[PATCH] x86_64: Implement evex512 version of memchr, rawmemchr and wmemchr

Noah Goldstein goldstein.w.n@gmail.com
Thu Sep 29 04:07:15 GMT 2022


On Wed, Sep 28, 2022 at 8:43 PM Sunil Pandey <skpgkp2@gmail.com> wrote:
>
> On Thu, Sep 22, 2022 at 8:57 PM Sunil Pandey <skpgkp2@gmail.com> wrote:
> >
> > Microbenchmark data collected on: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
> >
> >
> > On Wed, Sep 21, 2022 at 5:50 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
> > >
> > > On Wed, Sep 21, 2022 at 5:27 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.
> > > >
> > > > - memchr function using 512 bit vectors.
> > > > - rawmemchr function using 512 bit vectors.
> > > > - wmemchr function using 512 bit vectors.
> > > >
> > > > Code size data:
> > > >
> > > > memchr-evex.o           762 byte
> > > > memchr-evex512.o        570 byte (-25%)
> > > >
> > > > rawmemchr-evex.o        461 byte
> > > > rawmemchr-evex512.o     413 byte (-10%)
> > > >
> > > > wmemchr-evex.o          794 byte
> > > > wmemchr-evex512.o       568 byte (-28%)
> > > >
> > > > Placeholder function, not used by any processor at the moment.
> > > > ---
> > > >  sysdeps/x86_64/multiarch/Makefile            |   3 +
> > > >  sysdeps/x86_64/multiarch/ifunc-impl-list.c   |  15 +
> > > >  sysdeps/x86_64/multiarch/memchr-evex-base.S  | 306 +++++++++++++++++++
> > > >  sysdeps/x86_64/multiarch/memchr-evex512.S    |   7 +
> > > >  sysdeps/x86_64/multiarch/rawmemchr-evex512.S |   7 +
> > > >  sysdeps/x86_64/multiarch/wmemchr-evex512.S   |   8 +
> > > >  6 files changed, 346 insertions(+)
> > > >  create mode 100644 sysdeps/x86_64/multiarch/memchr-evex-base.S
> > > >  create mode 100644 sysdeps/x86_64/multiarch/memchr-evex512.S
> > > >  create mode 100644 sysdeps/x86_64/multiarch/rawmemchr-evex512.S
> > > >  create mode 100644 sysdeps/x86_64/multiarch/wmemchr-evex512.S
> > > >
> > > > diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile
> > > > index df4601c294..e974b1ad97 100644
> > > > --- a/sysdeps/x86_64/multiarch/Makefile
> > > > +++ b/sysdeps/x86_64/multiarch/Makefile
> > > > @@ -4,6 +4,7 @@ sysdep_routines += \
> > > >    memchr-avx2 \
> > > >    memchr-avx2-rtm \
> > > >    memchr-evex \
> > > > +  memchr-evex512 \
> > > >    memchr-evex-rtm \
> > > >    memchr-sse2 \
> > > >    memcmp-avx2-movbe \
> > > > @@ -36,6 +37,7 @@ sysdep_routines += \
> > > >    rawmemchr-avx2 \
> > > >    rawmemchr-avx2-rtm \
> > > >    rawmemchr-evex \
> > > > +  rawmemchr-evex512 \
> > > >    rawmemchr-evex-rtm \
> > > >    rawmemchr-sse2 \
> > > >    stpcpy-avx2 \
> > > > @@ -156,6 +158,7 @@ sysdep_routines += \
> > > >    wmemchr-avx2 \
> > > >    wmemchr-avx2-rtm \
> > > >    wmemchr-evex \
> > > > +  wmemchr-evex512 \
> > > >    wmemchr-evex-rtm \
> > > >    wmemchr-sse2 \
> > > >    wmemcmp-avx2-movbe \
> > > > diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > > > index a71444eccb..17f770318d 100644
> > > > --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > > > +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
> > > > @@ -63,6 +63,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > >                                       && CPU_FEATURE_USABLE (BMI2)),
> > > >                                      __memchr_evex)
> > > > +             X86_IFUNC_IMPL_ADD_V4 (array, i, memchr,
> > > > +                                    (CPU_FEATURE_USABLE (AVX512VL)
> > > > +                                     && CPU_FEATURE_USABLE (AVX512BW)
> > > > +                                     && CPU_FEATURE_USABLE (BMI2)),
> > > > +                                    __memchr_evex512)
> > > >               X86_IFUNC_IMPL_ADD_V4 (array, i, memchr,
> > > >                                      (CPU_FEATURE_USABLE (AVX512VL)
