x86_64: memset optimized with AVX512

Message ID CAMXFM3v6BMXYGD_EbJO_-GqKaEMaeCDHdrd9L+C8wxJRuofBvg@mail.gmail.com
State New, archived
Headers

Commit Message

Andrew Senkevich Dec. 11, 2015, 6:13 p.m. UTC
  2015-12-11 17:43 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>> Please make following changes:
>>
>> 1.  Change _avx512 to _avx512_no_vzeroupper.
>> 2.  Add a feature, Prefer_No_VZEROUPPER, to cpu-features.h, and set
>> it for KNL.
>> 3. Check Prefer_No_VZEROUPPER instead of AVX512DQ_Usable,
>> 4. Don't check AVX512DQ_Usable nor  Prefer_No_VZEROUPPER in
>> ifunc-impl-list.c.
>>
>
> I submitted a patch to enable SLM optimization for KNL:
>
> https://sourceware.org/ml/libc-alpha/2015-12/msg00221.html
>
> It is on hjl/32bit/master branch.   Please rebase your patch against
> mine since it adds KNL optimization.

Here ir rebased and updated version:

From f488d8572bc43c731b0ce054ce1f84db7d90eb61 Mon Sep 17 00:00:00 2001
From: Andrew Senkevich <andrew.senkevich@intel.com>
Date: Fri, 11 Dec 2015 20:58:57 +0300
Subject: [PATCH] Added memset optimized with AVX512 for KNL hardware.

It shows improvement up to 28% over AVX2 memset (performance results
attached at <https://sourceware.org/ml/libc-alpha/2015-12/msg00052.html>).

        * sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S: New file.
        * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new file.
        * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
        * sysdeps/x86_64/multiarch/memset.S: Added new IFUNC branch.
        * sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
        * sysdeps/x86/cpu-features.h (bit_Prefer_No_VZEROUPPER,
        index_Prefer_No_VZEROUPPER): New.
        * sysdeps/x86/cpu-features.c (init_cpu_features): Set the
        Prefer_No_VZEROUPPER for Knights Landing.
---
 ChangeLog                                          |  12 ++
 sysdeps/x86/cpu-features.c                         |   2 +
 sysdeps/x86/cpu-features.h                         |   4 +
 sysdeps/x86_64/multiarch/Makefile                  |   3 +-
 sysdeps/x86_64/multiarch/ifunc-impl-list.c         |  17 +-
 .../x86_64/multiarch/memset-avx512-no-vzeroupper.S | 194 +++++++++++++++++++++
 sysdeps/x86_64/multiarch/memset.S                  |   7 +
 sysdeps/x86_64/multiarch/memset_chk.S              |   7 +
 8 files changed, 243 insertions(+), 3 deletions(-)
 create mode 100644 sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S

--
1.8.3.1


--
WBR,
Andrew
  

Comments

H.J. Lu Dec. 16, 2015, 5:19 p.m. UTC | #1
On Fri, Dec 11, 2015 at 10:13 AM, Andrew Senkevich
<andrew.n.senkevich@gmail.com> wrote:
> 2015-12-11 17:43 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>>> Please make following changes:
>>>
>>> 1.  Change _avx512 to _avx512_no_vzeroupper.
>>> 2.  Add a feature, Prefer_No_VZEROUPPER, to cpu-features.h, and set
>>> it for KNL.
>>> 3. Check Prefer_No_VZEROUPPER instead of AVX512DQ_Usable,
>>> 4. Don't check AVX512DQ_Usable nor  Prefer_No_VZEROUPPER in
>>> ifunc-impl-list.c.
>>>
>>
>> I submitted a patch to enable SLM optimization for KNL:
>>
>> https://sourceware.org/ml/libc-alpha/2015-12/msg00221.html
>>
>> It is on hjl/32bit/master branch.   Please rebase your patch against
>> mine since it adds KNL optimization.
>
> Here ir rebased and updated version:
>
> From f488d8572bc43c731b0ce054ce1f84db7d90eb61 Mon Sep 17 00:00:00 2001
> From: Andrew Senkevich <andrew.senkevich@intel.com>
> Date: Fri, 11 Dec 2015 20:58:57 +0300
> Subject: [PATCH] Added memset optimized with AVX512 for KNL hardware.
>
> It shows improvement up to 28% over AVX2 memset (performance results
> attached at <https://sourceware.org/ml/libc-alpha/2015-12/msg00052.html>).
>
>         * sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S: New file.
>         * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new file.
>         * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
>         * sysdeps/x86_64/multiarch/memset.S: Added new IFUNC branch.
>         * sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
>         * sysdeps/x86/cpu-features.h (bit_Prefer_No_VZEROUPPER,
>         index_Prefer_No_VZEROUPPER): New.
>         * sysdeps/x86/cpu-features.c (init_cpu_features): Set the
>         Prefer_No_VZEROUPPER for Knights Landing.

