[PATCH v4 2/2] x86_64: roundeven with sse4.1 support
H.J. Lu
hjl.tools@gmail.com
Thu May 28 12:05:32 GMT 2020
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.
More information about the Libc-alpha
mailing list