From: Hongtao Liu <crazylht@gmail.com>
To: Joseph Myers <joseph@codesourcery.com>
Cc: Jakub Jelinek <jakub@redhat.com>,
liuhongt <hongtao.liu@intel.com>,
"gcc-patches@gcc.gnu.org" <gcc-patches@gcc.gnu.org>
Subject: Re: [PATCH 0/2] Initial support for AVX512FP16
Date: Wed, 7 Jul 2021 09:24:22 +0800 [thread overview]
Message-ID: <CAMZc-bx-8g5TA+P5fdXdNitKoa1s-yjXpwDSqPtWtgvYpEGvmA@mail.gmail.com> (raw)
In-Reply-To: <alpine.DEB.2.22.394.2107061758340.627411@digraph.polyomino.org.uk>
On Wed, Jul 7, 2021 at 2:11 AM Joseph Myers <joseph@codesourcery.com> wrote:
>
> On Tue, 6 Jul 2021, Hongtao Liu via Gcc-patches wrote:
>
> > There may be inconsistent behavior between soft-fp and avx512fp16
> > instructions if we emulate _Float16 w/ float .
> > i.e
> > 1) for a + b - c where b and c are variables with the same big value
> > and a + b is NAN at _Float16 and real value at float, avx512fp16
> > instruction will raise an exception but soft-fp won't(unless it's
> > rounded after every operation.)
>
> There are at least two variants of emulation using float:
>
> (a) Using the excess precision support, as on AArch64, which means the C
> front end converts the _Float16 operations to float ones, with explicit
> narrowing on assignment (and conversion as if by assignment - argument
> passing and return, casts, etc.). Excess precision indeed involves
> different semantics compared to doing each operation directly in the range
> and precision of _Float16.
>
Yes, set excess_precision_type to FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
could round after each operation.
> (b) Letting the expand/optabs code generate operations in a wider mode.
> My understanding is that the result should get converted back to the
> narrower mode after each operation (by the expand/optabs code /
> convert_move called by it generating such a conversion), meaning (for
> basic arithmetic operations) that the semantics end up the same as if the
> operation had been done directly on _Float16 (but with more truncation
> operations occurring than would be the case with excess precision support
> used).
Yes, just w/ different behavior related to exceptions..
>
> > 2) a / b where b is denormal value and AVX512FP16 won't flush it to
> > zero even w/ -Ofast, but when it's extended to float and using divss,
> > it will be flushed to zero and raise an exception when compiling w/
> > Ofast
>
> I don't think that's a concern, flush to zero is well outside the scope of
> standards defining _Float16 semantics.
Ok.
>
> > So the key point is that the soft-fp and avx512fp16 instructions may
> > do not behave the same on the exception, is this acceptable?
>
> As far as I understand it, all cases within the standards will behave as
> expected for exceptions, whether pure software floating-point is used,
> pure hardware _Float16 arithmetic or one of the forms of emulation listed
> above. (Where "as expected" itself depends on the value of
> FLT_EVAL_METHOD, i.e. whether excess precision is used for _Float16.)
> Flush to zero and trapping exceptions are outside the scope of the
> standards. Since trapping exceptions is outside the scope of the
> standards, so is anything that distinguishes whether an arithmetic
> operation raises the same exception more than once or the order in which
> it raises different exceptions (e.g. the possibility of "inexact" being
> raised more than once, both by arithmetic on float and by narrowing from
> float to _Float16).
>
Set excess_precision_type to FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 to
round after each operation could keep semantics right.
And I'll document the behavior difference between soft-fp and
AVX512FP16 instruction for exceptions.