Looks good to me.

Thanks.
  
Andrew Senkevich Dec. 18, 2015, 2:06 p.m. UTC | #2
2015-12-16 20:19 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
> On Fri, Dec 11, 2015 at 10:13 AM, Andrew Senkevich
> <andrew.n.senkevich@gmail.com> wrote:
>> 2015-12-11 17:43 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>>>> Please make following changes:
>>>>
>>>> 1.  Change _avx512 to _avx512_no_vzeroupper.
>>>> 2.  Add a feature, Prefer_No_VZEROUPPER, to cpu-features.h, and set
>>>> it for KNL.
>>>> 3. Check Prefer_No_VZEROUPPER instead of AVX512DQ_Usable,
>>>> 4. Don't check AVX512DQ_Usable nor  Prefer_No_VZEROUPPER in
>>>> ifunc-impl-list.c.
>>>>
>>>
>>> I submitted a patch to enable SLM optimization for KNL:
>>>
>>> https://sourceware.org/ml/libc-alpha/2015-12/msg00221.html
>>>
>>> It is on hjl/32bit/master branch.   Please rebase your patch against
>>> mine since it adds KNL optimization.
>>
>> Here ir rebased and updated version:
>>
>> From f488d8572bc43c731b0ce054ce1f84db7d90eb61 Mon Sep 17 00:00:00 2001
>> From: Andrew Senkevich <andrew.senkevich@intel.com>
>> Date: Fri, 11 Dec 2015 20:58:57 +0300
>> Subject: [PATCH] Added memset optimized with AVX512 for KNL hardware.
>>
>> It shows improvement up to 28% over AVX2 memset (performance results
>> attached at <https://sourceware.org/ml/libc-alpha/2015-12/msg00052.html>).
>>
>>         * sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S: New file.
>>         * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new file.
>>         * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
>>         * sysdeps/x86_64/multiarch/memset.S: Added new IFUNC branch.
>>         * sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
>>         * sysdeps/x86/cpu-features.h (bit_Prefer_No_VZEROUPPER,
>>         index_Prefer_No_VZEROUPPER): New.
>>         * sysdeps/x86/cpu-features.c (init_cpu_features): Set the
>>         Prefer_No_VZEROUPPER for Knights Landing.
>
> Looks good to me.
>
> Thanks.

Is it OK for trunk?


