diff mbox series

x86_64: Add sinh with FMA

Message ID 20230814151634.2860117-1-hjl.tools@gmail.com
State New
Headers show
Series x86_64: Add sinh with FMA | expand

Commit Message

H.J. Lu Aug. 14, 2023, 3:16 p.m. UTC 
On Skylake, it improves sinh bench performance by:

        Before       After     Improvement
max     64.06        60.416       7%
min     10.874       9.674        11%
mean    14.0546      12.2018      13%

NB: Add

extern long double __expm1l (long double);
extern long double __expm1f128 (long double);

for __typeof (__expm1l) and __typeof (__expm1f128) when __expm1 is
defined since __expm1 may be expanded in their declarations which
causes the build failure.
---
 sysdeps/ieee754/dbl-64/e_sinh.c           |  7 +++++
 sysdeps/x86_64/fpu/multiarch/Makefile     |  2 ++
 sysdeps/x86_64/fpu/multiarch/e_sinh-fma.c | 12 +++++++++
 sysdeps/x86_64/fpu/multiarch/e_sinh.c     | 32 +++++++++++++++++++++++
 4 files changed, 53 insertions(+)
 create mode 100644 sysdeps/x86_64/fpu/multiarch/e_sinh-fma.c
 create mode 100644 sysdeps/x86_64/fpu/multiarch/e_sinh.c

Comments

Andreas Schwab Aug. 14, 2023, 3:50 p.m. UTC | #1 
On Aug 14 2023, H.J. Lu via Libc-alpha wrote:

> +/* NB: __expm1 may be expanded to __expm1_fma in the following
> +   prototypes.  */

Did you mean __expm1l?

> +extern long double __expm1l (long double);
> +extern long double __expm1f128 (long double);
H.J. Lu Aug. 14, 2023, 3:57 p.m. UTC | #2 
On Mon, Aug 14, 2023 at 8:50 AM Andreas Schwab <schwab@linux-m68k.org> wrote:
>
> On Aug 14 2023, H.J. Lu via Libc-alpha wrote:
>
> > +/* NB: __expm1 may be expanded to __expm1_fma in the following
> > +   prototypes.  */
>
> Did you mean __expm1l?
>
> > +extern long double __expm1l (long double);
> > +extern long double __expm1f128 (long double);

It looks like

extern long double expm1l (long double __x) __attribute__
((__nothrow__ )); extern long double __expm1_sse2l (long double __x)
__attribute__ ((__nothrow__ ));
 extern _Float128 expm1f128 (_Float128 __x) __attribute__
((__nothrow__ )); extern _Float128 __expm1_sse2f128 (_Float128 __x)
__attribute__ ((__nothrow__ ));

where __expm1 is expanded to __expm1_sse2.
Andreas Schwab Aug. 14, 2023, 4:06 p.m. UTC | #3 
On Aug 14 2023, H.J. Lu wrote:

> On Mon, Aug 14, 2023 at 8:50 AM Andreas Schwab <schwab@linux-m68k.org> wrote:
>>
>> On Aug 14 2023, H.J. Lu via Libc-alpha wrote:
>>
>> > +/* NB: __expm1 may be expanded to __expm1_fma in the following
>> > +   prototypes.  */
>>
>> Did you mean __expm1l?
>>
>> > +extern long double __expm1l (long double);
>> > +extern long double __expm1f128 (long double);
>
> It looks like
>
> extern long double expm1l (long double __x) __attribute__
> ((__nothrow__ )); extern long double __expm1_sse2l (long double __x)
> __attribute__ ((__nothrow__ ));
>  extern _Float128 expm1f128 (_Float128 __x) __attribute__
> ((__nothrow__ )); extern _Float128 __expm1_sse2f128 (_Float128 __x)
> __attribute__ ((__nothrow__ ));
>
> where __expm1 is expanded to __expm1_sse2.

Ok, expanded is confusing here since this is a term associated with the
preprocessor, but __expm1 doesn't occur as an identifier.
Paul Zimmermann Aug. 16, 2023, 2:45 p.m. UTC | #4 
Hi H.J.,

I get no improvement in accuracy. The largest known ulp error is still
the same:

sinh 0 -1 -0x1.633c654fee2bap+9 [2] [1.93] 1.92222 1.922214006544865

Maybe the FMA is not used for large inputs?

