public inbox for libc-alpha@sourceware.org
 help / color / mirror / Atom feed
From: Noah Goldstein <goldstein.w.n@gmail.com>
To: Sunil K Pandey <skpgkp2@gmail.com>
Cc: GNU C Library <libc-alpha@sourceware.org>
Subject: Re: [PATCH] x86_64: Implement evex512 version of strchrnul, strchr and wcschr
Date: Wed, 21 Sep 2022 17:50:16 -0700	[thread overview]
Message-ID: <CAFUsyfJRaJR03MeEBSNAjRadKgQx0no7JM5_ksdrb8TcZ=Mfxg@mail.gmail.com> (raw)
In-Reply-To: <20220922001652.4039546-1-skpgkp2@gmail.com>

On Wed, Sep 21, 2022 at 5:17 PM Sunil K Pandey via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> This patch implements following evex512 version of string functions.
> evex512 version takes up to 30% less cycle as compared to evex,
> depending on length and alignment.

Please attach benchmark numbers.
>
> - strchrnul function using 512 bit vectors.
> - strchr function using 512 bit vectors.
> - wcschr function using 512 bit vectors.
>
> Code size data:
>
> strchrnul-evex.o        615 byte
> strchrnul-evex512.o     573 byte (-7%)
>
> strchr-evex.o           670 byte
> strchr-evex512.o        616 byte (-8%)
>
> wcschr-evex.o           678 byte
> wcschr-evex512.o        620 byte (-9%)
>
> Placeholder function, not used by any processor at the moment.
> ---
>  sysdeps/x86_64/multiarch/Makefile            |   3 +
>  sysdeps/x86_64/multiarch/ifunc-impl-list.c   |  12 +
>  sysdeps/x86_64/multiarch/strchr-evex-base.S  | 294 +++++++++++++++++++
>  sysdeps/x86_64/multiarch/strchr-evex512.S    |   7 +
>  sysdeps/x86_64/multiarch/strchrnul-evex512.S |   8 +
>  sysdeps/x86_64/multiarch/wcschr-evex512.S    |   8 +
>  6 files changed, 332 insertions(+)
>  create mode 100644 sysdeps/x86_64/multiarch/strchr-evex-base.S
>  create mode 100644 sysdeps/x86_64/multiarch/strchr-evex512.S
>  create mode 100644 sysdeps/x86_64/multiarch/strchrnul-evex512.S
>  create mode 100644 sysdeps/x86_64/multiarch/wcschr-evex512.S
>
> diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile
> index df4601c294..89b58fa557 100644
> --- a/sysdeps/x86_64/multiarch/Makefile
> +++ b/sysdeps/x86_64/multiarch/Makefile
> @@ -60,11 +60,13 @@ sysdep_routines += \
>    strchr-avx2 \
>    strchr-avx2-rtm \
>    strchr-evex \
> +  strchr-evex512 \
>    strchr-sse2 \
>    strchr-sse2-no-bsf \
>    strchrnul-avx2 \
>    strchrnul-avx2-rtm \
>    strchrnul-evex \
> +  strchrnul-evex512 \
>    strchrnul-sse2 \
>    strcmp-avx2 \
>    strcmp-avx2-rtm \
> @@ -129,6 +131,7 @@ sysdep_routines += \
>    wcschr-avx2 \
>    wcschr-avx2-rtm \
>    wcschr-evex \
> +  wcschr-evex512 \
>    wcschr-sse2 \
>    wcscmp-avx2 \
>    wcscmp-avx2-rtm \
> diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> index a71444eccb..bce1d15171 100644
> --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> @@ -518,6 +518,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>                                       && CPU_FEATURE_USABLE (AVX512BW)
>                                       && CPU_FEATURE_USABLE (BMI2)),
>                                      __strchr_evex)
> +             X86_IFUNC_IMPL_ADD_V4 (array, i, strchr,
> +                                    (CPU_FEATURE_USABLE (AVX512VL)
> +                                     && CPU_FEATURE_USABLE (AVX512BW)),
> +                                    __strchr_evex512)
>               X86_IFUNC_IMPL_ADD_V3 (array, i, strchr,
>                                      (CPU_FEATURE_USABLE (AVX2)
>                                       && CPU_FEATURE_USABLE (BMI2)),
> @@ -543,6 +547,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>                                       && CPU_FEATURE_USABLE (AVX512BW)
>                                       && CPU_FEATURE_USABLE (BMI2)),