--
WBR,
Andrew
  
H.J. Lu Dec. 18, 2015, 2:46 p.m. UTC | #3
On Fri, Dec 18, 2015 at 6:06 AM, Andrew Senkevich
<andrew.n.senkevich@gmail.com> wrote:
> 2015-12-16 20:19 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>> On Fri, Dec 11, 2015 at 10:13 AM, Andrew Senkevich
>> <andrew.n.senkevich@gmail.com> wrote:
>>> 2015-12-11 17:43 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>>>>> Please make following changes:
>>>>>
>>>>> 1.  Change _avx512 to _avx512_no_vzeroupper.
>>>>> 2.  Add a feature, Prefer_No_VZEROUPPER, to cpu-features.h, and set
>>>>> it for KNL.
>>>>> 3. Check Prefer_No_VZEROUPPER instead of AVX512DQ_Usable,
>>>>> 4. Don't check AVX512DQ_Usable nor  Prefer_No_VZEROUPPER in
>>>>> ifunc-impl-list.c.
>>>>>
>>>>
>>>> I submitted a patch to enable SLM optimization for KNL:
>>>>
>>>> https://sourceware.org/ml/libc-alpha/2015-12/msg00221.html
>>>>
>>>> It is on hjl/32bit/master branch.   Please rebase your patch against
>>>> mine since it adds KNL optimization.
>>>
>>> Here ir rebased and updated version:
>>>
>>> From f488d8572bc43c731b0ce054ce1f84db7d90eb61 Mon Sep 17 00:00:00 2001
>>> From: Andrew Senkevich <andrew.senkevich@intel.com>
>>> Date: Fri, 11 Dec 2015 20:58:57 +0300
>>> Subject: [PATCH] Added memset optimized with AVX512 for KNL hardware.
>>>
>>> It shows improvement up to 28% over AVX2 memset (performance results
>>> attached at <https://sourceware.org/ml/libc-alpha/2015-12/msg00052.html>).
>>>
>>>         * sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S: New file.
>>>         * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new file.
>>>         * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
>>>         * sysdeps/x86_64/multiarch/memset.S: Added new IFUNC branch.
>>>         * sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
>>>         * sysdeps/x86/cpu-features.h (bit_Prefer_No_VZEROUPPER,
>>>         index_Prefer_No_VZEROUPPER): New.
>>>         * sysdeps/x86/cpu-features.c (init_cpu_features): Set the
>>>         Prefer_No_VZEROUPPER for Knights Landing.
>>
>> Looks good to me.
>>
>> Thanks.
>
> Is it OK for trunk?
>
>

Yes, please check it in.

Thanks.
  

Patch

diff --git a/ChangeLog b/ChangeLog
index 80f4635..c530210 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,15 @@ 
+2015-12-11  Andrew Senkevich  <andrew.senkevich@intel.com>
+
+       * sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S: New file.
+       * sysdeps/x86_64/multiarch/Makefile (sysdep_routines): Added new file.
+       * sysdeps/x86_64/multiarch/ifunc-impl-list.c: Added new tests.
+       * sysdeps/x86_64/multiarch/memset.S: Added new IFUNC branch.
+       * sysdeps/x86_64/multiarch/memset_chk.S: Likewise.
+       * sysdeps/x86/cpu-features.h (bit_Prefer_No_VZEROUPPER,
+       index_Prefer_No_VZEROUPPER): New feature.
+       * sysdeps/x86/cpu-features.c (init_cpu_features): Set the
+       Prefer_No_VZEROUPPER for Knights Landing.
+
 2015-12-10  Aurelien Jarno  <aurelien@aurel32.net>
            Jakub Wilk  <jwilk@debian.org>

diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
index 6a132f7..424b203 100644
--- a/sysdeps/x86/cpu-features.c
+++ b/sysdeps/x86/cpu-features.c
@@ -107,6 +107,8 @@  init_cpu_features (struct cpu_features *cpu_features)

            case 0x57:
              /* Knights Landing.  Enable Silvermont optimizations.  */
+             cpu_features->feature[index_Prefer_No_VZEROUPPER]
+               |= bit_Prefer_No_VZEROUPPER;

            case 0x37:
            case 0x4a:
diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h
index 93bee69..5e09a2a 100644
--- a/sysdeps/x86/cpu-features.h
+++ b/sysdeps/x86/cpu-features.h
@@ -34,6 +34,7 @@ 
 #define bit_I586                       (1 << 14)
 #define bit_I686                       (1 << 15)
 #define bit_Prefer_MAP_32BIT_EXEC      (1 << 16)
+#define bit_Prefer_No_VZEROUPPER       (1 << 17)

 /* CPUID Feature flags.  */

@@ -99,6 +100,8 @@ 
 # define index_I586                    FEATURE_INDEX_1*FEATURE_SIZE
 # define index_I686                    FEATURE_INDEX_1*FEATURE_SIZE
 # define index_Prefer_MAP_32BIT_EXEC   FEATURE_INDEX_1*FEATURE_SIZE
+# define index_Prefer_No_VZEROUPPER    FEATURE_INDEX_1*FEATURE_SIZE
+