Best regards,
Paul
H.J. Lu Aug. 16, 2023, 5:23 p.m. UTC | #5 
On Wed, Aug 16, 2023 at 7:45 AM Paul Zimmermann
<Paul.Zimmermann@inria.fr> wrote:
>
>        Hi H.J.,
>
> I get no improvement in accuracy. The largest known ulp error is still
> the same:
>
> sinh 0 -1 -0x1.633c654fee2bap+9 [2] [1.93] 1.92222 1.922214006544865
>
> Maybe the FMA is not used for large inputs?
>

There is only one FMA insn in __ieee754_sinh_fma:

114: c4 e2 e9 9b 05 00 00 00 00 vfmsub132sd 0x0(%rip),%xmm2,%xmm0

Since it calls __ieee754_exp and __expm1, with

commit 1b214630ce6f7e0099b8b6f87246246739b079cf (master)
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Fri Aug 11 08:04:08 2023 -0700

    x86_64: Add expm1 with FMA

    On Skylake, it improves expm1 bench performance by:

            Before       After     Improvement
    max     70.204       68.054       3%
    min     20.709       16.2         22%
    mean    22.1221      16.7367      24%

There is very little improvement with a single FMA insn.

I will drop this patch.

Thanks.
Florian Weimer Aug. 17, 2023, 7:15 p.m. UTC | #6 
* H. J. Lu via Libc-alpha:

> On Wed, Aug 16, 2023 at 7:45 AM Paul Zimmermann
> <Paul.Zimmermann@inria.fr> wrote:
>>
>>        Hi H.J.,
>>
>> I get no improvement in accuracy. The largest known ulp error is still
>> the same:
>>
>> sinh 0 -1 -0x1.633c654fee2bap+9 [2] [1.93] 1.92222 1.922214006544865
>>
>> Maybe the FMA is not used for large inputs?
>>
>
> There is only one FMA insn in __ieee754_sinh_fma:
>
> 114: c4 e2 e9 9b 05 00 00 00 00 vfmsub132sd 0x0(%rip),%xmm2,%xmm0
>
> Since it calls __ieee754_exp and __expm1, with
>
> commit 1b214630ce6f7e0099b8b6f87246246739b079cf (master)
> Author: H.J. Lu <hjl.tools@gmail.com>
> Date:   Fri Aug 11 08:04:08 2023 -0700
>
>     x86_64: Add expm1 with FMA
>
>     On Skylake, it improves expm1 bench performance by:
>
>             Before       After     Improvement
>     max     70.204       68.054       3%
>     min     20.709       16.2         22%
>     mean    22.1221      16.7367      24%
>
> There is very little improvement with a single FMA insn.

If you specialize the implementation for FMA, could you avoid the
indirect call to __ieee754_exp and __expm1?

Thanks,
Florian
H.J. Lu Aug. 17, 2023, 7:21 p.m. UTC | #7 
On Thu, Aug 17, 2023 at 12:15 PM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> > On Wed, Aug 16, 2023 at 7:45 AM Paul Zimmermann
> > <Paul.Zimmermann@inria.fr> wrote:
> >>
> >>        Hi H.J.,
> >>
> >> I get no improvement in accuracy. The largest known ulp error is still
> >> the same:
> >>
> >> sinh 0 -1 -0x1.633c654fee2bap+9 [2] [1.93] 1.92222 1.922214006544865
> >>
> >> Maybe the FMA is not used for large inputs?
> >>
> >
> > There is only one FMA insn in __ieee754_sinh_fma:
> >
> > 114: c4 e2 e9 9b 05 00 00 00 00 vfmsub132sd 0x0(%rip),%xmm2,%xmm0
> >
> > Since it calls __ieee754_exp and __expm1, with
> >
> > commit 1b214630ce6f7e0099b8b6f87246246739b079cf (master)
> > Author: H.J. Lu <hjl.tools@gmail.com>
> > Date:   Fri Aug 11 08:04:08 2023 -0700
> >
> >     x86_64: Add expm1 with FMA
> >
> >     On Skylake, it improves expm1 bench performance by:
> >
> >             Before       After     Improvement
> >     max     70.204       68.054       3%
> >     min     20.709       16.2         22%
> >     mean    22.1221      16.7367      24%
> >
> > There is very little improvement with a single FMA insn.
>
> If you specialize the implementation for FMA, could you avoid the
> indirect call to __ieee754_exp and __expm1?
>

That is true.  Performance differences are

1. Non-FMA

  "sinh": {
   "": {
    "duration": 1.60856e+09,
    "iterations": 1.248e+08,
    "max": 173.747,
    "min": 10.706,
    "mean": 12.8891
   }
  }%

