-independent code.
>>>>
>>>
>>> Agree! It's not good to leave them there. I thought to wait and see
>>> if other targets which support vector with length can reuse this, or
>>> move it to target specific codes then if not sharable. But anyway
>>> it looks not good, let's fix it.
>
> In other generic places like this we simply use three generic scalar_stmt
> costs. At least I don't see how min_max should be different from it
> when shift can be the same as minus. Note this is also how we treat
Yeah, normally they (min/max/minus/shift) are taken as scalar_stmt, excepting
for fine-grain tuning like i386 port, they will use the same cost. On Power9,
to implement min/max it takes double cycles of the normal scalar operations
like add/shift, I was trying to model it more fine-grained since we probably
generate a few min/max here, if the loop body cost is small, I was worried
the decision isn't good enough. But yeah, in other generic places, the small
loop could also suffer this similar off, they are the same essentially.
> vectorization of MAX_EXPR - scalar cost is one scalar_stmt and
> vector cost is one vector_stmt. As you say below the add_stmt_cost
> hook can adjust based on the actual GIMPLE stmt -- if one is
> available (which indeed it isn't here).
>
> I'm somewhat lacking context here as well - we actually GIMPLE
> code-generate the min/max/shift/minus and only the eventual
> AND is defered to the target, right?
>
Yes, min/max/shift/minus are all GIMPLE code, targets like Power
will have its target specific cost for shift which moves length
to high bits 0:7.
One typical case is as below:
<bb 3> [local count: 105119324]:
_26 = n_11(D) * 4;
_37 = MAX_EXPR <_26, 16>;
_38 = _37 + 18446744073709551600;
_40 = MIN_EXPR <_26, 16>;
<bb 4> [local count: 630715945]:
# ivtmp_35 = PHI <0(3), ivtmp_36(4)>
# loop_len_30 = PHI <_40(3), _44(4)>
_19 = &MEM[base: a_12(D), index: ivtmp_35, offset: 0B];
vect_24 = .LEN_LOAD (_19, 4B, loop_len_30);
vect__3.7_23 = VIEW_CONVERT_EXPR<vector(4) unsigned int>(vect_24);
_1 = &MEM[base: b_13(D), index: ivtmp_35, offset: 0B];
vect_17 = .LEN_LOAD (_1, 4B, loop_len_30);
vect__5.10_9 = VIEW_CONVERT_EXPR<vector(4) unsigned int>(vect_17);
vect__7.11_8 = vect__5.10_9 + vect__3.7_23;
vect_28 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(vect__7.11_8);
_2 = &MEM[base: c_14(D), index: ivtmp_35, offset: 0B];
.LEN_STORE (_2, 4B, loop_len_30, vect_28);
_42 = MIN_EXPR <ivtmp_35, _38>;
_43 = _38 - _42;
_44 = MIN_EXPR <_43, 16>;
ivtmp_36 = ivtmp_35 + 16;
if (ivtmp_35 < _38)
goto <bb 4>; [83.33%]
else
goto <bb 5>; [16.67%]
>>> I had some concerns on vect_cost_for_stmt way, since it seems to allow
>>> more computations similar to min/max to be added like this, in long
>>> term it probably leads to the situtation like: scalar_min_max,
>>> scalar_div_expr, scalar_mul_expr ... an enum (cost types) bloat, it
>>> seems not good to maintain.
>>
>> I guess doing that doesn't seem so bad to me :-) I think it's been
>> a recurring problem that the current classification isn't fine-grained
>> enough for some cases.
>
> But we eventually want to get rid of this classification enum in favor
> of the add_stmt_cost hook ...
>
Nice, sounds like each target has to handle it fine-grain. :)
IIUC, the current modeling doesn't consider the instruction dependency and
execution resource etc. like scheduling, even if all costs are fine-grained
enough, the decision could be sub-optimal.
>>> I noticed that i386 port ix86_add_stmt_cost will check stmt_info->stmt,
>>> whether is assignment and the subcode of the expression, it provides the
>>> chance to check the statement more fine-grain, not just as normal
>>> scalar_stmt/vector_stmt.
>>>
>>> For the case here, we don't have the stmt_info, but we know the type
>>> of computation(expression), how about to extend the hook add_stmt_cost
>>> with one extra tree_code type argument, by default it can be some
>>> unmeaningful code, for some needs like here, we specify the tree_code
>>> as the code of computation, like {MIN,MAX}_EXPR, then target specific
>>> add_stmt_cost can check this tree_code and adjust the cost accordingly.
>>
>> If we do that, I guess we should “promote” code_helper out of
>> gimple-match.h and use that instead, so that we can handle
>> internal and built-in functions too.
>>
>> Would like to hear Richard's opinion on the best way forward here.
>
> I'd say defer this to a later patch and for now simply cost one scalar
> stmt for MIN/MAX. I agree that if we add a tree_code we want a
> code_helper instead. Note that I want to eventually have a
> full SLP tree for the final code generation where all info should be
> there (including SLP nodes for those min/max ops) and which the
> target could traverse. But I'm not sure if I can make enough progress
> on that SLP-only thing for GCC 11 even...
>
OK, I'm fine to take MIN/MAX as scalar_stmt here. Thank both of you!
This new SLP framework looks very promising and powerful. :-)
BR,
Kewen
next prev parent reply other threads:[~2020-07-31 12:37 UTC|newest]
Thread overview: 23+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-07-21 5:51 [PATCH] vect: " Kewen.Lin
2020-07-21 7:57 ` Richard Biener
2020-07-22 1:26 ` [PATCH v2] vect/rs6000: " Kewen.Lin
2020-07-22 6:38 ` Richard Biener
2020-07-22 7:08 ` Kewen.Lin
2020-07-22 9:11 ` Richard Sandiford
2020-07-22 15:48 ` [PATCH v3] " Kewen.Lin
2020-07-22 16:25 ` Kewen.Lin
2020-07-24 16:21 ` Richard Sandiford
2020-07-27 3:58 ` [PATCH v4] " Kewen.Lin
2020-07-27 13:40 ` Richard Sandiford
2020-07-28 8:36 ` Kewen.Lin
2020-07-31 11:03 ` Richard Sandiford
2020-07-31 11:20 ` Richard Biener
2020-07-31 12:37 ` Kewen.Lin [this message]
2020-07-31 13:01 ` Richard Biener
2020-07-31 13:21 ` Kewen.Lin
2020-07-31 14:51 ` [PATCH v5] " Kewen.Lin
2020-08-05 7:27 ` Richard Sandiford
2020-08-05 14:06 ` Segher Boessenkool
2020-08-06 6:47 ` Kewen.Lin
2020-07-22 17:49 ` [PATCH v2] " Segher Boessenkool
2020-07-27 3:44 ` Kewen.Lin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=4c7c935b-8eaa-1fd7-4357-72fd3239e710@linux.ibm.com \
--to=linkw@linux.ibm.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.guenther@gmail.com \
--cc=richard.sandiford@arm.com \
--cc=segher@kernel.crashing.org \
--cc=wschmidt@linux.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).