 # if defined (_LIBC) && !IS_IN (nonlib)
 #  ifdef __x86_64__
@@ -251,6 +254,7 @@  extern const struct cpu_features *__get_cpu_features (void)
 # define index_I586                    FEATURE_INDEX_1
 # define index_I686                    FEATURE_INDEX_1
 # define index_Prefer_MAP_32BIT_EXEC   FEATURE_INDEX_1
+# define index_Prefer_No_VZEROUPPER     FEATURE_INDEX_1

 #endif /* !__ASSEMBLER__ */

diff --git a/sysdeps/x86_64/multiarch/Makefile
b/sysdeps/x86_64/multiarch/Makefile
index bb811c2..b2e31ef 100644
--- a/sysdeps/x86_64/multiarch/Makefile
+++ b/sysdeps/x86_64/multiarch/Makefile
@@ -18,7 +18,8 @@  sysdep_routines += strncat-c stpncpy-c strncpy-c
strcmp-ssse3 \
                   stpcpy-sse2-unaligned stpncpy-sse2-unaligned \
                   strcat-sse2-unaligned strncat-sse2-unaligned \
                   strchr-sse2-no-bsf memcmp-ssse3 strstr-sse2-unaligned \
-                  strcspn-c strpbrk-c strspn-c varshift memset-avx2
+                  strcspn-c strpbrk-c strspn-c varshift memset-avx2 \
+                  memset-avx512-no-vzeroupper
 CFLAGS-varshift.c += -msse4
 CFLAGS-strcspn-c.c += -msse4
 CFLAGS-strpbrk-c.c += -msse4
diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
index 5c0c219..9771474 100644
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
@@ -20,6 +20,7 @@ 
 #include <string.h>
 #include <wchar.h>
 #include <ifunc-impl-list.h>
+#include <sysdep.h>
 #include "init-arch.h"

 /* Maximum number of IFUNC implementations.  */
@@ -76,14 +77,26 @@  __libc_ifunc_impl_list (const char *name, struct
libc_ifunc_impl *array,
                              __memset_chk_sse2)
              IFUNC_IMPL_ADD (array, i, __memset_chk,
                              HAS_ARCH_FEATURE (AVX2_Usable),
-                             __memset_chk_avx2))
+                             __memset_chk_avx2)
+#ifdef HAVE_AVX512_ASM_SUPPORT
+             IFUNC_IMPL_ADD (array, i, __memset_chk,
+                             HAS_ARCH_FEATURE (AVX512F_Usable),
+                             __memset_chk_avx512_no_vzeroupper)
+#endif
+             )

   /* Support sysdeps/x86_64/multiarch/memset.S.  */
   IFUNC_IMPL (i, name, memset,
              IFUNC_IMPL_ADD (array, i, memset, 1, __memset_sse2)
              IFUNC_IMPL_ADD (array, i, memset,
                              HAS_ARCH_FEATURE (AVX2_Usable),
-                             __memset_avx2))
+                             __memset_avx2)
+#ifdef HAVE_AVX512_ASM_SUPPORT
+             IFUNC_IMPL_ADD (array, i, memset,
+                             HAS_ARCH_FEATURE (AVX512F_Usable),
+                             __memset_avx512_no_vzeroupper)
+#endif
+            )

