From: "Liu, Hongtao" <hongtao.liu@intel.com>
To: "Beulich, Jan" <JBeulich@suse.com>,
"gcc-patches@gcc.gnu.org" <gcc-patches@gcc.gnu.org>
Cc: Kirill Yukhin <kirill.yukhin@gmail.com>
Subject: RE: [PATCH] x86: improve fast bfloat->float conversion
Date: Tue, 11 Jul 2023 06:45:34 +0000 [thread overview]
Message-ID: <SA1PR11MB6757BFE318BA27890B8041E6E531A@SA1PR11MB6757.namprd11.prod.outlook.com> (raw)
In-Reply-To: <d75a4d5a-8624-aa77-9f29-140767357b58@suse.com>
> -----Original Message-----
> From: Jan Beulich <jbeulich@suse.com>
> Sent: Tuesday, July 11, 2023 2:08 PM
> To: gcc-patches@gcc.gnu.org
> Cc: Liu, Hongtao <hongtao.liu@intel.com>; Kirill Yukhin
> <kirill.yukhin@gmail.com>
> Subject: [PATCH] x86: improve fast bfloat->float conversion
>
> There's nothing AVX512BW-ish in here, so no reason to use Yw as the
> constraints for the AVX alternative. Furthermore by using the 512-bit form of
> VPSSLD (in a new alternative) all 32 registers can be used directly by the insn
> without AVX512VL needing to be enabled.
Yes, the instruction vpslld doesn't need AVX512BW, the patch LGTM.
>
> Also adjust the originally last alternative's "prefix" attribute to maybe_evex.
>
> gcc/
>
> * config/i386/i386.md (extendbfsf2_1): Add new AVX512F
> alternative. Adjust original last alternative's "prefix"
> attribute to maybe_evex.
> ---
> The corresponding expander, "extendbfsf2", looks to have been dead since
> its introduction in a1ecc5600464 ("Fix incorrect _mm_cvtsbh_ss"): The builtin
> references the insn (extendbfsf2_1), not the expander. Can't the expander
> be deleted and the name of the insn then pruned of the _1 suffix? If so, that
> further raises the question of the significance of the "!HONOR_NANS
> (BFmode)" that the expander has, but the insn doesn't have. Which may
> instead suggest the builtin was meant to reference the expander. Yet then I
> can't see what would the builtin would expand to when HONOR_NANS
> (BFmode) it true.
Quote from what Jakub said in [1].
-------
This is not correct.
While using such code for _mm_cvtsbh_ss is fine if it is documented not to
raise exceptions and turn a sNaN into a qNaN, it is not fine for HONOR_NANS
(i.e. when -ffast-math is not on), because a __bf16 -> float conversion
on sNaN should raise invalid exception and turn it into a qNaN.
We could have extendbfsf2 expander that would FAIL; if HONOR_NANS and
emit extendbfsf2_1 otherwise.
-------
[1] https://gcc.gnu.org/pipermail/gcc-patches/2022-November/607108.html
>
> I further wonder whether the nearby "extendhfdf2" expander is really
> needed. It doesn't look to specify anything that the corresponding insn
> doesn't also specify.
>
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -5181,21 +5181,27 @@
> ;; Don't use float_extend since psrlld doesn't raise ;; exceptions and turn a
> sNaN into a qNaN.
> (define_insn "extendbfsf2_1"
> - [(set (match_operand:SF 0 "register_operand" "=x,Yw")
> + [(set (match_operand:SF 0 "register_operand" "=x,Yv,v")
> (unspec:SF
> - [(match_operand:BF 1 "register_operand" " 0,Yw")]
> + [(match_operand:BF 1 "register_operand" " 0,Yv,v")]
> UNSPEC_CVTBFSF))]
> "TARGET_SSE2"
> "@
> pslld\t{$16, %0|%0, 16}
> - vpslld\t{$16, %1, %0|%0, %1, 16}"
> - [(set_attr "isa" "noavx,avx")
> + vpslld\t{$16, %1, %0|%0, %1, 16}
> + vpslld\t{$16, %g1, %g0|%g0, %g1, 16}"
> + [(set_attr "isa" "noavx,avx,*")
> (set_attr "type" "sseishft1")
> (set_attr "length_immediate" "1")
> - (set_attr "prefix_data16" "1,*")
> - (set_attr "prefix" "orig,vex")
> - (set_attr "mode" "TI")
> - (set_attr "memory" "none")])
> + (set_attr "prefix_data16" "1,*,*")
> + (set_attr "prefix" "orig,maybe_evex,evex")
> + (set_attr "mode" "TI,TI,XI")
> + (set_attr "memory" "none")
> + (set (attr "enabled")
> + (if_then_else (eq_attr "alternative" "2")
> + (symbol_ref "TARGET_AVX512F && !TARGET_AVX512VL
> + && !TARGET_PREFER_AVX256")
> + (const_string "*")))])
>
> (define_expand "extend<mode>xf2"
> [(set (match_operand:XF 0 "nonimmediate_operand")
next prev parent reply other threads:[~2023-07-11 6:45 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-07-11 6:08 Jan Beulich
2023-07-11 6:45 ` Liu, Hongtao [this message]
2023-07-11 7:50 ` Jan Beulich
2023-07-11 8:36 ` Jakub Jelinek
2023-07-11 8:39 ` Liu, Hongtao
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=SA1PR11MB6757BFE318BA27890B8041E6E531A@SA1PR11MB6757.namprd11.prod.outlook.com \
--to=hongtao.liu@intel.com \
--cc=JBeulich@suse.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=kirill.yukhin@gmail.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).