> --
> Joseph S. Myers
> joseph@codesourcery.com
--
BR,
Hongtao
next prev parent reply other threads:[~2021-07-07 1:19 UTC|newest]
Thread overview: 138+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20210701054808.39000-1-hongtao.liu@intel.com>
2021-07-01 5:55 ` Hongtao Liu
2021-07-01 20:46 ` Joseph Myers
2021-07-06 8:53 ` Hongtao Liu
[not found] ` <20210701054808.39000-3-hongtao.liu@intel.com>
2021-07-01 5:55 ` [PATCH 2/2] AVX512FP16: Add HFmode support in libgcc Hongtao Liu
[not found] ` <20210701054808.39000-2-hongtao.liu@intel.com>
2021-07-01 5:55 ` [PATCH 1/2] AVX512FP16: Initial support for _Float16 type and AVX512FP16 feature Hongtao Liu
2021-07-01 11:10 ` [PATCH 0/2] Initial support for AVX512FP16 Uros Bizjak
2021-07-01 12:39 ` H.J. Lu
2021-07-01 12:58 ` Richard Biener
2021-07-01 13:03 ` Jakub Jelinek
2021-07-06 8:51 ` Hongtao Liu
2021-07-06 10:14 ` Richard Biener
2021-07-06 12:11 ` H.J. Lu
2021-07-06 18:20 ` Joseph Myers
2021-07-06 18:18 ` Joseph Myers
2021-07-06 18:11 ` Joseph Myers
2021-07-07 1:24 ` Hongtao Liu [this message]
2021-07-14 7:50 ` Hongtao Liu
2021-07-14 15:32 ` [llvm-dev] " Craig Topper
2021-07-15 2:07 ` Wang, Pengfei
2021-07-15 6:34 ` Hongtao Liu
2021-07-15 6:57 ` Wang, Pengfei
2021-07-15 7:49 ` Hongtao Liu
2021-07-21 7:43 ` [PATCH V2 00/10] " liuhongt
2021-07-21 7:43 ` [PATCH 01/10] Update hf soft-fp from glibc liuhongt
2021-07-21 7:43 ` [PATCH 02/10] [i386] Enable _Float16 type for TARGET_SSE2 and above liuhongt
2021-07-21 10:35 ` Uros Bizjak
2021-07-22 5:21 ` Hongtao Liu
2021-07-22 11:56 ` Richard Biener
2021-07-28 21:56 ` Joseph Myers
2021-07-29 4:53 ` Hongtao Liu
2021-07-29 5:34 ` Hongtao Liu
2021-07-29 21:30 ` Joseph Myers
2021-08-02 5:23 ` Hongtao Liu
2021-08-02 6:31 ` [PATCH V3 0/6] Initial support for AVX512FP16 liuhongt
2021-08-02 6:31 ` [PATCH 1/6] Update hf soft-fp from glibc liuhongt
2021-08-02 6:31 ` [PATCH 2/6] [i386] Enable _Float16 type for TARGET_SSE2 and above liuhongt
2021-08-04 2:45 ` Hongtao Liu
2021-08-04 11:28 ` Richard Biener
2021-08-05 7:31 ` Hongtao Liu
2021-08-05 7:39 ` Hongtao Liu
2021-08-05 9:24 ` Richard Biener
2021-08-05 9:49 ` Hongtao Liu
2021-08-05 10:14 ` Richard Biener
2021-08-06 3:32 ` [PATCH] Make sure we're playing with integral modes before call extract_integral_bit_field liuhongt
2021-08-06 3:44 ` Andrew Pinski
2021-08-06 4:59 ` Hongtao Liu
2021-08-06 5:52 ` Hongtao Liu
2021-08-06 6:59 ` Richard Biener
2021-08-06 6:57 ` Richard Biener
2021-08-06 9:05 ` Richard Sandiford
2021-08-06 11:27 ` Richard Biener
2021-08-09 8:34 ` Hongtao Liu
2021-08-17 1:52 ` Hongtao Liu
2021-08-24 9:40 ` Hongtao Liu
2021-08-24 9:44 ` Hongtao Liu
2021-08-24 11:38 ` Richard Biener
2021-08-26 1:17 ` Hongtao Liu
2021-08-25 23:16 ` Jeff Law
2021-08-26 2:05 ` Hongtao Liu
2021-08-26 7:11 ` Richard Biener
2021-08-26 9:06 ` Richard Sandiford
2021-08-26 10:14 ` Richard Biener
2021-08-26 10:50 ` Richard Sandiford
2021-08-26 11:09 ` Richard Biener
2021-08-27 4:56 ` Hongtao Liu
2021-08-30 19:09 ` Joseph Myers
2021-08-30 21:15 ` Jeff Law
2021-08-31 6:10 ` Richard Biener
2021-08-31 6:30 ` Hongtao Liu
2021-08-31 6:48 ` Hongtao Liu
2021-08-31 11:16 ` Richard Biener
2021-08-31 11:17 ` [PATCH 0/2] Get rid of all float-int special cases in validate_subreg liuhongt
2021-08-31 11:17 ` [PATCH 1/2] Revert "Make sure we're playing with integral modes before call extract_integral_bit_field." liuhongt
2021-08-31 11:17 ` [PATCH 2/2] Get rid of all float-int special cases in validate_subreg liuhongt