   /* Support sysdeps/x86_64/multiarch/stpncpy.S.  */
   IFUNC_IMPL (i, name, stpncpy,
diff --git a/sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S
b/sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S
new file mode 100644
index 0000000..eca8ca0
--- /dev/null
+++ b/sysdeps/x86_64/multiarch/memset-avx512-no-vzeroupper.S
@@ -0,0 +1,194 @@ 
+/* memset optimized with AVX512 for KNL hardware.
+   Copyright (C) 2015 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
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+#if defined HAVE_AVX512_ASM_SUPPORT && IS_IN (libc)
+
+#include "asm-syntax.h"
+#ifndef MEMSET
+# define MEMSET __memset_avx512_no_vzeroupper
+# define MEMSET_CHK __memset_chk_avx512_no_vzeroupper
+#endif
+
+       .section .text,"ax",@progbits
+#if defined PIC
+ENTRY (MEMSET_CHK)
+       cmpq    %rdx, %rcx
+       jb      HIDDEN_JUMPTARGET (__chk_fail)
+END (MEMSET_CHK)
+#endif
+
+ENTRY (MEMSET)
+       vpxor   %xmm0, %xmm0, %xmm0
+       vmovd   %esi, %xmm1
+       lea     (%rdi, %rdx), %rsi
+       mov     %rdi, %rax
+       vpshufb %xmm0, %xmm1, %xmm0
+       cmp     $16, %rdx
+       jb      L(less_16bytes)
+       cmp     $512, %rdx
+       vbroadcastss    %xmm0, %zmm2
+       ja      L(512bytesormore)
+       cmp     $256, %rdx
+       jb      L(less_256bytes)
+       vmovups %zmm2, (%rdi)
+       vmovups %zmm2, 0x40(%rdi)
+       vmovups %zmm2, 0x80(%rdi)
+       vmovups %zmm2, 0xC0(%rdi)
+       vmovups %zmm2, -0x100(%rsi)
+       vmovups %zmm2, -0xC0(%rsi)
+       vmovups %zmm2, -0x80(%rsi)
+       vmovups %zmm2, -0x40(%rsi)
+       ret
+
+L(less_256bytes):
+       cmp     $128, %dl
+       jb      L(less_128bytes)
+       vmovups %zmm2, (%rdi)
+       vmovups %zmm2, 0x40(%rdi)
+       vmovups %zmm2, -0x80(%rsi)
+       vmovups %zmm2, -0x40(%rsi)
+       ret
+
+L(less_128bytes):
+       cmp     $64, %dl
+       jb      L(less_64bytes)
+       vmovups %zmm2, (%rdi)
+       vmovups %zmm2, -0x40(%rsi)
+       ret
+
+L(less_64bytes):
+       cmp     $32, %dl
+       jb      L(less_32bytes)
+       vmovdqu %ymm2, (%rdi)
+       vmovdqu %ymm2, -0x20(%rsi)
+       ret
+
+L(less_32bytes):
+       vmovdqu %xmm0, (%rdi)
+       vmovdqu %xmm0, -0x10(%rsi)
+       ret
+
+L(less_16bytes):
+       cmp     $8, %dl
+       jb      L(less_8bytes)
+       vmovq   %xmm0, (%rdi)
+       vmovq   %xmm0, -0x08(%rsi)
+       ret
+
+L(less_8bytes):
+       vmovd   %xmm0, %ecx
+       cmp     $4, %dl
+       jb      L(less_4bytes)
+       mov     %ecx, (%rdi)
+       mov     %ecx, -0x04(%rsi)
+       ret
+
+L(less_4bytes):
+       cmp     $2, %dl
+       jb      L(less_2bytes)
+       mov     %cx, (%rdi)
+       mov     %cx, -0x02(%rsi)
+       ret
+
+L(less_2bytes):
+       cmp     $1, %dl
+       jb      L(less_1bytes)
+       mov     %cl, (%rdi)
+L(less_1bytes):
+       ret
+
+L(512bytesormore):
+       mov     __x86_shared_cache_size_half(%rip), %rcx
+       cmp     %rcx, %rdx
+       ja      L(preloop_large)
+       cmp     $1024, %rdx
+       ja      L(1024bytesormore)
+
+       vmovups %zmm2, (%rdi)
+       vmovups %zmm2, 0x40(%rdi)
+       vmovups %zmm2, 0x80(%rdi)
+       vmovups %zmm2, 0xC0(%rdi)