2. FMA

  "sinh": {
   "": {
    "duration": 1.61017e+09,
    "iterations": 1.275e+08,
    "max": 100.218,
    "min": 9.678,
    "mean": 12.6288
   }
  }%

FMA is a little faster.  Should I repost it with updated performance
numbers?
Florian Weimer Aug. 20, 2023, 2:28 p.m. UTC | #8 
* H. J. Lu via Libc-alpha:

> On Thu, Aug 17, 2023 at 12:15 PM Florian Weimer <fweimer@redhat.com> wrote:
>>
>> * H. J. Lu via Libc-alpha:
>> > There is only one FMA insn in __ieee754_sinh_fma:
>> >
>> > 114: c4 e2 e9 9b 05 00 00 00 00 vfmsub132sd 0x0(%rip),%xmm2,%xmm0
>> >
>> > Since it calls __ieee754_exp and __expm1, with
>> >
>> > commit 1b214630ce6f7e0099b8b6f87246246739b079cf (master)
>> > Author: H.J. Lu <hjl.tools@gmail.com>
>> > Date:   Fri Aug 11 08:04:08 2023 -0700
>> >
>> >     x86_64: Add expm1 with FMA
>> >
>> >     On Skylake, it improves expm1 bench performance by:
>> >
>> >             Before       After     Improvement
>> >     max     70.204       68.054       3%
>> >     min     20.709       16.2         22%
>> >     mean    22.1221      16.7367      24%
>> >
>> > There is very little improvement with a single FMA insn.
>>
>> If you specialize the implementation for FMA, could you avoid the
>> indirect call to __ieee754_exp and __expm1?
>>
>
> That is true.  Performance differences are
>
> 1. Non-FMA
>
>   "sinh": {
>    "": {
>     "duration": 1.60856e+09,
>     "iterations": 1.248e+08,
>     "max": 173.747,
>     "min": 10.706,
>     "mean": 12.8891
>    }
>   }%
>
> 2. FMA
>
>   "sinh": {
>    "": {
>     "duration": 1.61017e+09,
>     "iterations": 1.275e+08,
>     "max": 100.218,
>     "min": 9.678,
>     "mean": 12.6288
>    }
>   }%
>
> FMA is a little faster.  Should I repost it with updated performance
> numbers?

So the baseline is the before column from the original patch?

> On Skylake, it improves sinh bench performance by:
> 
>         Before       After     Improvement
> max     64.06        60.416       7%
> min     10.874       9.674        11%
> mean    14.0546      12.2018      13%

So we go from 14.0546 to 12.6288 for the mean case?
H.J. Lu Aug. 20, 2023, 2:29 p.m. UTC | #9 
On Sun, Aug 20, 2023 at 7:28 AM Florian Weimer <fw@deneb.enyo.de> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> > On Thu, Aug 17, 2023 at 12:15 PM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * H. J. Lu via Libc-alpha:
> >> > There is only one FMA insn in __ieee754_sinh_fma:
> >> >
> >> > 114: c4 e2 e9 9b 05 00 00 00 00 vfmsub132sd 0x0(%rip),%xmm2,%xmm0
> >> >
> >> > Since it calls __ieee754_exp and __expm1, with
> >> >
> >> > commit 1b214630ce6f7e0099b8b6f87246246739b079cf (master)
> >> > Author: H.J. Lu <hjl.tools@gmail.com>
> >> > Date:   Fri Aug 11 08:04:08 2023 -0700
> >> >
> >> >     x86_64: Add expm1 with FMA
> >> >
> >> >     On Skylake, it improves expm1 bench performance by:
> >> >
> >> >             Before       After     Improvement
> >> >     max     70.204       68.054       3%
> >> >     min     20.709       16.2         22%
> >> >     mean    22.1221      16.7367      24%
> >> >
> >> > There is very little improvement with a single FMA insn.
> >>
> >> If you specialize the implementation for FMA, could you avoid the
> >> indirect call to __ieee754_exp and __expm1?
> >>
> >
> > That is true.  Performance differences are
> >
> > 1. Non-FMA
> >
> >   "sinh": {
> >    "": {
> >     "duration": 1.60856e+09,
> >     "iterations": 1.248e+08,
> >     "max": 173.747,
> >     "min": 10.706,
> >     "mean": 12.8891
> >    }
> >   }%
> >
> > 2. FMA
> >
> >   "sinh": {
> >    "": {
> >     "duration": 1.61017e+09,
> >     "iterations": 1.275e+08,
> >     "max": 100.218,
> >     "min": 9.678,
> >     "mean": 12.6288
> >    }
> >   }%
> >
> > FMA is a little faster.  Should I repost it with updated performance
> > numbers?
>
> So the baseline is the before column from the original patch?
>
> > On Skylake, it improves sinh bench performance by:
> >
> >         Before       After     Improvement
> > max     64.06        60.416       7%
> > min     10.874       9.674        11%
> > mean    14.0546      12.2018      13%
>
> So we go from 14.0546 to 12.6288 for the mean case?