>                                      __strchrnul_evex)
> +             X86_IFUNC_IMPL_ADD_V4 (array, i, strchrnul,
> +                                    (CPU_FEATURE_USABLE (AVX512VL)
> +                                     && CPU_FEATURE_USABLE (AVX512BW)),
> +                                    __strchrnul_evex512)
>               X86_IFUNC_IMPL_ADD_V3 (array, i, strchrnul,
>                                      (CPU_FEATURE_USABLE (AVX2)
>                                       && CPU_FEATURE_USABLE (BMI2)),
> @@ -753,6 +761,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
>                                       && CPU_FEATURE_USABLE (AVX512BW)
>                                       && CPU_FEATURE_USABLE (BMI2)),
>                                      __wcschr_evex)
> +             X86_IFUNC_IMPL_ADD_V4 (array, i, wcschr,
> +                                    (CPU_FEATURE_USABLE (AVX512VL)
> +                                     && CPU_FEATURE_USABLE (AVX512BW)),
> +                                    __wcschr_evex512)
>               X86_IFUNC_IMPL_ADD_V3 (array, i, wcschr,
>                                      (CPU_FEATURE_USABLE (AVX2)
>                                       && CPU_FEATURE_USABLE (BMI2)),
> diff --git a/sysdeps/x86_64/multiarch/strchr-evex-base.S b/sysdeps/x86_64/multiarch/strchr-evex-base.S
> new file mode 100644
> index 0000000000..919dafc8b6
> --- /dev/null
> +++ b/sysdeps/x86_64/multiarch/strchr-evex-base.S
> @@ -0,0 +1,294 @@
> +/* Placeholder function, not used by any processor at the moment.
> +   Copyright (C) 2022 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/>.  */
> +
> +/* UNUSED. Exists purely as reference implementation.  */
> +
> +#include <isa-level.h>
> +
> +#if ISA_SHOULD_BUILD (4)
> +
> +# include <sysdep.h>
> +
> +# ifdef USE_AS_WCSCHR
> +#  define CHAR_REG     esi
> +#  define CHAR_SIZE    4
> +#  define VPBROADCAST   vpbroadcastd
> +#  define VPCMP                vpcmpd
> +#  define VPMINU       vpminud
> +#  define VPTESTN      vptestnmd
> +# else
> +#  define CHAR_REG     sil
> +#  define CHAR_SIZE    1
> +#  define VPBROADCAST   vpbroadcastb
> +#  define VPCMP                vpcmpb
> +#  define VPMINU       vpminub
> +#  define VPTESTN      vptestnmb
> +# endif
> +
> +# define PAGE_SIZE     4096
> +# define CHAR_PER_VEC  (VEC_SIZE / CHAR_SIZE)
> +# define XMM1           xmm17
> +
> +# if VEC_SIZE == 64
> +#  define KMOV         kmovq
> +#  define KORTEST      kortestq
> +#  define RAX          rax
> +#  define RCX          rcx
> +#  define RDX          rdx
> +#  define SHR          shrq
> +#  define TEXTSUFFIX   evex512
> +#  define VMM0         zmm16
> +#  define VMM1         zmm17
> +#  define VMM2         zmm18
> +#  define VMM3         zmm19
> +#  define VMM4         zmm20
> +#  define VMM5         zmm21
> +#  define VMOVA                vmovdqa64
> +#  define VMOVU                vmovdqu64
> +
> +# elif VEC_SIZE == 32
> +/* Currently Unused.  */
> +#  define KMOV         kmovd
> +#  define KORTEST      kortestd
> +#  define RAX          eax
> +#  define RCX          ecx
> +#  define RDX          edx
> +#  define SHR          shrl
> +#  define TEXTSUFFIX   evex256
> +#  define VMM0         ymm16
> +#  define VMM1         ymm17
> +#  define VMM2         ymm18
> +#  define VMM3         ymm19
> +#  define VMM4         ymm20
> +#  define VMM5         ymm21
> +#  define VMOVA                vmovdqa32
> +#  define VMOVU                vmovdqu32
> +# endif
> +
> +       .section .text.TEXTSUFFIX, "ax", @progbits
> +/* Aligning entry point to 64 byte, provides better performance for
> +   one vector length string.  */
> +ENTRY_P2ALIGN (STRCHR, 6)
> +
> +       /* Broadcast CHAR to VMM0.  */