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > > @@ -329,6 +334,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > >                                       && CPU_FEATURE_USABLE (BMI2)),
> > > >                                      __rawmemchr_evex)
> > > > +             X86_IFUNC_IMPL_ADD_V4 (array, i, rawmemchr,
> > > > +                                    (CPU_FEATURE_USABLE (AVX512VL)
> > > > +                                     && CPU_FEATURE_USABLE (AVX512BW)
> > > > +                                     && CPU_FEATURE_USABLE (BMI2)),
> > > > +                                    __rawmemchr_evex512)
> > > >               X86_IFUNC_IMPL_ADD_V4 (array, i, rawmemchr,
> > > >                                      (CPU_FEATURE_USABLE (AVX512VL)
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > > @@ -903,6 +913,11 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > >                                       && CPU_FEATURE_USABLE (BMI2)),
> > > >                                      __wmemchr_evex)
> > > > +             X86_IFUNC_IMPL_ADD_V4 (array, i, wmemchr,
> > > > +                                    (CPU_FEATURE_USABLE (AVX512VL)
> > > > +                                     && CPU_FEATURE_USABLE (AVX512BW)
> > > > +                                     && CPU_FEATURE_USABLE (BMI2)),
> > > > +                                    __wmemchr_evex512)
> > > >               X86_IFUNC_IMPL_ADD_V4 (array, i, wmemchr,
> > > >                                      (CPU_FEATURE_USABLE (AVX512VL)
> > > >                                       && CPU_FEATURE_USABLE (AVX512BW)
> > > > diff --git a/sysdeps/x86_64/multiarch/memchr-evex-base.S b/sysdeps/x86_64/multiarch/memchr-evex-base.S
> > > > new file mode 100644
> > > > index 0000000000..524f0809b5
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/multiarch/memchr-evex-base.S
> > > > @@ -0,0 +1,306 @@
> > > > +/* 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_WMEMCHR
> > > > +#  define CHAR_SIZE    4
> > > > +#  define VPBROADCAST   vpbroadcastd
> > > > +#  define VPCMP                vpcmpd
> > > > +# else
> > > > +#  define CHAR_SIZE    1
> > > > +#  define VPBROADCAST   vpbroadcastb
> > > > +#  define VPCMP                vpcmpb
> > > > +# endif
> > > > +
> > > > +# define PAGE_SIZE     4096
> > > > +# define CHAR_PER_VEC  (VEC_SIZE / CHAR_SIZE)
> > > > +# define XMM1           xmm17
> > > > +
> > > > +# if VEC_SIZE == 64
> > > > +#  define KMOV         kmovq
> > > > +#  define KOR          korq
> > > > +#  define KORTEST      kortestq
> > > > +#  define RAX          rax
> > > > +#  define RCX          rcx
> > > > +#  define SHR          shrq
> > > > +#  define SARX         sarxq
> > > > +#  define TEXTSUFFIX   evex512
> > > > +#  define VMM0         zmm16
> > > > +# elif VEC_SIZE == 32
> > > > +/* Currently Unused.  */
> > > > +#  define KMOV         kmovd
> > > > +#  define KOR          kord
> > > > +#  define KORTEST      kortestd
> > > > +#  define RAX          eax
> > > > +#  define RCX          ecx
> > > > +#  define SHR          shrl
> > > > +#  define SARX         sarxl
> > > > +#  define TEXTSUFFIX   evex256
> > > > +#  define VMM0         ymm16
> > > > +# endif
> > > > +
> > > > +       .section .text.TEXTSUFFIX, "ax", @progbits
> > > > +/* Aligning entry point to 64 byte, provides better performance for
> > > > +   one vector length string.  */
> > > > +ENTRY_P2ALIGN (MEMCHR, 6)
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       /* Check for zero length.  */
> > > > +       test    %RDX_LP, %RDX_LP
> > > > +       jz      L(zero)
> > > > +
> > > > +#  ifdef __ILP32__
> > > > +       /* Clear the upper 32 bits.  */
> > > > +       movl    %edx, %edx
> > > > +#  endif
> > > > +# endif
> > > > +
> > > > +       /* 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.  */