2021-08-31 11:57 ` Richard Biener
2021-09-02 17:55 ` Segher Boessenkool
2021-09-03 15:05 ` Andreas Schwab
2021-09-07 23:19 ` Segher Boessenkool
2021-09-08 0:55 ` Hongtao Liu
2021-09-03 12:42 ` [PATCH 2/6] [i386] Enable _Float16 type for TARGET_SSE2 and above Jakub Jelinek
2021-09-06 2:05 ` Hongtao Liu
2021-09-06 12:13 ` Jakub Jelinek
2021-09-07 1:52 ` Hongtao Liu
2021-09-07 7:17 ` Jakub Jelinek
2021-09-07 10:08 ` Hongtao Liu
2021-09-07 10:10 ` Jakub Jelinek
2021-08-02 6:31 ` [PATCH 3/6] [i386] libgcc: Enable hfmode soft-sf/df/xf/tf extensions and truncations liuhongt
2021-08-02 6:31 ` [PATCH 4/6] Support -fexcess-precision=16 which will enable FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 when backend supports _Float16 liuhongt
2021-08-02 19:34 ` Joseph Myers
2021-08-03 2:44 ` Hongtao Liu
2021-08-06 6:06 ` Hongtao Liu
2021-08-17 1:53 ` Hongtao Liu
2021-08-24 9:39 ` Hongtao Liu
2021-09-02 6:06 ` Hongtao Liu
2021-08-02 6:39 ` [PATCH 6/6] AVX512FP16: Support vector init/broadcast/set/extract for FP16 liuhongt
2021-08-02 6:44 ` [PATCH 5/6] AVX512FP16: Initial support for AVX512FP16 feature and scalar _Float16 instructions liuhongt
2021-08-04 2:40 ` Hongtao Liu
2021-08-04 9:55 ` Uros Bizjak
2021-09-02 6:06 ` [PATCH V3 0/6] Initial support for AVX512FP16 Hongtao Liu
2021-09-02 11:30 ` Iain Sandoe
2021-09-02 15:18 ` Hongtao Liu
2021-09-02 16:44 ` Iain Sandoe
2021-09-02 20:03 ` Joseph Myers
2021-09-03 7:51 ` Iain Sandoe
2021-09-03 15:33 ` Iain Sandoe
2021-09-21 20:11 ` Joseph Myers
2021-09-21 20:25 ` Iain Sandoe
2021-09-22 7:08 ` Iain Sandoe
2021-09-22 19:50 ` Joseph Myers
2021-09-02 15:30 ` H.J. Lu
2021-09-02 15:50 ` Hongtao Liu
2021-09-02 19:45 ` Joseph Myers
2021-07-21 7:43 ` [PATCH 03/10] [i386] libgcc: Enable hfmode soft-sf/df/xf/tf extensions and truncations liuhongt
2021-07-21 10:51 ` Uros Bizjak
2021-07-22 12:14 ` Richard Biener
2021-07-27 5:32 ` Hongtao Liu
2021-07-29 20:57 ` Joseph Myers
2021-08-02 5:10 ` Hongtao Liu
2021-07-21 7:43 ` [PATCH 04/10] AVX512FP16: Initial support for AVX512FP16 feature and scalar _Float16 instructions liuhongt
2021-07-22 8:49 ` Uros Bizjak
2021-07-27 7:31 ` Hongtao Liu
2021-07-21 7:43 ` [PATCH 05/10] AVX512FP16: Support vector init/broadcast/set/extract for FP16 liuhongt
2021-07-22 5:24 ` Hongtao Liu
2021-07-21 7:43 ` [PATCH 06/10] AVX512FP16: Add testcase for vector init and broadcast intrinsics liuhongt
2021-07-21 7:43 ` [PATCH 07/10] AVX512FP16: Add tests for vector passing in variable arguments liuhongt
2021-07-21 7:43 ` [PATCH 08/10] AVX512FP16: Add ABI tests for xmm liuhongt
2021-07-21 7:43 ` [PATCH 09/10] AVX512FP16: Add ABI test for ymm liuhongt
2021-07-21 7:43 ` [PATCH 10/10] AVX512FP16: Add abi test for zmm liuhongt
2021-09-08 2:54 ` [PATCH V2 00/10] Initial support for AVX512FP16 Hongtao Liu
2021-09-08 3:02 ` Hongtao Liu
2021-07-01 12:58 ` [PATCH 0/2] " Uros Bizjak
2021-07-01 21:40 ` Joseph Myers
2021-07-02 6:30 ` Hongtao Liu
2021-07-02 8:03 ` Uros Bizjak
2021-07-02 8:19 ` Richard Biener
2021-07-03 14:44 ` Hongtao Liu
2021-07-05 1:25 ` Hongtao Liu
2021-07-05 11:02 ` Richard Biener
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=CAMZc-bx-8g5TA+P5fdXdNitKoa1s-yjXpwDSqPtWtgvYpEGvmA@mail.gmail.com \
--to=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hongtao.liu@intel.com \
--cc=jakub@redhat.com \
--cc=joseph@codesourcery.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).