[v4,2/2] x86_64: roundeven with sse4.1 support
Commit Message
This patch adds support for the sse4.1 hardware floating point
roundeven.
Here is a benchmark result on my AMD Ryzen 9 3900X system:
* benchmark result before this commit
| | roundeven | roundevenf |
|------------|---------------|--------------|
| duration | 3.77783e+09 | 3.77792e+09 |
| iterations | 3.75706e+08 | 3.80448e+08 |
| max | 158.498 | 88.539 |
| min | 6.802 | 7.676 |
| mean | 10.0553 | 9.93018 |
* benchmark result after this commit
| | roundeven | roundevenf |
|------------|---------------|---------------|
| duration | 3.77242e+09 | 3.77238e+09 |
| iterations | 5.18681e+08 | 5.2425e+08 |
| max | 127.338 | 172.102 |
| min | 7.03 | 7.03 |
| mean | 7.27311 | 7.19577 |
---
sysdeps/x86_64/fpu/multiarch/Makefile | 5 +--
sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c | 2 ++
.../x86_64/fpu/multiarch/s_roundeven-sse4_1.S | 26 ++++++++++++++++
sysdeps/x86_64/fpu/multiarch/s_roundeven.c | 31 +++++++++++++++++++
sysdeps/x86_64/fpu/multiarch/s_roundevenf-c.c | 3 ++
.../fpu/multiarch/s_roundevenf-sse4_1.S | 26 ++++++++++++++++
sysdeps/x86_64/fpu/multiarch/s_roundevenf.c | 31 +++++++++++++++++++
7 files changed, 122 insertions(+), 2 deletions(-)
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven-sse4_1.S
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven.c
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf-c.c
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf-sse4_1.S
create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf.c
Comments
On Sat, May 2, 2020 at 8:06 AM Shen-Ta Hsieh via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> This patch adds support for the sse4.1 hardware floating point
> roundeven.
Do you have FSF paper on file?
> Here is a benchmark result on my AMD Ryzen 9 3900X system:
Since we don't know or may not care SSE4 machines without AVX,
should we make it to AVX only?
> * benchmark result before this commit
> | | roundeven | roundevenf |
> |------------|---------------|--------------|
> | duration | 3.77783e+09 | 3.77792e+09 |
> | iterations | 3.75706e+08 | 3.80448e+08 |
> | max | 158.498 | 88.539 |
> | min | 6.802 | 7.676 |
> | mean | 10.0553 | 9.93018 |
>
> * benchmark result after this commit
> | | roundeven | roundevenf |
> |------------|---------------|---------------|
> | duration | 3.77242e+09 | 3.77238e+09 |
> | iterations | 5.18681e+08 | 5.2425e+08 |
> | max | 127.338 | 172.102 |
> | min | 7.03 | 7.03 |
> | mean | 7.27311 | 7.19577 |
> ---
> sysdeps/x86_64/fpu/multiarch/Makefile | 5 +--
> sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c | 2 ++
> .../x86_64/fpu/multiarch/s_roundeven-sse4_1.S | 26 ++++++++++++++++
> sysdeps/x86_64/fpu/multiarch/s_roundeven.c | 31 +++++++++++++++++++
> sysdeps/x86_64/fpu/multiarch/s_roundevenf-c.c | 3 ++
> .../fpu/multiarch/s_roundevenf-sse4_1.S | 26 ++++++++++++++++
> sysdeps/x86_64/fpu/multiarch/s_roundevenf.c | 31 +++++++++++++++++++
> 7 files changed, 122 insertions(+), 2 deletions(-)
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven-sse4_1.S
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundeven.c
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf-c.c
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf-sse4_1.S
> create mode 100644 sysdeps/x86_64/fpu/multiarch/s_roundevenf.c