> > > > +       VPCMP   $0, (%rdi), %VMM0, %k0
> > > > +
> > > > +       KMOV    %k0, %RAX
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       bsf     %RAX, %RCX
> > > > +       jz      L(align_more)
> > > > +       xor     %eax, %eax
> > > > +#  ifdef USE_AS_WMEMCHR
> > > > +       leaq    (%rdi, %rcx, CHAR_SIZE), %rdi
> > > > +#  else
> > > > +       addq    %rcx, %rdi
> > > > +#  endif
> > > > +       cmp     %rcx, %rdx
> > > > +       cmova   %rdi, %rax
> > > > +# else
> > > > +       bsf     %RAX, %RAX
> > > > +       jz      L(align_more)
> > > > +       add     %rdi, %rax
> > > > +# endif
> > > > +       ret
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +L(zero):
> > > > +       xorl    %eax, %eax
> > > > +       ret
> > > > +# endif
> > > > +
> > > > +       .p2align 5,,5
> > > > +L(page_cross):
> > > > +       movq    %rdi, %rcx
> > > > +       andq    $-VEC_SIZE, %rcx
> > > > +
> > > > +       VPCMP   $0, (%rcx), %VMM0, %k0
> > > > +       KMOV    %k0, %RCX
> > > > +       SARX    %RAX, %RCX, %RAX
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       bsf     %RAX, %RCX
> > > > +       jz      L(align_more)
> > > > +       xor     %eax, %eax
> > > > +#  ifdef USE_AS_WMEMCHR
> > > > +       leaq    (%rdi, %rcx, CHAR_SIZE), %rdi
> > > > +#  else
> > > > +       addq    %rcx, %rdi
> > > > +#  endif
> > > > +       cmp     %rcx, %rdx
> > > > +       cmovae  %rdi, %rax
> > > > +
> > > > +# else
> > > > +       bsf     %rax, %rax
> > > > +       jz      L(align_more)
> > > > +       add     %rdi, %rax
> > > > +# endif
> > > > +       ret
> > > > +
> > > > +L(ret_vec_x2):
> > > > +       subq    $-VEC_SIZE, %rdi
> > > > +L(ret_vec_x1):
> > > > +       bsf     %RAX, %RAX
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       jz      L(zero)
> > > > +       cmp     %rax, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +# ifdef USE_AS_WMEMCHR
> > > > +       leaq    (%rdi, %rax, CHAR_SIZE), %rax
> > > > +# else
> > > > +       add     %rdi, %rax
> > > > +# endif
> > > > +       ret
> > > > +
> > > > +       .p2align 5,,10
> > > > +L(align_more):
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       xor     %eax, %eax
> > > > +       subq    %rdi, %rax
> > > > +# endif
> > > > +
> > > > +       subq    $-VEC_SIZE, %rdi
> > > > +       /* Align rdi to VEC_SIZE.  */
> > > > +       andq    $-VEC_SIZE, %rdi
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       addq    %rdi, %rax
> > > > +#  ifdef USE_AS_WMEMCHR
> > > > +       sarl    $2, %eax
> > > > +#  endif
> > > > +       subq    %rax, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       /* Loop unroll 4 times for 4 vector loop.  */
> > > > +       VPCMP   $0, (%rdi), %VMM0, %k0
> > > > +
> > > > +       KMOV    %k0, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x1)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       VPCMP   $0, VEC_SIZE(%rdi), %VMM0, %k0
> > > > +
> > > > +       KMOV    %k0, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x2)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       VPCMP   $0, (VEC_SIZE * 2)(%rdi), %VMM0, %k0
> > > > +
> > > > +       KMOV    %k0, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x3)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       VPCMP   $0, (VEC_SIZE * 3)(%rdi), %VMM0, %k0
> > > > +
> > > > +       KMOV    %k0, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x4)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +       /* Save pointer to find alignment adjustment.  */
> > > > +       movq    %rdi, %rax
> > > > +# endif
> > > > +       /* Align address to VEC_SIZE * 4 for loop.  */
> > > > +       andq    $-(VEC_SIZE * 4), %rdi
> > > > +
> > > > +       /* Add alignment difference to rdx.  */
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    %rdi, %rax
