From: "Lucas A. M. Magalhaes" <lamm@linux.ibm.com>
To: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>,
libc-alpha@sourceware.org
Cc: anton@ozlabs.org
Subject: Re: [PATCH] powerpc64le: Optimize memcpy for POWER10
Date: Wed, 28 Apr 2021 11:22:40 -0300 [thread overview]
Message-ID: <161961976056.1414308.9345952117933428187@localhost.localdomain> (raw)
In-Reply-To: <20210428023138.795316-1-tuliom@linux.ibm.com>
Hi Tulio,
LGTM, All tests pass.
I have just a single suggestion of change.
Quoting Tulio Magno Quites Machado Filho via Libc-alpha (2021-04-27 23:31:38)
> This implementation is based on __memcpy_power8_cached and integrates
> suggestions from Anton Blanchard.
> It benefits from loads and stores with length for short lengths and for
> tail code, simplifying the code.
>
> All unaligned memory accesses use instructions that do not generate
> alignment interrupts on POWER10, making it safe to use on
> caching-inhibited memory.
>
> The main loop has also been modified in order to increase instruction
> throughput by reducing the dependency on updates from previous iterations.
>
> On average, this implementation provides around 30% improvement when
> compared to __memcpy_power7 and 10% improvement in comparison to
> __memcpy_power8_cached.
> ---
> sysdeps/powerpc/powerpc64/le/power10/memcpy.S | 198 ++++++++++++++++++
> sysdeps/powerpc/powerpc64/multiarch/Makefile | 3 +-
> .../powerpc64/multiarch/ifunc-impl-list.c | 6 +
> .../powerpc64/multiarch/memcpy-power10.S | 26 +++
> sysdeps/powerpc/powerpc64/multiarch/memcpy.c | 7 +
> 5 files changed, 239 insertions(+), 1 deletion(-)
> create mode 100644 sysdeps/powerpc/powerpc64/le/power10/memcpy.S
> create mode 100644 sysdeps/powerpc/powerpc64/multiarch/memcpy-power10.S
>
> diff --git a/sysdeps/powerpc/powerpc64/le/power10/memcpy.S b/sysdeps/powerpc/powerpc64/le/power10/memcpy.S
> new file mode 100644
> index 0000000000..f84acabec5
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/le/power10/memcpy.S
> @@ -0,0 +1,198 @@
> +/* Optimized memcpy implementation for POWER10.
> + Copyright (C) 2021 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>
> +
> +
> +#ifndef MEMCPY
> +# define MEMCPY memcpy
> +#endif
> +
> +/* __ptr_t [r3] memcpy (__ptr_t dst [r3], __ptr_t src [r4], size_t len [r5]);
> + Returns 'dst'. */
> +
> + .machine power9
> +ENTRY_TOCLESS (MEMCPY, 5)
> + CALL_MCOUNT 3
> +
> + /* Copy up to 16 bytes. */
> + sldi r6,r5,56 /* Prepare [l|st]xvl counter. */
> + lxvl v10,r4,r6
> + stxvl v10,r3,r6
> + subic. r6,r5,16 /* Return if len <= 16. */
> + blelr
> +
> + /* If len >= 256, assume nothing got copied before and copy
> + again. This might cause issues with overlapped memory, but memcpy
> + is not expected to treat overlapped memory. */
> + cmpdi r5,256
> + bge L(copy_ge_256)
> + /* 16 < len < 256 and the first 16 bytes have already been copied. */
> + addi r10,r3,16 /* Keep r3 intact as return value. */
> + addi r4,r4,16
> + subic r5,r5,16
> + b L(copy_lt_256) /* Avoid the main loop if len < 256. */
> +
Ok.
> + .p2align 5
> +L(copy_ge_256):
> + mr r10,r3 /* Keep r3 intact as return value. */
> + /* Align dst to 16 bytes. */
> + andi. r9,r10,0xf
> + beq L(dst_is_align_16)
> + lxv v10,0(r4)
> + subfic r12,r9,16
> + subf r5,r12,r5
> + add r4,r4,r12
> + stxv v10,0(r10)
> + add r10,r10,r12
You could use r3 for previous instructions here and just set up r10 here.
> +
> +L(dst_is_align_16):
> + srdi r9,r5,7 /* Divide by 128. */
> + mtctr r9
> + addi r6,r4,64
> + addi r7,r10,64
> +
Ok.
> +
> + /* Main loop, copy 128 bytes per iteration.