> +       VPBROADCAST %esi, %VMM0
> +       movl    %edi, %eax
> +       andl    $(PAGE_SIZE - 1), %eax
> +       cmpl    $(PAGE_SIZE - VEC_SIZE), %eax
> +       ja      L(page_cross)
> +
> +       /* Compare [w]char for null, mask bit will be set for match.  */
> +       VMOVU   (%rdi), %VMM1
> +
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +
> +       KMOV    %k0, %RAX
> +# ifndef USE_AS_STRCHRNUL
> +       test    %RAX, %RAX
> +       jz      L(align_more)
> +       bsf     %RAX, %RAX
> +# else
> +       /* For strchnul, using bsf, if string is less than 64 byte,
> +          entire logic will fit in 64 byte cache line and offset
> +          the perf gap as compared to evex version.  Even though
> +          using bsf as condition can save code size but it is not
> +          preferred for conditional jump for 2 reason.  1) It's
> +          latency is 3. 2) Unlike test, it can't be micro-fused
> +          with jump.  */
> +       bsf     %RAX, %RAX
> +       jz      L(align_more)
> +# endif
> +
> +# ifdef USE_AS_WCSCHR
> +       leaq    (%rdi, %rax, CHAR_SIZE), %rax
> +# else
> +       add     %rdi, %rax
> +# endif
> +# ifndef USE_AS_STRCHRNUL
> +       cmp     (%rax), %CHAR_REG
> +       jne     L(zero)
> +# endif
> +       ret
> +
> +# ifndef USE_AS_STRCHRNUL
> +L(zero):
> +       xorl    %eax, %eax
> +       ret
> +# endif
> +
> +L(ret_vec_x2):
> +       subq    $-VEC_SIZE, %rax
> +L(ret_vec_x1):
> +       bsf     %RCX, %RCX
> +# ifdef USE_AS_WCSCHR
> +       leaq    (%rax, %rcx, CHAR_SIZE), %rax
> +# else
> +       add     %rcx, %rax
> +# endif
> +
> +# ifndef USE_AS_STRCHRNUL
> +       cmp     (%rax), %CHAR_REG
> +       jne     L(zero)
> +# endif
> +       ret
> +
> +L(align_more):
> +       leaq    VEC_SIZE(%rdi), %rax
> +       /* Align rax to VEC_SIZE.  */
> +       andq    $-VEC_SIZE, %rax
> +
> +       /* Loop unroll 4 times for 4 vector loop.  */
> +       VMOVA   (%rax), %VMM1
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x1)
> +
> +       VMOVA   VEC_SIZE(%rax), %VMM1
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x2)
> +
> +       VMOVA   (VEC_SIZE * 2)(%rax), %VMM1
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x3)
> +
> +       VMOVA   (VEC_SIZE * 3)(%rax), %VMM1
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x4)
> +
> +       /* Align address to VEC_SIZE * 4 for loop.  */
> +       andq    $-(VEC_SIZE * 4), %rax
> +
> +       .p2align 4,,11
> +L(loop):
> +       /* VPMINU and VPCMP combination provide better performance as
> +          compared to alternative combinations.  */
> +       VMOVA   (VEC_SIZE * 4)(%rax), %VMM1
> +       VMOVA   (VEC_SIZE * 5)(%rax), %VMM2
> +       VMOVA   (VEC_SIZE * 6)(%rax), %VMM3
> +       VMOVA   (VEC_SIZE * 7)(%rax), %VMM4
> +
> +       vpxorq  %VMM1, %VMM0, %VMM5
> +       VPMINU  %VMM5, %VMM1, %VMM1
> +
> +       VPCMP   $4, %VMM0, %VMM2, %k1
> +       VPMINU  %VMM1, %VMM2, %VMM2{%k1}{z}
> +
> +       VPCMP   $4, %VMM0, %VMM3, %k2
> +       VPMINU  %VMM2, %VMM3, %VMM3{%k2}{z}
> +
> +       VPCMP   $4, %VMM0, %VMM4, %k3
> +       VPMINU  %VMM3, %VMM4, %VMM4{%k3}{z}
> +
> +       VPTESTN %VMM4, %VMM4, %k3
> +
> +       subq    $-(VEC_SIZE * 4), %rax
> +       KORTEST %k3, %k3
> +       jz      L(loop)
> +
> +       VPTESTN %VMM1, %VMM1, %k0
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x1)
> +
> +       VPTESTN %VMM2, %VMM2, %k0
> +       KMOV    %k0, %RCX
> +       /* At this point, if k1 is non zero, null char must be in the
> +          second vector.  */
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x2)
> +
> +       VPTESTN %VMM3, %VMM3, %k0
> +       KMOV    %k0, %RCX
> +       test    %RCX, %RCX
> +       jnz     L(ret_vec_x3)