>
> diff --git a/sysdeps/x86_64/fpu/multiarch/Makefile b/sysdeps/x86_64/fpu/multiarch/Makefile
> index 3836574f48..7e3a3f78cb 100644
> --- a/sysdeps/x86_64/fpu/multiarch/Makefile
> +++ b/sysdeps/x86_64/fpu/multiarch/Makefile
> @@ -1,11 +1,12 @@
> ifeq ($(subdir),math)
> libm-sysdep_routines += s_floor-c s_ceil-c s_floorf-c s_ceilf-c \
> s_rint-c s_rintf-c s_nearbyint-c s_nearbyintf-c \
> - s_trunc-c s_truncf-c
> + s_roundeven-c s_roundevenf-c s_trunc-c s_truncf-c
>
> libm-sysdep_routines += s_ceil-sse4_1 s_ceilf-sse4_1 s_floor-sse4_1 \
> s_floorf-sse4_1 s_nearbyint-sse4_1 \
> - s_nearbyintf-sse4_1 s_rint-sse4_1 s_rintf-sse4_1 \
> + s_nearbyintf-sse4_1 s_roundeven-sse4_1 \
> + s_roundevenf-sse4_1 s_rint-sse4_1 s_rintf-sse4_1 \
> s_trunc-sse4_1 s_truncf-sse4_1
>
> libm-sysdep_routines += e_exp-fma e_log-fma e_pow-fma s_atan-fma \
> diff --git a/sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c b/sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c
> new file mode 100644
> index 0000000000..c7be43cb22
> --- /dev/null
> +++ b/sysdeps/x86_64/fpu/multiarch/s_roundeven-c.c
> @@ -0,0 +1,2 @@
> +#define __roundeven __roundeven_c
> +#include <sysdeps/ieee754/dbl-64/s_roundeven.c>
> diff --git a/sysdeps/x86_64/fpu/multiarch/s_roundeven-sse4_1.S b/sysdeps/x86_64/fpu/multiarch/s_roundeven-sse4_1.S
> new file mode 100644
> index 0000000000..6db88a1649
> --- /dev/null
> +++ b/sysdeps/x86_64/fpu/multiarch/s_roundeven-sse4_1.S
> @@ -0,0 +1,26 @@
> +/* Round to nearest integer value, rounding halfway cases to even.
> + double version.
> + Copyright (C) 2019 Free Software Foundation, Inc.
Please replace all 2019 with 2020.
* H. J. Lu via Libc-alpha:
>> Here is a benchmark result on my AMD Ryzen 9 3900X system:
>
> Since we don't know or may not care SSE4 machines without AVX,
> should we make it to AVX only?
What about Goldmont/Tremont? Those are current CPUs which do not
support AVX, but I think they have sufficient SSE4 support levels for
this change.
Thanks,
Florian
On Thu, May 28, 2020 at 5:22 AM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu via Libc-alpha:
>
> >> Here is a benchmark result on my AMD Ryzen 9 3900X system:
> >
> > Since we don't know or may not care SSE4 machines without AVX,
> > should we make it to AVX only?
>
> What about Goldmont/Tremont? Those are current CPUs which do not
> support AVX, but I think they have sufficient SSE4 support levels for
> this change.
>
Good point. Lili, please collect glibc micro benchmark roundeven/roundevenf
data before and after:
https://sourceware.org/pipermail/libc-alpha/2020-May/113533.html
on Tremont.
> -----Original Message-----
> From: H.J. Lu <hjl.tools@gmail.com>
> Sent: Thursday, May 28, 2020 8:32 PM
> To: Florian Weimer <fweimer@redhat.com>; Cui, Lili <lili.cui@intel.com>
> Cc: H.J. Lu via Libc-alpha <libc-alpha@sourceware.org>; Shen-Ta Hsieh
> <ibmibmibm.tw@gmail.com>
> Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
>
> On Thu, May 28, 2020 at 5:22 AM Florian Weimer <fweimer@redhat.com<mailto:fweimer@redhat.com>>
> wrote:
> >
> > * H. J. Lu via Libc-alpha:
> >
> > >> Here is a benchmark result on my AMD Ryzen 9 3900X system:
> > >
> > > Since we don't know or may not care SSE4 machines without AVX,
> > > should we make it to AVX only?