+       vmovups %zmm2, 0x100(%rdi)
+       vmovups %zmm2, 0x140(%rdi)
+       vmovups %zmm2, 0x180(%rdi)
+       vmovups %zmm2, 0x1C0(%rdi)
+       vmovups %zmm2, -0x200(%rsi)
+       vmovups %zmm2, -0x1C0(%rsi)
+       vmovups %zmm2, -0x180(%rsi)
+       vmovups %zmm2, -0x140(%rsi)
+       vmovups %zmm2, -0x100(%rsi)
+       vmovups %zmm2, -0xC0(%rsi)
+       vmovups %zmm2, -0x80(%rsi)
+       vmovups %zmm2, -0x40(%rsi)
+       ret
+
+/* Align on 64 and loop with aligned stores.  */
+L(1024bytesormore):
+       sub     $0x100, %rsi
+       vmovups %zmm2, (%rax)
+       and     $-0x40, %rdi
+       add     $0x40, %rdi
+
+L(gobble_256bytes_loop):
+       vmovaps %zmm2, (%rdi)
+       vmovaps %zmm2, 0x40(%rdi)
+       vmovaps %zmm2, 0x80(%rdi)
+       vmovaps %zmm2, 0xC0(%rdi)
+       add     $0x100, %rdi
+       cmp     %rsi, %rdi
+       jb      L(gobble_256bytes_loop)
+       vmovups %zmm2, (%rsi)
+       vmovups %zmm2, 0x40(%rsi)
+       vmovups %zmm2, 0x80(%rsi)
+       vmovups %zmm2, 0xC0(%rsi)
+       ret
+
+/* Align on 128 and loop with non-temporal stores.  */
+L(preloop_large):
+       and     $-0x80, %rdi
+       add     $0x80, %rdi
+       vmovups %zmm2, (%rax)
+       vmovups %zmm2, 0x40(%rax)
+       sub     $0x200, %rsi
+
+L(gobble_512bytes_nt_loop):
+       vmovntdq %zmm2, (%rdi)
+       vmovntdq %zmm2, 0x40(%rdi)
+       vmovntdq %zmm2, 0x80(%rdi)
+       vmovntdq %zmm2, 0xC0(%rdi)
+       vmovntdq %zmm2, 0x100(%rdi)
+       vmovntdq %zmm2, 0x140(%rdi)
+       vmovntdq %zmm2, 0x180(%rdi)
+       vmovntdq %zmm2, 0x1C0(%rdi)
+       add     $0x200, %rdi
+       cmp     %rsi, %rdi
+       jb      L(gobble_512bytes_nt_loop)
+       sfence
+       vmovups %zmm2, (%rsi)
+       vmovups %zmm2, 0x40(%rsi)
+       vmovups %zmm2, 0x80(%rsi)
+       vmovups %zmm2, 0xC0(%rsi)
+       vmovups %zmm2, 0x100(%rsi)
+       vmovups %zmm2, 0x140(%rsi)
+       vmovups %zmm2, 0x180(%rsi)
+       vmovups %zmm2, 0x1C0(%rsi)
+       ret
+END (MEMSET)
+#endif
diff --git a/sysdeps/x86_64/multiarch/memset.S
b/sysdeps/x86_64/multiarch/memset.S
index dbc00d2..5fa360d 100644
--- a/sysdeps/x86_64/multiarch/memset.S
+++ b/sysdeps/x86_64/multiarch/memset.S
@@ -30,6 +30,13 @@  ENTRY(memset)
        HAS_ARCH_FEATURE (AVX2_Usable)
        jz      2f
        leaq    __memset_avx2(%rip), %rax
+#ifdef HAVE_AVX512_ASM_SUPPORT
+       HAS_ARCH_FEATURE (AVX512F_Usable)
+       jz      2f
+       HAS_ARCH_FEATURE (Prefer_No_VZEROUPPER)
+       jz      2f
+       leaq    __memset_avx512_no_vzeroupper(%rip), %rax
+#endif
 2:     ret
 END(memset)
 #endif
diff --git a/sysdeps/x86_64/multiarch/memset_chk.S
b/sysdeps/x86_64/multiarch/memset_chk.S
index e2abb15..d461a24 100644
--- a/sysdeps/x86_64/multiarch/memset_chk.S
+++ b/sysdeps/x86_64/multiarch/memset_chk.S
@@ -30,6 +30,13 @@  ENTRY(__memset_chk)
        HAS_ARCH_FEATURE (AVX2_Usable)
        jz      2f
        leaq    __memset_chk_avx2(%rip), %rax
+#ifdef HAVE_AVX512_ASM_SUPPORT
+       HAS_ARCH_FEATURE (AVX512F_Usable)
+       jz      2f
+       HAS_ARCH_FEATURE (Prefer_No_VZEROUPPER)
+       jz      2f
+       leaq    __memset_chk_avx512_no_vzeroupper(%rip), %rax
+#endif
 2:     ret
 END(__memset_chk)