> + Use r6=src+64 and r7=dest+64 in order to reduce the dependency on
> + r4 and r10. */
> + .p2align 5
> +L(copy_128):
> +
> + lxv v10, 0(r4)
> + lxv v11, 16(r4)
> + lxv v12, 32(r4)
> + lxv v13, 48(r4)
> +
> + addi r4,r4,128
> +
> + stxv v10, 0(r10)
> + stxv v11, 16(r10)
> + stxv v12, 32(r10)
> + stxv v13, 48(r10)
> +
> + addi r10,r10,128
> +
> + lxv v10, 0(r6)
> + lxv v11, 16(r6)
> + lxv v12, 32(r6)
> + lxv v13, 48(r6)
> +
> + addi r6,r6,128
> +
> + stxv v10, 0(r7)
> + stxv v11, 16(r7)
> + stxv v12, 32(r7)
> + stxv v13, 48(r7)
> +
> + addi r7,r7,128
> +
> + bdnz L(copy_128)
> +
> + clrldi. r5,r5,64-7 /* Have we copied everything? */
> + beqlr
> +
Ok.
> + .p2align 5
> +L(copy_lt_256):
> + cmpdi r5,16
> + ble L(copy_le_16)
> + srdi. r9,r5,5 /* Divide by 32. */
> + beq L(copy_lt_32)
> + mtctr r9
> + /* Use r6=src+32, r7=dest+32, r8=src+64, r9=dest+64 in order to reduce
> + the dependency on r4 and r10. */
> + addi r6,r4,32
> + addi r7,r10,32
> + addi r8,r4,64
> + addi r9,r10,64
Ok.
> +
> + .p2align 5
> + /* Copy 32 bytes at a time, unaligned.
> + The loop is unrolled 3 times in order to reduce the dependency on
> + r4 and r10, copying up-to 96 bytes per iteration. */
> +L(copy_32):
> + lxv v10, 0(r4)
> + lxv v11, 16(r4)
> + stxv v10, 0(r10)
> + stxv v11, 16(r10)
> + bdz L(end_copy_32a)
> + addi r4,r4,96
> + addi r10,r10,96
> +
> + lxv v10, 0(r6)
> + lxv v11, 16(r6)
> + addi r6,r6,96
> + stxv v10, 0(r7)
> + stxv v11, 16(r7)
> + bdz L(end_copy_32b)
> + addi r7,r7,96
> +
> + lxv v12, 0(r8)
> + lxv v13, 16(r8)
> + addi r8,r8,96
> + stxv v12, 0(r9)
> + stxv v13, 16(r9)
> + addi r9,r9,96
> + bdnz L(copy_32)
> +
> + clrldi. r5,r5,64-5 /* Have we copied everything? */
> + beqlr
> + cmpdi r5,16
> + ble L(copy_le_16)
> + b L(copy_lt_32)
> +
Ok.
> + .p2align 5
> +L(end_copy_32a):
> + clrldi. r5,r5,64-5 /* Have we copied everything? */
> + beqlr
> + /* 32 bytes have been copied since the last update of r4 and r10. */
> + addi r4,r4,32
> + addi r10,r10,32
> + cmpdi r5,16
> + ble L(copy_le_16)
> + b L(copy_lt_32)
> +
Ok.
> + .p2align 5
> +L(end_copy_32b):
> + clrldi. r5,r5,64-5 /* Have we copied everything? */
> + beqlr
> + /* The last iteration of the loop copied 64 bytes. Update r4 and r10
> + accordingly. */
> + addi r4,r4,-32
> + addi r10,r10,-32
> + cmpdi r5,16
> + ble L(copy_le_16)
> +
Ok.
> + .p2align 5
> +L(copy_lt_32):
> + lxv v10, 0(r4)
> + stxv v10, 0(r10)
> + addi r4,r4,16
> + addi r10,r10,16
> + subic r5,r5,16
> +
Ok.
> + .p2align 5
> +L(copy_le_16):
> + sldi r6,r5,56
> + lxvl v10,r4,r6
> + stxvl v10,r10,r6
> + blr
Ok.