> >
> > What about Goldmont/Tremont? Those are current CPUs which do not
> > support AVX, but I think they have sufficient SSE4 support levels for
> > this change.
> >
>
> Good point. Lili, please collect glibc micro benchmark roundeven/roundevenf
> data before and after:
>
> https://sourceware.org/pipermail/libc-alpha/2020-May/113533.html
>
> on Tremont.
>
> --
> H.J.
Hi H.J,
Result is here.
benchmark result before this commit on Tremont
[X]
benchmark result after this commit on Tremont
On Fri, May 29, 2020 at 1:48 AM Cui, Lili <lili.cui@intel.com> wrote:
>
>
> > -----Original Message-----
> > From: H.J. Lu <hjl.tools@gmail.com>
> > Sent: Thursday, May 28, 2020 8:32 PM
> > To: Florian Weimer <fweimer@redhat.com>; Cui, Lili <lili.cui@intel.com>
> > Cc: H.J. Lu via Libc-alpha <libc-alpha@sourceware.org>; Shen-Ta Hsieh
> > <ibmibmibm.tw@gmail.com>
> > Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
> >
> > On Thu, May 28, 2020 at 5:22 AM Florian Weimer <fweimer@redhat.com>
> > wrote:
> > >
> > > * H. J. Lu via Libc-alpha:
> > >
> > > >> Here is a benchmark result on my AMD Ryzen 9 3900X system:
> > > >
> > > > Since we don't know or may not care SSE4 machines without AVX,
> > > > should we make it to AVX only?
> > >
> > > What about Goldmont/Tremont? Those are current CPUs which do not
> > > support AVX, but I think they have sufficient SSE4 support levels for
> > > this change.
> > >
> >
> > Good point. Lili, please collect glibc micro benchmark
> roundeven/roundevenf
> > data before and after:
> >
> > https://sourceware.org/pipermail/libc-alpha/2020-May/113533.html
> >
> > on Tremont.
> >
> > --
> > H.J.
>
> Hi H.J,
>
> Result is here.
> benchmark result before this commit on Tremont
>
>
>
> benchmark result after this commit on Tremont
>
>
>
>
Hi Lili,
The results are empty.
From: H.J. Lu <hjl.tools@gmail.com>
Sent: Friday, May 29, 2020 7:30 PM
To: Cui, Lili <lili.cui@intel.com>
Cc: Florian Weimer <fweimer@redhat.com>; H.J. Lu via Libc-alpha <libc-alpha@sourceware.org>; Shen-Ta Hsieh <ibmibmibm.tw@gmail.com>
Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
On Fri, May 29, 2020 at 1:48 AM Cui, Lili <lili.cui@intel.com<mailto:lili.cui@intel.com>> wrote:
> -----Original Message-----
> From: H.J. Lu <hjl.tools@gmail.com<mailto:hjl.tools@gmail.com>>
> Sent: Thursday, May 28, 2020 8:32 PM
> To: Florian Weimer <fweimer@redhat.com<mailto:fweimer@redhat.com>>; Cui, Lili <lili.cui@intel.com<mailto:lili.cui@intel.com>>
> Cc: H.J. Lu via Libc-alpha <libc-alpha@sourceware.org<mailto:libc-alpha@sourceware.org>>; Shen-Ta Hsieh
> <ibmibmibm.tw@gmail.com<mailto:ibmibmibm.tw@gmail.com>>
> Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
>
> On Thu, May 28, 2020 at 5:22 AM Florian Weimer <fweimer@redhat.com<mailto:fweimer@redhat.com>>
> wrote:
> >
> > * H. J. Lu via Libc-alpha:
> >
> > >> Here is a benchmark result on my AMD Ryzen 9 3900X system:
> > >
> > > Since we don't know or may not care SSE4 machines without AVX,
> > > should we make it to AVX only?
> >
> > What about Goldmont/Tremont? Those are current CPUs which do not
> > support AVX, but I think they have sufficient SSE4 support levels for
> > this change.
> >
>
> Good point. Lili, please collect glibc micro benchmark roundeven/roundevenf
> data before and after:
>
> https://sourceware.org/pipermail/libc-alpha/2020-May/113533.html
>
> on Tremont.