> > > > +#  ifdef USE_AS_WMEMCHR
> > > > +       SHR     $2, %RAX
> > > > +#  endif
> > > > +       addq    %rax, %rdx
> > > > +       jmp     L(loop_entry)
> > > > +# endif
> > > > +
> > > > +       /* 4 vector loop.  */
> > > > +       .p2align 5,,11
> > > > +L(loop):
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $(CHAR_PER_VEC * 4), %rdx
> > > > +       jbe     L(zero)
> > > > +L(loop_entry):
> > > > +# endif
> > > > +       VPCMP   $0, (VEC_SIZE * 4)(%rdi), %VMM0, %k1
> > > > +       VPCMP   $0, (VEC_SIZE * 5)(%rdi), %VMM0, %k2
> > > > +       VPCMP   $0, (VEC_SIZE * 6)(%rdi), %VMM0, %k3
> > > > +       VPCMP   $0, (VEC_SIZE * 7)(%rdi), %VMM0, %k4
> > > > +       KOR     %k1, %k2, %k5
> > > > +       KOR     %k3, %k4, %k6
> > > > +
> > > > +       subq    $-(VEC_SIZE * 4), %rdi
> > > > +       KORTEST %k5, %k6
> > > > +       jz      L(loop)
> > > > +
> > > > +       KMOV    %k1, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x1)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       KMOV    %k2, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x2)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       KMOV    %k3, %RAX
> > > > +       test    %RAX, %RAX
> > > > +       jnz     L(ret_vec_x3)
> > > > +
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       subq    $CHAR_PER_VEC, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +
> > > > +       /* At this point null [w]char must be in the fourth vector so no
> > > > +          need to check.  */
> > > > +       KMOV    %k4, %RAX
> > > > +
> > > > +L(ret_vec_x4):
> > > > +       bsf     %RAX, %RAX
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       cmp     %rax, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +       leaq    (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
> > > > +       ret
> > > > +
> > > > +       .p2align 5,,5
> > > > +L(ret_vec_x3):
> > > > +       bsf     %RAX, %RAX
> > > > +# ifndef USE_AS_RAWMEMCHR
> > > > +       cmp     %rax, %rdx
> > > > +       jbe     L(zero)
> > > > +# endif
> > > > +       leaq    (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
> > > > +       ret
> > > > +
> > > > +END (MEMCHR)
> > > > +#endif
> > > > diff --git a/sysdeps/x86_64/multiarch/memchr-evex512.S b/sysdeps/x86_64/multiarch/memchr-evex512.S
> > > > new file mode 100644
> > > > index 0000000000..47349d817a
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/multiarch/memchr-evex512.S
> > > > @@ -0,0 +1,7 @@
> > > > +# ifndef MEMCHR
> > > > +#  define MEMCHR       __memchr_evex512
> > > > +# endif
> > > > +
> > > > +#define VEC_SIZE        64
> > > > +
> > > > +#include "memchr-evex-base.S"
> > > > diff --git a/sysdeps/x86_64/multiarch/rawmemchr-evex512.S b/sysdeps/x86_64/multiarch/rawmemchr-evex512.S
> > > > new file mode 100644
> > > > index 0000000000..302d3cb055
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/multiarch/rawmemchr-evex512.S
> > > > @@ -0,0 +1,7 @@
> > > > +#ifndef RAWMEMCHR
> > > > +# define RAWMEMCHR     __rawmemchr_evex512
> > > > +#endif
> > > > +#define USE_AS_RAWMEMCHR       1
> > > > +#define MEMCHR RAWMEMCHR
> > > > +
> > > > +#include "memchr-evex512.S"
> > > > diff --git a/sysdeps/x86_64/multiarch/wmemchr-evex512.S b/sysdeps/x86_64/multiarch/wmemchr-evex512.S
> > > > new file mode 100644
> > > > index 0000000000..f45ed1db75
> > > > --- /dev/null
> > > > +++ b/sysdeps/x86_64/multiarch/wmemchr-evex512.S
> > > > @@ -0,0 +1,8 @@
> > > > +#ifndef WMEMCHR
> > > > +# define WMEMCHR       __wmemchr_evex512
> > > > +#endif
> > > > +
> > > > +#define MEMCHR WMEMCHR
> > > > +#define USE_AS_WMEMCHR 1
> > > > +
> > > > +#include "memchr-evex512.S"
> > > > --
> > > > 2.36.1
> > > >
>
> ping
see my reply to strrchr.


More information about the Libc-alpha mailing list