No.  It is from 12.8891 to 12.6288.
diff mbox series

Patch

diff --git a/sysdeps/ieee754/dbl-64/e_sinh.c b/sysdeps/ieee754/dbl-64/e_sinh.c
index b4b5857ddd..08611c94df 100644
--- a/sysdeps/ieee754/dbl-64/e_sinh.c
+++ b/sysdeps/ieee754/dbl-64/e_sinh.c
@@ -41,6 +41,11 @@  static char rcsid[] = "$NetBSD: e_sinh.c,v 1.7 1995/05/10 20:46:13 jtc Exp $";
 
 static const double one = 1.0, shuge = 1.0e307;
 
+#ifndef SECTION
+# define SECTION
+#endif
+
+SECTION
 double
 __ieee754_sinh (double x)
 {
@@ -90,4 +95,6 @@  __ieee754_sinh (double x)
   /* |x| > overflowthresold, sinh(x) overflow */
   return math_narrow_eval (x * shuge);
 }
+#ifndef __ieee754_sinh
 libm_alias_finite (__ieee754_sinh, __sinh)
+#endif
diff --git a/sysdeps/x86_64/fpu/multiarch/Makefile b/sysdeps/x86_64/fpu/multiarch/Makefile
index add339a876..15c44fc343 100644
--- a/sysdeps/x86_64/fpu/multiarch/Makefile
+++ b/sysdeps/x86_64/fpu/multiarch/Makefile
@@ -36,6 +36,7 @@  libm-sysdep_routines += \
   e_log-fma \
   e_log2-fma \
   e_pow-fma \
+  e_sinh-fma \
   s_atan-fma \
   s_expm1-fma \
   s_sin-fma \
@@ -49,6 +50,7 @@  CFLAGS-e_exp-fma.c = -mfma -mavx2
 CFLAGS-e_log-fma.c = -mfma -mavx2
 CFLAGS-e_log2-fma.c = -mfma -mavx2
 CFLAGS-e_pow-fma.c = -mfma -mavx2
+CFLAGS-e_sinh-fma.c = -mfma -mavx2 -march=x86-64-v3
 CFLAGS-s_atan-fma.c = -mfma -mavx2
 CFLAGS-s_expm1-fma.c = -mfma -mavx2
 CFLAGS-s_sin-fma.c = -mfma -mavx2
diff --git a/sysdeps/x86_64/fpu/multiarch/e_sinh-fma.c b/sysdeps/x86_64/fpu/multiarch/e_sinh-fma.c
new file mode 100644
index 0000000000..e0e1e39a7a
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_sinh-fma.c
@@ -0,0 +1,12 @@ 
+#define __ieee754_sinh __ieee754_sinh_fma
+#define __ieee754_exp __ieee754_exp_fma
+#define __expm1 __expm1_fma
+
+/* NB: __expm1 may be expanded to __expm1_fma in the following
+   prototypes.  */
+extern long double __expm1l (long double);
+extern long double __expm1f128 (long double);
+
+#define SECTION __attribute__ ((section (".text.fma")))
+
+#include <sysdeps/ieee754/dbl-64/e_sinh.c>
diff --git a/sysdeps/x86_64/fpu/multiarch/e_sinh.c b/sysdeps/x86_64/fpu/multiarch/e_sinh.c
new file mode 100644
index 0000000000..4c42adeb21
--- /dev/null
+++ b/sysdeps/x86_64/fpu/multiarch/e_sinh.c
@@ -0,0 +1,32 @@ 
+/* Multiple versions of sinh.
+   Copyright (C) 2023 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <libm-alias-double.h>
+#include <libm-alias-finite.h>
+
+extern double __redirect_ieee754_sinh (double);
+
+#define SYMBOL_NAME ieee754_sinh
+#include "ifunc-fma.h"
+
+libc_ifunc_redirected (__redirect_ieee754_sinh, __ieee754_sinh,
+		       IFUNC_SELECTOR ());
+libm_alias_finite (__ieee754_sinh, __sinh)
+
+#define __ieee754_sinh __ieee754_sinh_sse2
+#include <sysdeps/ieee754/dbl-64/e_sinh.c>