>
> --
> H.J.
Hi H.J,
Result is here.
benchmark result before this commit on Tremont
benchmark result after this commit on Tremont
Hi Lili,
The results are empty.
--
H.J.
Hi H.J,
Sorry for that my format has some problems, data is here.
benchmark result before this commit on Tremont
"roundeven": "roundevenf":
"duration": 2.19422e+09, "duration": 2.19402e+09,
"iterations": 1.44514e+08, "iterations": 1.4184e+08,
"max": 43.258, "max": 53.07,
"min": 11.052, "min": 12.052,
"mean": 15.1835 "mean": 15.4683
benchmark result after this commit on Tremont
"roundeven": "roundevenf":
"duration": 2.19144e+09, "duration": 2.19218e+09,
"iterations": 2.17075e+08, "iterations": 1.97982e+08,
"max": 395.428, "max": 34.928,
"min": 10.044, "min": 11.02,
"mean": 10.0953 "mean": 11.0726
Thanks,
Lili.
On Sun, May 31, 2020 at 6:28 PM Cui, Lili <lili.cui@intel.com> wrote:
>
>
>
>
>
> From: H.J. Lu <hjl.tools@gmail.com>
> Sent: Friday, May 29, 2020 7:30 PM
> To: Cui, Lili <lili.cui@intel.com>
> Cc: Florian Weimer <fweimer@redhat.com>; H.J. Lu via Libc-alpha <libc-alpha@sourceware.org>; Shen-Ta Hsieh <ibmibmibm.tw@gmail.com>
> Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
>
>
>
> On Fri, May 29, 2020 at 1:48 AM Cui, Lili <lili.cui@intel.com> wrote:
>
>
>
>
>
> > -----Original Message-----
>
> > From: H.J. Lu <hjl.tools@gmail.com>
>
> > Sent: Thursday, May 28, 2020 8:32 PM
>
> > To: Florian Weimer <fweimer@redhat.com>; Cui, Lili <lili.cui@intel.com>
>
> > Cc: H.J. Lu via Libc-alpha <libc-alpha@sourceware.org>; Shen-Ta Hsieh
>
> > <ibmibmibm.tw@gmail.com>
>
> > Subject: Re: [PATCH v4 2/2] x86_64: roundeven with sse4.1 support
>
> >
>
> > On Thu, May 28, 2020 at 5:22 AM Florian Weimer <fweimer@redhat.com>
>
> > wrote:
>
> > >
>
> > > * H. J. Lu via Libc-alpha:
>
> > >
>
> > > >> Here is a benchmark result on my AMD Ryzen 9 3900X system:
>
> > > >
>
> > > > Since we don't know or may not care SSE4 machines without AVX,
>
> > > > should we make it to AVX only?
>
> > >
>
> > > What about Goldmont/Tremont? Those are current CPUs which do not
>
> > > support AVX, but I think they have sufficient SSE4 support levels for
>
> > > this change.
>
> > >
>
> >
>
> > Good point. Lili, please collect glibc micro benchmark roundeven/roundevenf
>
> > data before and after:
>
> >
>
> > https://sourceware.org/pipermail/libc-alpha/2020-May/113533.html
>
> >
>
> > on Tremont.
>
> >
>
> > --
>
> > H.J.
>
>
>
> Hi H.J,
>
>
>
> Result is here.
>
> benchmark result before this commit on Tremont
>
>
>
>
>
>
>
> benchmark result after this commit on Tremont
>
>
>
>
>
>
>
>
>
> Hi Lili,
>
>
>
> The results are empty.
>
>
>
> --
>
> H.J.
>
>
>
> Hi H.J,
>
>
>
> Sorry for that my format has some problems, data is here.