> +
> +
> +END_GEN_TB (MEMCPY,TB_TOCLESS)
> +libc_hidden_builtin_def (memcpy)
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/Makefile b/sysdeps/powerpc/powerpc64/multiarch/Makefile
> index 8aa46a3702..fdaa5ddb24 100644
> --- a/sysdeps/powerpc/powerpc64/multiarch/Makefile
> +++ b/sysdeps/powerpc/powerpc64/multiarch/Makefile
> @@ -1,5 +1,6 @@
> ifeq ($(subdir),string)
> -sysdep_routines += memcpy-power8-cached memcpy-power7 memcpy-a2 memcpy-power6 \
> +sysdep_routines += memcpy-power10 \
> + memcpy-power8-cached memcpy-power7 memcpy-a2 memcpy-power6 \
> memcpy-cell memcpy-power4 memcpy-ppc64 \
> memcmp-power8 memcmp-power7 memcmp-power4 memcmp-ppc64 \
> memset-power7 memset-power6 memset-power4 \
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> index 1a6993616f..7bb3028676 100644
> --- a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> +++ b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
> @@ -51,6 +51,12 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
> #ifdef SHARED
> /* Support sysdeps/powerpc/powerpc64/multiarch/memcpy.c. */
> IFUNC_IMPL (i, name, memcpy,
> +#ifdef __LITTLE_ENDIAN__
> + IFUNC_IMPL_ADD (array, i, memcpy,
> + hwcap2 & PPC_FEATURE2_ARCH_3_1
> + && hwcap & PPC_FEATURE_HAS_VSX,
> + __memcpy_power10)
> +#endif
> IFUNC_IMPL_ADD (array, i, memcpy, hwcap2 & PPC_FEATURE2_ARCH_2_07,
> __memcpy_power8_cached)
> IFUNC_IMPL_ADD (array, i, memcpy, hwcap & PPC_FEATURE_HAS_VSX,
Ok.
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/memcpy-power10.S b/sysdeps/powerpc/powerpc64/multiarch/memcpy-power10.S
> new file mode 100644
> index 0000000000..70e0fc3ed6
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/multiarch/memcpy-power10.S
> @@ -0,0 +1,26 @@
> +/* Optimized memcpy implementation for POWER10.
> + Copyright (C) 2021 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/>. */
> +
> +#if defined __LITTLE_ENDIAN__ && IS_IN (libc)
> +#define MEMCPY __memcpy_power10
> +
> +#undef libc_hidden_builtin_def
> +#define libc_hidden_builtin_def(name)
> +
> +#include <sysdeps/powerpc/powerpc64/le/power10/memcpy.S>
> +#endif
> diff --git a/sysdeps/powerpc/powerpc64/multiarch/memcpy.c b/sysdeps/powerpc/powerpc64/multiarch/memcpy.c
> index 5733192932..53ab32ef26 100644
> --- a/sysdeps/powerpc/powerpc64/multiarch/memcpy.c
> +++ b/sysdeps/powerpc/powerpc64/multiarch/memcpy.c
> @@ -36,8 +36,15 @@ extern __typeof (__redirect_memcpy) __memcpy_power6 attribute_hidden;
> extern __typeof (__redirect_memcpy) __memcpy_a2 attribute_hidden;
> extern __typeof (__redirect_memcpy) __memcpy_power7 attribute_hidden;
> extern __typeof (__redirect_memcpy) __memcpy_power8_cached attribute_hidden;
> +# if defined __LITTLE_ENDIAN__
> +extern __typeof (__redirect_memcpy) __memcpy_power10 attribute_hidden;
> +# endif
>
> libc_ifunc (__libc_memcpy,
> +# if defined __LITTLE_ENDIAN__
> + (hwcap2 & PPC_FEATURE2_ARCH_3_1 && hwcap & PPC_FEATURE_HAS_VSX)
> + ? __memcpy_power10 :
> +# endif
> ((hwcap2 & PPC_FEATURE2_ARCH_2_07) && use_cached_memopt)
> ? __memcpy_power8_cached :
> (hwcap & PPC_FEATURE_HAS_VSX)
Ok.
---
Lucas A. M. Magalhães
next prev parent reply other threads:[~2021-04-28 14:22 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2021-04-28 2:31 Tulio Magno Quites Machado Filho
2021-04-28 13:31 ` Raphael M Zinsly
2021-04-28 14:22 ` Lucas A. M. Magalhaes [this message]
2021-04-28 18:13 ` Raoni Fassina Firmino
2021-04-30 0:24 ` [PATCHv2] " Tulio Magno Quites Machado Filho
2021-04-30 12:56 ` Raoni Fassina Firmino
2021-04-30 21:19 ` Tulio Magno Quites Machado Filho
2021-04-30 12:10 ` [PATCH] " Tulio Magno Quites Machado Filho
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=161961976056.1414308.9345952117933428187@localhost.localdomain \
--to=lamm@linux.ibm.com \
--cc=anton@ozlabs.org \
--cc=libc-alpha@sourceware.org \
--cc=tuliom@linux.ibm.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).