> +       /* At this point null [w]char must be in the fourth vector so no
> +          need to check.  */
> +       KMOV    %k3, %RCX
> +
> +L(ret_vec_x4):
> +       bsf     %RCX, %RCX
> +       leaq    (VEC_SIZE * 3)(%rax, %rcx, CHAR_SIZE), %rax
> +# ifndef USE_AS_STRCHRNUL
> +       cmp     (%rax), %CHAR_REG
> +       jne     L(zero)
> +# endif
> +       ret
> +
> +L(ret_vec_x3):
> +       bsf     %RCX, %RCX
> +       leaq    (VEC_SIZE * 2)(%rax, %rcx, CHAR_SIZE), %rax
> +# ifndef USE_AS_STRCHRNUL
> +       cmp     (%rax), %CHAR_REG
> +       jne     L(zero)
> +# endif
> +       ret
> +
> +L(page_cross):
> +       movl    %eax, %ecx
> +# ifdef USE_AS_WCSCHR
> +       /* Calculate number of compare result bits to be skipped for
> +          wide string alignment adjustment.  */
> +       andl    $(VEC_SIZE - 1), %ecx
> +       sarl    $2, %ecx
> +# endif
> +       /* ecx contains number of w[char] to be skipped as a result
> +          of address alignment.  */
> +       xorq    %rdi, %rax
> +       VMOVA   (PAGE_SIZE - VEC_SIZE)(%rax), %VMM1
> +       vpxorq  %VMM1, %VMM0, %VMM2
> +       VPMINU  %VMM2, %VMM1, %VMM2
> +       VPTESTN %VMM2, %VMM2, %k0
> +       KMOV    %k0, %RAX
> +       /* Ignore number of character for alignment adjustment.  */
> +       SHR     %cl, %RAX
> +       jz      L(align_more)
> +
> +       bsf     %RAX, %RAX
> +# ifdef USE_AS_WCSCHR
> +       leaq    (%rdi, %rax, CHAR_SIZE), %rax
> +# else
> +       addq    %rdi, %rax
> +# endif
> +
> +# ifndef USE_AS_STRCHRNUL
> +       cmp     (%rax), %CHAR_REG
> +       jne     L(zero)
> +# endif
> +       ret
> +
> +END (STRCHR)
> +#endif
> diff --git a/sysdeps/x86_64/multiarch/strchr-evex512.S b/sysdeps/x86_64/multiarch/strchr-evex512.S
> new file mode 100644
> index 0000000000..4079bf387d
> --- /dev/null
> +++ b/sysdeps/x86_64/multiarch/strchr-evex512.S
> @@ -0,0 +1,7 @@
> +# ifndef STRCHR
> +#  define STRCHR       __strchr_evex512
> +# endif
> +
> +#define VEC_SIZE       64
> +
> +#include "strchr-evex-base.S"
> diff --git a/sysdeps/x86_64/multiarch/strchrnul-evex512.S b/sysdeps/x86_64/multiarch/strchrnul-evex512.S
> new file mode 100644
> index 0000000000..1be0b12f38
> --- /dev/null
> +++ b/sysdeps/x86_64/multiarch/strchrnul-evex512.S
> @@ -0,0 +1,8 @@
> +#ifndef STRCHRNUL
> +# define STRCHRNUL     __strchrnul_evex512
> +#endif
> +
> +#define STRCHR STRCHRNUL
> +#define USE_AS_STRCHRNUL 1
> +
> +#include "strchr-evex512.S"
> diff --git a/sysdeps/x86_64/multiarch/wcschr-evex512.S b/sysdeps/x86_64/multiarch/wcschr-evex512.S
> new file mode 100644
> index 0000000000..50c87ab1e5
> --- /dev/null
> +++ b/sysdeps/x86_64/multiarch/wcschr-evex512.S
> @@ -0,0 +1,8 @@
> +#ifndef WCSCHR
> +# define WCSCHR        __wcschr_evex512
> +#endif
> +
> +#define STRCHR WCSCHR
> +#define USE_AS_WCSCHR 1
> +
> +#include "strchr-evex512.S"
> --
> 2.36.1
>

  reply	other threads:[~2022-09-22  0:50 UTC|newest]

Thread overview: 10+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-09-22  0:16 Sunil K Pandey
2022-09-22  0:50 ` Noah Goldstein [this message]
2022-09-23  3:57   ` Sunil Pandey
2022-09-29  3:41     ` Sunil Pandey
2022-09-29  4:07       ` Noah Goldstein
2022-10-21 21:23         ` [PATCH v2] " Sunil K Pandey
2022-10-25 23:35           ` [PATCH v3] " Sunil K Pandey
2022-10-26  1:35             ` Noah Goldstein
2022-10-26  2:06               ` [PATCH v4] " Sunil K Pandey
2022-10-26  4:11                 ` Noah Goldstein

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='CAFUsyfJRaJR03MeEBSNAjRadKgQx0no7JM5_ksdrb8TcZ=Mfxg@mail.gmail.com' \
    --to=goldstein.w.n@gmail.com \
    --cc=libc-alpha@sourceware.org \
    --cc=skpgkp2@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).