>
>
>
> benchmark result before this commit on Tremont
>
>
>
> "roundeven": "roundevenf":
>
> "duration": 2.19422e+09, "duration": 2.19402e+09,
>
> "iterations": 1.44514e+08, "iterations": 1.4184e+08,
>
> "max": 43.258, "max": 53.07,
>
> "min": 11.052, "min": 12.052,
>
> "mean": 15.1835 "mean": 15.4683
>
>
>
> benchmark result after this commit on Tremont
>
>
>
> "roundeven": "roundevenf":
>
> "duration": 2.19144e+09, "duration": 2.19218e+09,
>
> "iterations": 2.17075e+08, "iterations": 1.97982e+08,
>
> "max": 395.428, "max": 34.928,
>
> "min": 10.044, "min": 11.02,
>
> "mean": 10.0953 "mean": 11.0726
>
>
>
Looks good.
Thanks.
@@ -1,11 +1,12 @@
ifeq ($(subdir),math)
libm-sysdep_routines += s_floor-c s_ceil-c s_floorf-c s_ceilf-c \
s_rint-c s_rintf-c s_nearbyint-c s_nearbyintf-c \
- s_trunc-c s_truncf-c
+ s_roundeven-c s_roundevenf-c s_trunc-c s_truncf-c
libm-sysdep_routines += s_ceil-sse4_1 s_ceilf-sse4_1 s_floor-sse4_1 \
s_floorf-sse4_1 s_nearbyint-sse4_1 \
- s_nearbyintf-sse4_1 s_rint-sse4_1 s_rintf-sse4_1 \
+ s_nearbyintf-sse4_1 s_roundeven-sse4_1 \
+ s_roundevenf-sse4_1 s_rint-sse4_1 s_rintf-sse4_1 \
s_trunc-sse4_1 s_truncf-sse4_1
libm-sysdep_routines += e_exp-fma e_log-fma e_pow-fma s_atan-fma \
new file mode 100644
@@ -0,0 +1,2 @@
+#define __roundeven __roundeven_c
+#include <sysdeps/ieee754/dbl-64/s_roundeven.c>
new file mode 100644
@@ -0,0 +1,26 @@
+/* Round to nearest integer value, rounding halfway cases to even.
+ double version.
+ Copyright (C) 2019 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 <sysdep.h>
+
+ .section .text.sse4.1,"ax",@progbits
+ENTRY(__roundeven_sse41)
+ roundsd $8, %xmm0, %xmm0
+ ret
+END(__roundeven_sse41)
new file mode 100644
@@ -0,0 +1,31 @@
+/* Multiple versions of __roundeven.
+ Copyright (C) 2019 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>
+
+#define roundeven __redirect_roundeven
+#define __roundeven __redirect___roundeven
+#include <math.h>
+#undef roundeven
+#undef __roundeven
+
+#define SYMBOL_NAME roundeven
+#include "ifunc-sse4_1.h"
+
+libc_ifunc_redirected (__redirect_roundeven, __roundeven, IFUNC_SELECTOR ());
+libm_alias_double (__roundeven, roundeven)
new file mode 100644
@@ -0,0 +1,3 @@
+#undef __roundevenf
+#define __roundevenf __roundevenf_c
+#include <sysdeps/ieee754/flt-32/s_roundevenf.c>
new file mode 100644
@@ -0,0 +1,26 @@
+/* Round to nearest integer value, rounding halfway cases to even.
+ float version.
+ Copyright (C) 2019 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 <sysdep.h>
+
+ .section .text.sse4.1,"ax",@progbits
+ENTRY(__roundevenf_sse41)
+ roundss $8, %xmm0, %xmm0
+ ret
+END(__roundevenf_sse41)
new file mode 100644
@@ -0,0 +1,31 @@
+/* Multiple versions of __roundevenf.
+ Copyright (C) 2019 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-float.h>
+
+#define roundevenf __redirect_roundevenf
+#define __roundevenf __redirect___roundevenf
+#include <math.h>
+#undef roundevenf
+#undef __roundevenf
+
+#define SYMBOL_NAME roundevenf
+#include "ifunc-sse4_1.h"
+
+libc_ifunc_redirected (__redirect_roundevenf, __roundevenf, IFUNC_SELECTOR ());
+libm_alias_float (__roundeven, roundeven)